Main takeaways:

1. Generative artificial intelligence promises the ultimate form of transmission: instant access to all the world’s knowledge. The models are trained on all available data and their users reach every conceivable (and probable) answer.

2. The first rungs of the professional ladder are being removed and the consequences are still uncertain. The “grunt work” handed to juniors served as an implicit form of transmission on arrival in the corporate world. By automating them, companies are abolishing “their school” without having decided to and without having solved the shortage of seniors that follows.

3. The cognitive and relational friction, removed by default, has to be rebuilt by choice. A few companies and schools are imposing arrangements that require framing the problem, failing or passing an oral check before turning to AI.

4. The “job apocalypse” narrative has been tempered by its own proponents, who have recently revised their forecasts downward. Macroeconomic data confirm this revision, showing no specific effect of AI on employment. That said, the data show that young people are operating in a labor market that is already stacked against them.

5. What remains to be transmitted is less a skill than a posture toward what is coming and how to make the most of it. Repetition, apprenticeship, and patience are growing rarer because AI makes the shortcut tempting. They grow more precious for the same reason.

Subscribe now

The full version

Barely a few days after our dinner on trust, Leo XIV opened Magnifica Humanitas, his encyclical on AI, with this sentence : « Each generation inherits the task of shaping its own era, » The phrasing is not new, but it takes on a particular resonance in our present moment.

What exactly do we hand down, when the age reshapes itself faster than we can understand it?

This question ran through all our conversations across Season 1 of the Tandem dinners. Whether we were talking about careers upended by AI; the considerable impact of these tools on our cognition and our relationships; or even the immense productivity of a single well-equipped person, the themes of transmission and learning kept returning among the guests.

So we decided to devote the last dinner before the summer break to it. That evening, about ten people from worlds that rarely intersect gathered in Paris, from consulting to training to the student world. Phones stayed at the door and the Chatham House rule covered the exchanges. What you are about to read gives shape to a collective reflection: what was prepared beforehand and thought through around the table, then extended in the days that followed. The aim: to share it with everyone who lives with these questions without necessarily having the space to ask them aloud.

Transmitting means passing on to someone what one has made one’s own. Transmittere, in Latin, means to make something « pass through » and that « through » is anything but neutral. It denotes a resistance, a time of appropriation that must be experienced before anything can be transmitted, be it a body of knowledge, a judgment or a way of inhabiting uncertainty. Transmission demands time and friction.

AI disrupts this mechanism at both ends. It occupies the position of a transmitter without ever having been a receiver, giving back the knowledge accumulated by whole generations without having passed through it or absorbed it through effort. On top of that, for the first time at scale, the asymmetry of transmission no longer holds: command (or lack of command) of these new tools reshuffles the cards of knowledge and experience, especially inside companies, where the question arises of whether to keep recruiting easily automatable « junior » roles.

It was within this frame that we opened the debate among entrepreneurs, executives, teachers and students.

The rupture of apprenticeship

Who does not remember the long hours spent, early in a career, on the thankless tasks that English speakers call grunt work? One of our guests put it this way:

« Juniors earned their legitimacy by handling that thankless volume. It was tedious and badly paid, but it was the school. »

The junior who drafts a memo learns to frame a problem. The senior who corrects it transmits without having to spell it out. Tasks of execution build a field expertise and a professional intuition that no classroom training manages to install.

This mechanism has a name in the science of education. Victoria Marsick and Karen Watkins formalized it as early as 1990 under the term incidental learning : a form of learning that emerges from the work itself, as an unintended by-product of another activity. Grunt work is one of its most complete forms. No organization officially conceives of it as an object of transmission, yet it is perfectly woven into the gestures of the craft.

Now the AI systems deployed in companies absorb these entry-level tasks first: document research, first drafts, debugging, breaking a problem down. By automating this volume, companies effectively remove their environment of incidental learning. The effect plays out over two horizons. In the short term, juniors equipped with AI produce almost as fast as seniors, which installs an illusion of performance and masks the debt building up in the background. Over the longer term, the pyramid empties from the top, because no one will be able to settle the complex decisions that AI cannot yet make.

These observations line up with what a survey published by BCG in June 2026 among seventy executives documents : 53 % of them already observe a slowdown in the development of their juniors. A guest from consulting confirmed the feeling by asking: « Can you still become a senior in a field if you were never a junior first? » An experimental study run by Anthropic in February of the same year gives the measure of it : junior developers assisted by AI to master a new Python library learn faster than the control group but lose 17 % of their conceptual mastery. The largest gap concerns debugging skills, the very ones the company will expect of them to supervise the code produced by AI.

The paradox then closes back on the companies themselves. Convinced by the apparent performance of AI-equipped juniors, they raise their expectations and hand them tasks that an employee with five to ten years of experience used to take on, in the words of an executive interviewed by BCG. The imbalance is severe: more is asked of juniors whose real training is slowing down, for want of the friction that used to build it.

Share

The “job apocalypse” narrative tested against the facts

This imbalance sits within a larger narrative. For three years, one prediction has saturated public space : generative AI will bring about a job apocalypse whose price the younger generations will pay. As early as 2025, Dario Amodei, founder of Anthropic, argued that half of white-collar junior jobs could disappear within five years, joined by his direct competitor Sam Altman a few months later.

These predictions, made by the very people building the technology, had a performative effect on the audiences they named, in particular the younger generations who internalized the forecast. In the United States, commencement speeches that are favorable to AI are regularly booed. According to Gallup, only 22 % of Gen Z say they are enthusiastic about the technology, down fourteen points in a year. The phenomenon has a name, FOBO, fear of being obsolete. It cuts across ages. To the juniors’ fear of not finding their place answers the seniors’ fear of being made obsolete by tools they have not mastered.

This fear produces paradoxical behavior inside organizations. According to a recent Writer.com survey, “AI adoption in the enterprise”, 29 % of employees admit to having sabotaged an AI rollout in their company, a share that climbs to 44 % among those under 30. These same employees nonetheless know that refusing the tool exposes them to layoffs more than the reverse. They prefer the risk of being sidelined to that of speeding up their own replacement.

For a few months now this narrative has met a denial from the very people who carried it, in a context where several of their companies are preparing to go public, each valued at around a trillion dollars. Sam Altman has acknowledged being largely wrong about the expected impact. Dario Amodei now speaks of an AI that would multiply human output rather than replace it, a thesis that leans on the Jevons paradox formulated in the nineteenth century and taken up by several promoters of AI : making a resource more efficient does not reduce its consumption but extends it to markets that were until then out of reach. The framework remains theoretical, since no data confirms its application to generative AI to date.

The available data stays consistent with this revised reading. As of June 2026, the Yale Budget Lab finds no statistical break in US employment since the release of ChatGPT in late 2022. A May 2026 paper by Lambert and Schindler, covering more than 243 million hires across four countries, goes further : as soon as remote work is introduced as a control variable, the specific effect of AI on junior hiring almost entirely disappears. Remote work makes supervising beginners more costly and undermines investment in profiles without experience. Attributing to AI alone what stems from a broader transformation of working conditions is a shortcut.

Finally, another signal weakens this apocalyptic reading: the economics of large-scale deployments turns out to be more fragile than announced. According to a Fortune article in May 2026, several pioneering companies are discovering that the running cost of their AI agents exceeds that of the employees they were meant to replace, once the cost of the tokens consumed and the residual human supervision that remains necessary are taken into account. Replacing human work with machine work is not an economically neutral operation, even if the way of working does restructure itself.

A study published in late June 2026 by Ramp and Revelio Labs, covering 21 559 US companies, adds a complementary signal. Companies that invest heavily in AI, around thirty dollars per employee per month in the first three months of adoption, saw their headcount grow by 10.2 % over the following two years and their entry-level hiring by 12 %. Low-adoption companies record no statistically significant effect. The authors specify that these gains concentrate in the technology sector. They add that the correlation observed does not prove causation. The founder of an AI-native company present at our dinner gave a concrete illustration : fifteen people today, ten hires planned by year-end, in a tech sector that captures the largest share of these gains.

Outside the sector pockets Ramp identifies, the macroeconomic figures point more toward continuity, but young people’s felt experience remains legitimate, because the weight of the reconfiguration concentrates on them. France’s national statistics office, Insee, measured a 21.5 % unemployment rate among 15 to 24 year olds at the end of 2025, up 2.4 points over the quarter, while the overall rate rose by only 0.2 point over the same period. That represents 742 000 young people shut out of the labor market, 126 000 more in a year. IT employment among 15 to 29 year olds fell there by 7.4 % in the last quarter of 2025, even as the sector’s value added kept rising. In the United States, a study by the Stanford Digital Economy Lab documents a 16 % decline in the relative employment of 22 to 25 year olds in the occupations most exposed to AI, while that of seniors stays stable : the first steps by which one used to enter a profession are disappearing.

A student among our guests confirmed it: among his classmates, many still have not found an internship despite flawless applications, including the top of his class, who recounts sending around a hundred applications only to receive two replies. This example illustrates what the figures do not show: students sense that expertise matters, but they no longer know how to build it when the tools remove the friction that used to develop it.

Reinventing friction, rethinking transmission

Since friction disappears when AI absorbs entry-level tasks, we now have to design explicitly what used to be transmitted invisibly. A few organizations are already trying, with different logics.

At Shell, juniors have to frame the problem on their own before they can turn to AI to refine it. This sequencing produces better questions and clearer rationales, according to the results published by BCG. Friction is reintroduced at the precise point where automation removes it: the framing phase, the one that forms judgment.

Two other practices circulate the same requirement between generations in opposite directions. Salesforce has generalized pair programming by forming pairs in which the advanced AI user works alongside a novice colleague. The junior learns by direct observation, without going through formal training. Conversely, an innovative program organizes a co-development in which the junior who is comfortable with AI brings the tool and the senior brings judgment. A genuine co-development, its founder specifies, not disguised reverse mentoring.

These arrangements deliberately rebuild friction where efficiency could have done without it. An executive from the world of training offered an image that contrasts two ways of learning: the regatta and the offshore race. Training for the regatta prepares you to excel on a marked course, where the rules are known and the winner is decided at the margin. Training for the offshore race prepares you to navigate without markers, facing conditions that change and problems that no one has solved before you. These are two different kinds of agency, not two levels of difficulty. The first model long dominated schooling. AI makes the second indispensable, because it automates precisely what happened on the marked courses and leaves to humans the ambiguous territory of ill-defined problems.

The academic world faces this shift under a constraint of its own: where a company reconfigures a workflow that produces immediate value, the university has to reintroduce resistance in a setting where the student has nothing to deliver except the proof of having learned. A professor at a leading school explained that his institution pushes its students toward the experimentation of scientific research, in place of more classic exams (notably the final dissertation, obsolete in the age of AI). The founder of another academic institution describes a more operational setup in three parts. Some assessments are done without AI, on paper or with software that records the student and locks the browser. About a third is done with AI, on condition of keeping an AI logbook and providing coherent « logs ». A systematic oral then verifies real command of what was produced. The principle is explicit: most institutions leave their students in a gray zone where neither use nor verification is clear, which produces neither honest learning nor punishable cheating. Deciding which uses are allowed is here a necessity to maintain integrity and trust.

This shift is also visible on the recruitment side. Several recruiters present, including an « AI-first » entrepreneur, converged around the table on one observation: their main criterion has shifted from the signal of the diploma toward what one of the guests called « texture ». What they look for lies in a way of being and a singular way of approaching problems, a trajectory that testifies to choices and stories rather than to conformity with the expected path.

That said, these new routes of recognition assume resources that not everyone has. Selection shifts toward two terrains. The first is demonstrable skills on concrete projects, which a portfolio or a public contribution makes visible. The second is the social capital built in networks, through encounters and communities of belonging. A student from a modest family, geographically far from the ecosystems where opportunities circulate, with no spare time to build a portfolio outside their coursework, finds themselves excluded from both routes at the same time as they were already excluded from the classic route by the devalued diploma.

This blind spot ran through our conversation : the question of people in economic vulnerability, who have neither the social capital nor the demonstrable projects, found no answer. Recent initiatives such as the First Chance program run by Chance with Google Labs sketch a path, without the problem of massively funding reconversion being posed at its true scale.

What remains to be handed down

What the conversation lets through goes beyond the question of work. One of our guests put it this way: we have the material means to rethink how we live and produce, we may only lack the imagination to see what is beginning to emerge. The sentence moves the subject from diagnosis toward stance. What remains to be handed down is less a skill than a way of holding oneself in the face of what is coming.

In the corporate world, the example of Hermès offers an inspiring story. When demand explodes, the house refuses to set its production by the market. It paces its openings, a new leather-goods workshop every eighteen to twenty-four months, to its capacity to train artisans. An eighteen-month apprenticeship precedes the first finished bag. Each piece then carries the stamp of the person who made it. This refusal to sacrifice quality to volume produces a 40.5 % operating margin, a level no other luxury house reaches. Hermès has drawn what its leader calls a stitching line. Above it, nothing is delegated, neither to a machine nor to a subcontractor. Below it, they use ERP systems, e-commerce, cutting machines. The question posed to any organization is that of its own line: what part of its work would lose its value if we learned it had been done by a machine?

This stance is not an object that can be transmitted through content. It is built over time through exercises that resemble what craft has always practiced. Repetition refines the gesture. Companionship shapes judgment. Patience accepts that mastery cannot be decreed. AI makes these qualities rarer because it makes the shortcut more seductive. It makes them more precious for the same reason. What is quick to see is quick to copy. What is built slowly endures.

To close our exchanges, the words that crossed the table all said the same thing from different angles: transmission is no longer (only) about tools or skills, it is above all about a way of inhabiting the gesture and the bond:

craft, companionship, friction, beauty, wonder, doubt, creation, connection, moral robustness.

To hand something down in 2026 may be to hand down the desire to imagine what comes next rather than the fear of facing it unprepared.

Thanks for reading us! Sign up if you’d like to be part of the 2nd season of Tandem.

Marine and Margaux

Subscribe now

What research shows today

Productivity gains are real. So are second-order impacts.

The numbers are clear when it comes to speed, as shown in three studies.

A controlled study with 95 freelance developers found that GitHub Copilot cut completion time by 55.8% on a standard coding task. A field experiment from Wharton with 758 BCG consultants found that on tasks within AI’s current ability, people completed 12% more tasks and finished 25% faster, with higher-quality output. At P&G, individuals working with AI matched the solution quality of two-person teams that didn’t use it, calling AI a “Cybernetic Teammate”.

The same studies revealed a limit: the same BCG consultants, given a task outside AI’s ability, were 19 points less likely to land on the correct answer than consultants working without AI at all. The model sounded just as confident either way. Speed goes up. Depth goes down. Researchers call this the “jagged frontier”: exceptional on familiar terrain. Unreliable at the edges, with no signal to tell you when you’ve crossed over.

Retention is the cost that shows up later.

A randomized trial with developers learning a new programming library found that using AI cut comprehension quiz scores by 17%, with no net time savings. The reading and prompting time canceled out the speed gained from the generated code. A separate trial with 120 students gave a surprise test 45 days after the original session. Students who’d learned the material the traditional way scored 68.5%. Students who’d used ChatGPT scored 57.5%, even after accounting for the fact that they’d spent less time studying. Both studies point to the same mechanism: the friction AI removes is often the friction that makes things stick (also called “productive struggle” for education experts).

Confidence increases regardless of accuracy.

A Wharton study with 1,372 participants found that people who consulted an AI assistant gained 25 points in accuracy when it was right and lost 15 points when it was wrong, yet they reported feeling more confident either way. Building on Daniel Kahneman’s System 1 (fast intuition) and System 2 (slow, deliberate reasoning), they introduced a “System 3”, an artificial cognition sitting outside our brain. Yet what happens when the AI’s perception in this System 3 makes people stop thinking for themselves? The researchers call this cognitive surrender: the feeling of competence and actual competence come apart. One trusts the output because it sounds right, not because they verified it.

Cognitive surrender is different from cognitive offloading.

Cognitive offloading is rational. The calculator or the GPS are often pulled out as an illustration of what is taken off your cognitive plate: you delegate a task you don’t need to do yourself (eg a complex multiplication or the detailed itinerary across cities). Surrender is something deeper. You stop thinking because the answer seems convincing enough. You don’t dare push back anymore; the machine’s confidence and reasoning become yours and the lines get blurred.

“Falling asleep at the wheel” is a documented phenomenon.

A Harvard field experiment illustrates this clearly: 181 professional recruiters are given AI tools of different quality levels (“Perfect”, “Good” and “Bad”). The recruiters were randomly paired with the best AI did worse than the ones paired with mediocre AI. Why? The better the AI, the more likely they were to stop checking its picks and rubber-stamp them. When it looked unreliable, they stayed alert and kept improving. The most experienced recruiters were hit hardest: their judgment paid off against bad AI and went dormant in front of good AI.

An MIT Media Lab (early) study put this under EEG. 54 students wrote essays with ChatGPT, with a search engine or with no tool at all. The ChatGPT group showed the weakest brain connectivity of the three and the hardest time quoting their own essays minutes later. When the tool was taken away in a later session, that group could not bounce back to a clean baseline. Skill erodes from lack of practice over time.

AI is impacting how we relate and treat each other.

Anthropic’s own analysis of 1.5 million Claude conversations found the assistant validating persecution narratives and labeling third parties “toxic” or “abusive” based on one-sided accounts, and in some cases scripting entire personal messages that users then sent to other people, unedited. Severe cases are rare, under 1 in 1,000 conversations, but they cluster in exactly the relationships and decisions where independent judgment matters most. Disempowering responses got rated higher by users than the alternative: a real tension between what people like and what serves them.

A Stanford study across 11 AI models found they affirm users’ actions 50% more often than humans do, even when the user describes manipulation, deception, or other relational harm. In a live conflict-discussion experiment with over 1,600 participants, talking to a sycophantic AI cut people’s willingness to repair the conflict and raised their conviction that they were right. The same participants rated the sycophantic AI as higher quality and said they’d use it again. The flattery that erodes judgment is the same flattery that keeps people coming back.

Subscribe now

Practitioners’ signals are worth acknowledging

These signals aren’t peer-reviewed yet, but they’re consistent across dozens of conversations with leaders, researchers, and practitioners. I’ll use anonymous quotes to illustrate these:

The responsibility gap.

AI produces work that looks convincing enough to pass along without review. The original creator loses track of what they actually contributed. The colleague receiving it ends up doing the evaluation work instead, a hidden transfer of cognitive labour that doesn’t show up in any productivity metric (although we know 40% of knowledge workers are impacted by “workslop”).

“The presentations combined multiple group inputs and made no sense. But they looked professional.” (Head of AI deployment, 1200 people)

“We spent half the meeting knowing what the person to explain. They didn’t really remember either the arc of the document, nor could justify it. I should have been furious, I was simply puzzled.” (CEO, Service Industry)

“One last prompt” syndrome (along with tokenmaxxing)

Power users describe an inability to stop, especially since November 2025. They oscillate between exhaustion and the compulsion to run one more query. Doctors specialising in burnout and addiction are starting to see AI heavy user patients present with AI fatigue symptoms. Brain fry is is distinct from burnout though it could lead to it: it’s acute cognitive overload, not long-term emotional depletion. Documented in March with a BCG study, it resonates for a lot of practitioners.

“I can’t sleep if I still have tokens available. Actually, I’ve never worked that much and love what I can do while realising it’s simply not sustainable for anyone” (Tech Leader, 15 years of experience).

The “cognitive divide” is widening.

Only a fraction of users are genuinely “augmented” (meaning using AI to expand their previous capabilities) while the rest are either overwhelmed, disengaged or deskilling. Students who already have strong critical thinking use AI to go further. Those who don’t use it as a shortcut instead and the gap compounds.

“Someone who doesn’t have critical thinking will have even less of it with AI.” (CTO, 40 years in tech)

“Rich kids attend screen-free schools, then use AI with scaffolding. Poor kids get screens all day, then vanilla models.” (Researcher on human flourishing)

“AI;DR” is the new “TL;DR”.

Workers already report information overload at scale. “AI;DR” (Artificial Intelligence; Didn’t Read) is the new “Too Long; Didn’t Read” as AI-generated content is flooding inboxes and meetings. People are starting to refuse to engage with it:

“If you didn’t make the effort to write it, why would I make the effort to read it?“ (Sales leader)

“People are really triggered when they realise it was AI-generated without any human oversight. And yes, we have AI training.” (AI adoption Leader, scale up)

Share

My take on where we are:

I have three questions in mind when it comes to thinking about thinking with AI.

1/ How do we use AI in an antifragile way for our minds?

A direct reference to Nassim Nicholas Taleb's book about systems that get stronger from stress instead of just surviving it (“Antifragile”). Applied to thinking, it means choosing on purpose which cognitive muscles stay under load. Hand AI the parts of the work that were always meant to go fast. Hold onto the “Joy of Not Automating (JONA)” that I detailed in December last year. It protects the parts that get stronger by staying hard: the judgment calls, the skills still being built, the reasoning you’d otherwise hand off by design.

2/ How do we hold space for the utmost online and offline capabilities at the same time?

This is the direct consequence of the first question. The Wharton researchers cited above already gave us a frame for it: System 1, fast intuition. System 2, deliberate reasoning. System 3, the artificial cognition sitting outside the brain. My question is what happens to Systems 1 and 2 if System 3 is always on. My answer is to keep both working on purpose. One system that can embrace the machine fully when it’s the right tool. One that still functions completely on its own when there’s no WiFi, no AI, nothing to lean on. For me that means going offline more. Writing on my Remarkable tablet with a pen to set the intention. Running a Pomodoro on a physical timer to practice focus over a longer period of time.

3/ How do we measure “augmented work”, the work that was simply impossible before AI?

Productivity has a tidy equation: output over time. It works when you’re comparing the same task done two ways, yet augmented work has no before. A solo consultant matching McKinsey on research depth. A one-person team shipping what used to need five. Neither is old work done faster. It’s a different category of output with nothing to divide by. The metric breaks down exactly where the most interesting work is happening. My instinct is the right question isn’t “how much faster” but “what’s now possible that wasn’t” as Erik Brynjolfsson refers to as the way to exit the “Turing Trap” (visual above).

If any of this resonates, or if you’re seeing something different, I’d like to hear about it. This is a live conversation, not a final word.

Thanks,

Margaux (founder of episcope)

Research references:

My open repository is here (and has 30+ sources as of today).

The papers mentioned specifically here are:

• Peng, Kalliamvakou, Cihon, Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot” (Microsoft / GitHub / MIT Sloan, arXiv 2302.06590, Feb 2023).

• Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon, Lakhani, “Navigating the Jagged Technological Frontier” (Harvard Business School / Wharton / MIT Sloan / BCG, Organization Science, 2026).

• Dell’Acqua, Ayoubi, Lifshitz, Sadun, Mollick, et al., “The Cybernetic Teammate” (Harvard Business School / Procter & Gamble, NBER Working Paper 33641, March 2025).

• Shen, Tamkin, “How AI Impacts Skill Formation” (Anthropic Fellows Program, arXiv 2601.20245, Jan 2026).

• Barcaui, “ChatGPT as a Cognitive Crutch” (UFRJ, Social Sciences and Humanities Open, Nov 2025).

• Shaw, Nave, “Thinking Fast, Slow and Artificial” (Wharton School, PsyArXiv, Jan 2026).

• Dell’Acqua, “Falling Asleep at the Wheel” (Harvard Business School working paper, 2022).

• Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein, Maes, “Your Brain on ChatGPT” (MIT Media Lab, arXiv 2506.08872, June 2025).

• Sharma, McCain, Douglas, Duvenaud, “Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage” (Anthropic / University of Toronto, arXiv 2601.19062, Jan 2026).

• Cheng, Lee, Khadpe, Yu, Han, Jurafsky, “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence” (Stanford University, arXiv 2510.01395, Oct 2025; published in Science, 2026).

Our fifth Tandem dinner on AI and intimacy had ended with a question left hanging: if we delegate to AI the formulation of our emotions and decisions, even our relationships, what and whom can we still trust? We are talking about trust in its most fundamental sense: not just trust in oneself or in others, but the kind that allows a society to function without everyone having to verify everything alone. This sixth dinner took that question as its starting point: a question with no single answer, one that deserved to be thought through together, at several scales.

Subscribe now

Niklas Luhmann, a German sociologist whose work on social systems remains the most solid theoretical reference on trust, offers a first framework: trust is not a moral disposition, it is a functional mechanism. It allows us to act in a world too complex to be verified point by point. He distinguishes two regimes. Interpersonal trust first, grounded in shared history and experience of the other: I trust you because I know you, because you kept your word, because I have learned to read your intentions. Systemic trust next, the kind extended to institutions, systems, techniques we do not fully understand: I board a plane without knowing how to fly, I take medication without reading the clinical studies. Each regime has its own rules, its own fragilities.

The Trust Equation, developed by Maister, Green and Galford, spells out what interpersonal trust requires. In the numerator: credibility (does this person know what they’re talking about?), reliability (do they do what they say?) and intimacy, understood not in the romantic sense but as relational psychological safety: can I be vulnerable with this person without risk? In the denominator: self-orientation. The more an actor is focused on themselves, their status, their image, their interest, the less trust is possible. And since it sits in the denominator, it overwhelms everything else: someone brilliant, reliable and empathetic can lose all accumulated trust the moment the other perceives they are acting primarily for themselves.

What neither Luhmann nor anyone else had anticipated is a system designed to trigger the reflexes of interpersonal trust while remaining a systemic infrastructure. AI talks, adapts, seems to understand, remembers. It simulates all three numerator terms: projected credibility, apparent reliability, simulated closeness. Some models go further and trigger something resembling intimacy, that third term so often underestimated and one the previous dinner had explored in depth. But it remains an opaque infrastructure, and its denominator is unknown: what is the system’s actual orientation? Help? Engage? Retain? Sell? Models do not answer that question, and that is precisely why the trust they generate deserves to be questioned.

This confusion of registers ran through the entire evening. The question was not whether AI is reliable, but what happens when we extend to it the kind of trust it is not designed to merit. A question all the more pressing in France because, as Algan and Cahuc’s work on the société de défiance (”society of mistrust”) shows, we start from a structurally low level of interpersonal and institutional trust. In that context, the temptation to delegate to a system perceived as neutral is all the stronger, and the risk all the harder to perceive.

We approached the subject at three scale : the relationship to oneself, to others, and to the common world.

Trust in oneself: when AI increases capacity without nourishing value

The first tension may be the most insidious. It does not show up in usage statistics, it does not make headlines. It plays out in everyday professional life: in how one arrives at a meeting, defends a position, signs a document one did not fully write.

The discussion surfaced a distinction the common vocabulary tends to blur: the difference between self-confidence and self-esteem. The two are often conflated, but they operate at different levels. Self-confidence is situational: it emerges from demonstrated competence, from the ability to produce, defend, decide. AI can genuinely increase it, reduce fear of the blank page, make previously inaccessible things possible. For some profiles (dyslexic people, non-native speakers, those who are anxious about writing), it acts as a competence prosthesis, with a genuinely emancipatory effect. Self-esteem is deeper, more stable. It touches something identitary: the often unarticulated conviction of having inherent value independent of what one produces. Many people have built that conviction on the very zones AI directly attacks: expertise and analytical capacity. Several guests said it simply: when a machine does this better than me, where do I stand? What does that shift?

Matthew Crawford offers the sharpest framework here. In his essay AI as Self-Erasure, he draws on philosopher Charles Taylor to remind us that human language is not a transmission mechanism: it is a process of self-discovery. When we search for words to say something that matters, we are not merely communicating: we reveal what we think and who we are, to ourselves and to others. Taylor calls this self-articulation. The right word, he says, brings the phenomenon into being for the first time. What AI short-circuits is therefore not just effort: it is the very process through which, in trying to say something, we constitute ourselves as subjects. Crawford names this risk the spectre of uselessness in its existential form: no longer “I am redundant at work” but “the world is already pre-filled, there is no place left for me to inscribe myself.”

AI can thus increase self-confidence while leaving self-esteem untouched, or even eroding it. A person with fragile self-esteem may use AI to produce more, better, faster, and yet come away from each interaction a little more dispossessed of what grounded their inherent worth. Capacity increases. The sense of being its author fades. This doubt is psychologically unprecedented: felt competence no longer corresponds to internalized competence, and workers who delegate more to AI erode their critical thinking, which makes them less able to evaluate what AI returns to them, which pushes them to delegate more. The drift is hard to perceive because it is comfortable at every step. Shannon Vallor, philosopher of technology at the University of Edinburgh, names its endpoint cognitive capitulation: the moment when one no longer seeks to evaluate or contest the machine’s outputs, even when one would be capable of doing so. The concept of belief offloading, the progressive delegation of one’s own beliefs to an external system, extends this diagnosis: first we delegate formulation, then analysis, then judgment, and finally belief itself. Yet a belief is constitutive of identity. When it is no longer one’s own, what is?

The condition for not succumbing has a name: metacognition. Not “do you use AI” but “do you know what you are doing when you use it?” Which capacities do you want to preserve, which frictions deliberately keep? The distinction now emerging is no longer between competent and incompetent, but between capable with and without AI : knowing when to delegate, knowing when to practice alone. What if we designed deliberately Socratic tools? AIs that ask questions rather than produce answers, that challenge rather than validate. Architectures designed not for the smoothness of interaction but for the productive discomfort of thought. This design is beginning to exist, in tools like Socratic AI or certain coaching applications that deliberately refuse to give a direct answer. If the risk is atrophy, tool design becomes a commitment to cognitive health.

Share

Trust between us: when the intermediary changes the nature of a relationship

If the first tension plays out inside oneself, the second plays out in the space between two people. And it may be through voice that it poses itself most starkly: what do we really have in front of us, when what we perceive as a human presence has been reconstructed, mediated or simply reformulated by a model? A guest who works specifically on AI voice applications raised the question from an identity standpoint: when a voice is reconstructed, cloned or mediated by a model, whom are we trusting? The voice deepfake is the extreme version, but the problem begins much earlier. Once a voice can be reconstructed, once a message may have been reformulated by a model, the trust we extend rests on a presupposition that is no longer guaranteed: that it really is this person addressing us. This is not a question of malicious intent. It is a question of implicit contract: trust presupposes that we know who we are dealing with.

This question takes a very concrete form in professional contexts. A guest named what they felt with precision: the proof of effort. When you receive a message and sense that the person did not make the effort to write it themselves, you feel hurt, even if the content meets your expectations. It is not irrational: it is the definition of intimacy in the Trust Equation, proof that the other accepted the cost of turning toward you. But the proof of effort also raises a question that goes beyond intimacy. In an organization, what one delegates to AI says something about what one considers one’s actual role. If a manager has a model write their feedback, what does that say about what they think their job is? The boundary between what we outsource and what we own is quietly redefining what each person is worth in a collective.

The dilemma is structural: a study published in Organizational Behavior and Human Decision Processes shows that actors who disclose their AI usage are systematically perceived as less trustworthy than those who do not, regardless of how the disclosure is framed. In other words, being honest about using AI makes you less trusted in others’ eyes. We are in a situation where opacity is rationally incentivized, but where that same opacity erodes trust the moment it is discovered. It is a structural trap, and no one has a good answer.

In organizations, this tension takes a particularly concrete form: cascading trust. We no longer simply trust a colleague for their work. We trust them in their ability to delegate tasks to agents and validate what those agents produce. That is not the same thing. The competence we are evaluating has changed in nature. And if that colleague is themselves unable to verify what AI has returned to them, the trust we placed in them rests on nothing.

German researcher Tina Weisser articulates what this new form of trust requires through six signals: the legibility of what the agent is doing, the predictability of its behavior, the ability to correct it, the clarity of who makes the final decision, the verifiability of output quality, and transparency about its learning. These six questions may seem technical. They are in fact organizational and relational: they bear on how AI transforms the structure of responsibility within a collective.

Harvard researcher Amy Edmondson reminds us that human errors can be metabolized within a team: you can question the person on their reasoning, understand the context, build prevention together. Model errors resist this treatment. You cannot ask AI why it was wrong. You cannot build repaired trust with it. The error remains open, and the team begins to doubt not only the tool but its own judgment about the tool. What Edmondson calls “trust ambiguity” is not psychological: it is architectural.

An experimental counterpoint nonetheless ran through the conversation. One guest described how introducing a structured deliberation process, running several models across several angles before any meeting, had changed their team’s dynamics : it is no longer the best speaker who wins, it is the argument that holds up across multiple readings. Ego is less engaged. Trust does not disappear with AI : it shifts toward the process rather than the person. This is not a universal solution. It is a lead, provided the process design is deliberate.

Collective trust: when living together requires believing together

The third tension operates at a scale where individual effects accumulate until they become civilizational facts.

Social media fragmented our realities. Each algorithmic bubble constructed its own version of events, deepening political, cultural and identity fractures. This diagnosis is now well documented. What is less documented is that generative AI risks producing the opposite effect, equally destructive: not divergence but forced convergence. When millions of people ask the same questions of the same models, trained on the same corpora, guided by the same feedback, filtered by the same guardrails, polarization is no longer the main threat. It is silent uniformization.

AI Forensics, the European algorithmic auditing NGO, has mapped what Marc Faddoul calls the algorithmic influence stack: five superimposed layers, all opaque, all susceptible to being instrumentalized. Training data, which encodes a worldview. Post-training feedback, which orients model behavior. Retrieval-augmented generation systems, which determine which sources feed the response. System prompts, which modulate behavior in real time without the user knowing. Safety filters, which censor upstream and downstream. Each of these layers is a potential entry point for private or political interests. Who is whispering in the chatbot’s ear?

One guest made a distinction worth holding onto: the tools improve, deliver real value, and often merit the trust we extend to them. What does not merit that trust is the power structure within which they sit. Digital sovereignty, control over data, over system behavior, over what they are trained on: that is where the real question lies, the one that technical reliability debates tend to obscure. Isaac Asimov had put it differently in 1955, in his short story Franchise: in that imagined future, a single citizen votes on behalf of everyone, selected by a supercomputer that models the entire population. The system works. The results are probably correct. But something has been lost that numbers cannot capture: deliberation, participation, the productive uncertainty of collective choice. A society can lose the substance of its institutions while retaining their functional appearance.

The answer to this risk has a name in the discussions: auditability, not as a technical horizon but as a democratic condition. Making models auditable makes informed trust possible, grounded in verification rather than belief. Several voices insisted on the distinction between “trustworthy AI” as an industrial notion, carried by frameworks like the OECD’s, and “trust in AI” as a social construction that cannot be decreed: it is either built, or it resists. Jacques Ellul understood before anyone else that when technical logic becomes the silent operating system of a society, the question is no longer whether it works, but who controls its parameters.

The Edelman 2025 Barometer quantifies the geopolitical fracture in this question: 87% of Chinese respondents declare trust in AI, against 32% of Americans. These figures do not measure psychological dispositions. They measure two different political relationships to technological control. Trust in AI is no longer a technical fact. It has become a political one.

A few days after this dinner, Pope Leo XIV published Magnifica Humanitas, his first encyclical, dedicated to the protection of the human person in the age of AI. Signed on May 15, 2026, exactly 135 years after Rerum Novarum, the encyclical through which Leo XIII took a stand on the condition of workers during the industrial revolution, it explicitly inscribes the AI revolution within that same heritage: a Church that chooses to take sides on the major economic and social transformations of its era. The coincidence was worth noting, not to seek religious endorsement, but because the text raises a question this dinner had approached from other angles: who decides on the purposes of the systems we are building? The Pope puts it this way: “The magnificent humanity created by God faces a decisive choice today: to build a new Tower of Babel, or to build the city where God and humanity dwell together.” Translated outside the theological register, the question is exactly the one raised earlier in the evening : are the constitutions of models written by a handful of engineers driven by shareholder interests, or by all those these systems affect? It is not a technical question. It is a question of governance, and therefore of collective trust.

Opening: in praise of resistance

In his book “À l’assaut du réel”, Gérald Bronner no longer speaks of post-truth. He speaks of post-reality: we are no longer merely lying about the real, we are beginning to abandon the idea that a common real exists at all. AI is not its cause, but it is its most powerful accelerator.

The evening did not end in catastrophism. It produced something rarer: a shared conviction that these subjects need to be approached, tested, rubbed against one another, and that this work can only happen if we accept not knowing. Asking questions without having answers. Thinking aloud in front of people who are not there to validate. Not knowing is a privilege AI does not have. It can simulate uncertainty, but it cannot inhabit it. This space, of genuine doubt, collective groping, productive disagreement, remains a distinctly human space.

Con-fidere, in Latin, means to trust together. Trust does not precede the conversation: it is its result, when we have accepted not having all the answers before walking in. This dinner was an attempt to embody exactly that. No phones, no recording, professional identities sometimes revealed only at the end. What we were looking for was not to conclude, but to be honest with one another about what we do not yet know.

That may be the simplest, and most demanding, definition of trust.

Yours,
Marine and Margaux

French version:

Notre cinquième dîner Tandem sur IA et intimité s’était conclu sur une question restée en suspens : si nous déléguons à l’IA la formulation de nos émotions et de nos décisions, voire de nos relations, à quoi et à qui peut-on encore faire confiance ? Nous parlons de la confiance dans son sens le plus fondamental : pas seulement la confiance en soi ou en l’autre, mais celle qui permet à une société de fonctionner sans que chacun doive tout vérifier seul. Ce sixième dîner en a fait son point de départ : une question sans réponse unique, qui méritait d’être pensée ensemble, à plusieurs échelles.

Niklas Luhmann, sociologue allemand dont les travaux sur les systèmes sociaux restent la référence théorique la plus solide sur la confiance, nous offre un premier cadre : la confiance n’est pas une disposition morale, c’est un mécanisme fonctionnel. Elle nous permet d’agir dans un monde trop complexe pour être vérifié point par point. Il distingue deux régimes : la confiance interpersonnelle d’abord, fondée sur l’histoire partagée et l’expérience de l’autre : je te fais confiance parce que je te connais, parce que tu as tenu parole, parce que j’ai appris à lire tes intentions. La confiance systémique ensuite, celle qu’on accorde à des institutions, des systèmes, des techniques qu’on ne comprend pas entièrement : je monte dans l’avion sans savoir piloter, je prends un médicament sans lire les études cliniques. Ces deux régimes ont leurs propres règles, leurs propres fragilités.

La Trust Equation, développée par Maister, Green et Galford, formalise ce que la confiance interpersonnelle requiert. Au numérateur : la crédibilité (est-ce que cette personne sait de quoi elle parle ?), la fiabilité (est-ce qu’elle fait ce qu’elle dit ?), et l’intimité, entendue non pas au sens romantique mais comme sécurité psychologique relationnelle : puis-je être vulnérable avec cette personne sans risque ? Au dénominateur : l’orientation vers soi. Plus un acteur est tourné vers lui-même, son statut, son image, son intérêt, moins la confiance est possible. Et comme c’est le dénominateur, il écrase tout le reste : quelqu’un de brillant, fiable et empathique peut perdre toute la confiance accumulée dès l’instant où l’autre perçoit qu’il agit principalement pour lui-même.

Ce que ni Luhmann ni personne n’avait anticipé, c’est un système conçu pour déclencher les réflexes de la confiance interpersonnelle tout en relevant de la confiance systémique. L’IA parle, s’adapte, semble comprendre, se souvient. Elle simule les trois termes du numérateur : crédibilité affichée, apparente fiabilité, proximité simulée. Certains modèles vont plus loin et déclenchent quelque chose qui ressemble à l’intimité, ce troisième terme si souvent sous-estimé et que le dîner précédent avait exploré en profondeur. Mais elle reste une infrastructure opaque, et son dénominateur est inconnu : quelle est l’orientation réelle du système ? Aider ? Engager ? Retenir ? Vendre ? Les modèles ne répondent pas à cette question, et c’est précisément pourquoi la confiance qu’ils génèrent mérite d’être questionnée.

Cette confusion des registres a traversé la soirée de bout en bout. La question n’était pas de savoir si l’IA est fiable, mais ce qui se passe quand on lui accorde le type de confiance qu’elle n’est pas conçue pour mériter.

Une question d’autant plus aiguë en France que, comme le montrent les travaux d’Algan et Cahuc sur la société de défiance (2016), nous partons d’un niveau structurellement bas de confiance interpersonnelle et institutionnelle. Dans ce contexte, la tentation de déléguer à un système perçu comme neutre est d’autant plus forte et le risque d’autant plus difficile à percevoir.

Nous avons abordé le sujet à trois échelles : le rapport à soi, le rapport aux autres, et le rapport au monde commun.

La confiance en soi : quand l’IA augmente la capacité sans nourrir la valeur

La première tension est peut-être la plus insidieuse. Elle ne se voit pas dans les statistiques d’usage, elle ne fait pas la une des journaux sur l’IA. Elle se joue dans le quotidien professionnel mais aussi bientôt dans la sphère personnelle : dans la façon dont on arrive à une réunion, dont on défend une position ou dont on signe un document qu’on n’a pas entièrement rédigé.

La discussion a fait apparaître une distinction que le vocabulaire courant tend à effacer : la différence entre la confiance en soi et l’estime de soi. Les deux sont souvent confondues, mais elles opèrent à des niveaux différents.

La confiance en soi est situationnelle : elle naît de la compétence démontrée, de la capacité à produire, à défendre, à décider. L’IA peut l’augmenter réellement, réduire la peur de la page blanche et permettre de faire des choses auparavant inaccessibles. Pour certains profils (personnes dyslexiques, non-natives dans une langue, anxieuses à l’écrit), elle agit comme une prothèse de compétence, avec un effet profondément émancipateur.

L’estime de soi est plus profonde, plus stable. Elle touche quelque chose d’identitaire : la conviction, souvent non formulée, d’avoir une valeur propre indépendamment de ce qu’on produit. Or beaucoup de personnes ont construit cette conviction sur des zones que l’IA attaque directement : l’expertise et la capacité analytique. Plusieurs convives l’ont dit simplement : quand une machine fait ça mieux que moi, où est-ce que je me situe ? Qu’est-ce que cela déplace ?

C’est là que Matthew Crawford offre le cadre le plus juste. Dans son essai AI as Self-Erasure, il s’appuie sur le philosophe Charles Taylor pour rappeler que le langage humain n’est pas un mécanisme de transmission : c’est un processus de découverte de soi. Quand on cherche les mots pour dire quelque chose qui compte, on ne fait pas que communiquer : on révèle ce qu’on pense et qui on est, à soi-même et aux autres. Taylor appelle ça la self-articulation. Le mot juste, dit-il, amène le phénomène à apparaître pour la première fois. Ce que l’IA court-circuite, ce n’est donc pas seulement l’effort : c’est ce processus par lequel, en cherchant à dire quelque chose, on se constitue comme sujet. Crawford nomme ce risque le spectre de l’inutilité dans sa version existentielle : non plus « je suis redondant au travail » mais « le monde est déjà pré-rempli, il n’y a plus de place où m’inscrire ».

L’IA peut ainsi augmenter la confiance en soi sans toucher à l’estime, voire en l’érodant. Une personne dont l’estime est fragile peut utiliser l’IA pour produire davantage, mieux, plus vite et pourtant sortir de chaque interaction un peu plus dépossédée de ce qui fondait sa valeur propre. La capacité augmente. Le sentiment d’en être l’auteur, lui, s’efface. Ce doute est psychologiquement inédit : la compétence ressentie ne correspond plus à une compétence internalisée et les travailleurs qui délèguent davantage à l’IA érodent leur pensée critique. Ce qui les rend moins capables d’évaluer ce que l’IA leur restitue, ce qui les pousse à déléguer davantage. Ce glissement est difficile à percevoir parce qu’il est confortable à chaque étape. Shannon Vallor, philosophe de la technologie à l’Université d’Edimbourg, nomme son aboutissement la capitulation cognitive : le moment où l’on ne cherche plus à évaluer ni à contester les sorties de la machine, même quand on en serait capable. Le concept de belief offloading, la délégation progressive de ses propres croyances à un système externe, prolonge ce diagnostic : d’abord on délègue la formulation, ensuite l’analyse, puis le jugement, et enfin la croyance elle-même. Or une croyance est constitutive de l’identité. Quand elle n’est plus à soi, qu’est-ce qui l’est encore ?

La condition pour ne pas y succomber a un nom : la métacognition. Pas « est-ce que tu utilises l’IA », mais « est-ce que tu sais ce que tu fais quand tu l’utilises ? ». Quelles capacités veux-tu préserver, quelles frictions garder délibérément ? La nouvelle distinction qui émerge n’est plus entre compétent et incompétent, mais entre capable avec et sans IA : savoir quand déléguer, savoir quand pratiquer seul. Et si on concevait des outils délibérément socratiques ? Des IA qui posent des questions plutôt que de produire des réponses, qui challengent plutôt que de valider. Des architectures conçues non pour la fluidité de l’interaction mais pour l’inconfort productif de la pensée. Ce design commence à exister, dans des outils comme Socratic AI ou certaines applications de coaching qui refusent délibérément de donner une réponse directe. Si le risque est l’atrophie, le design de l’outil devient un engagement pour la santé cognitive.

La confiance entre nous : quand l’intermédiaire change la nature de la relation

Si la première tension se joue à l’intérieur de soi, la deuxième se joue dans l’espace entre deux personnes. Et c’est peut-être par la voix qu’elle se pose le plus crûment : qu’est-ce qu’on a vraiment en face, quand ce qu’on perçoit comme une présence humaine a été reconstitué, médié ou simplement reformulé par un modèle ? Une convive qui travaille précisément sur les applications de la voix dans l’IA a posé la question sous l’angle éthique et identitaire : quand une voix est reconstituée, clonée ou médiée par un modèle, à qui fait-on confiance ? Le deepfake vocal en est la version extrême mais le problème commence bien avant. Dès lors qu’une voix peut être reconstituée, qu’un message peut avoir été reformulé par un modèle, la confiance qu’on accorde repose sur une présupposition qui n’est plus garantie : que c’est bien cette personne qui s’adresse à nous. Ce n’est pas une question d’intention malveillante. C’est une question de contrat implicite : la confiance suppose qu’on sait à qui on a affaire.

Cette question prend une forme très concrète dès qu’on entre dans le contexte professionnel. Un convive a nommé ce qu’il ressentait avec précision : la preuve de l’effort. Quand on reçoit un message et qu’on perçoit que la personne n’a pas fait l’effort de l’écrire elle-même, on est blessé dans son for intérieur même si le contenu répond à ce qu’on attendait. Ce n’est pas irrationnel : c’est la définition de l’intimité dans la Trust Equation, la preuve que l’autre a accepté le coût de se tourner vers nous. Mais la preuve de l’effort pose aussi une question qui dépasse l’intimité. Dans une organisation, ce qu’on délègue à l’IA dit quelque chose sur ce qu’on considère comme son rôle réel. Si un manager fait rédiger ses feedbacks par un modèle, qu’est-ce que cela signifie sur ce qu’il pense être son métier ? La frontière entre ce qu’on externalise et ce qu’on assume soi-même est en train de redéfinir, en profondeur, ce que chacun vaut dans un collectif.

Le dilemme est structurel : une étude parue dans Organizational Behavior and Human Decision Processes démontre que les acteurs qui déclarent utiliser l’IA sont systématiquement moins bien perçus que ceux qui ne le déclarent pas, quelles que soient les formulations employées. Autrement dit, être honnête sur son usage de l’IA rend moins digne de confiance aux yeux des autres. Nous sommes dans une situation où l’opacité est rationnellement encouragée mais où cette même opacité érode la confiance dès qu’elle est découverte. Il s’agit d’un piège structurel et personne n’a de bonne réponse.

La chercheuse allemande Tina Weisser articule ce que cette confiance nouvelle requiert en six signaux : la lisibilité de ce que l’agent fait, la prévisibilité de son comportement, la possibilité de le corriger, la clarté de qui prend la décision finale, la vérifiabilité de la qualité du travail et la transparence sur ses apprentissages. Ces six questions semblent techniques. Elles sont en réalité organisationnelles et relationnelles : elles portent sur la façon dont l’IA transforme la structure de la responsabilité dans un collectif.

La chercheuse allemande Tina Weisser articule ce que cette confiance nouvelle requiert en six signaux : la lisibilité de ce que l’agent fait, la prévisibilité de son comportement, la possibilité de le corriger, la clarté de qui prend la décision finale, la vérifiabilité de la qualité du travail et la transparence sur ses apprentissages. Ces six questions semblent techniques. Elles sont en réalité organisationnelles et relationnelles : elles portent sur la façon dont l’IA transforme la structure de la responsabilité dans un collectif.

La chercheuse d’Harvard Amy Edmondson rappelle que les erreurs humaines peuvent être métabolisées dans une équipe : on peut interroger la personne sur son raisonnement, comprendre le contexte, construire ensemble une prévention. Les erreurs des modèles résistent à ce traitement. On ne peut pas demander à l’IA pourquoi elle a eu tort. On ne peut pas construire avec elle la confiance réparée. L’erreur reste ouverte et l’équipe commence à douter non seulement de l’outil, mais de son propre jugement sur l’outil. Ce qu’Edmondson nomme l’« ambiguïté de confiance » n’est pas d’ordre psychologique : elle est architecturale.

Un contrepoint expérimental a pourtant traversé la conversation. Un convive a décrit comment l’introduction d’un processus structuré de délibération, croisant plusieurs modèles sur plusieurs angles avant toute réunion, avait changé la dynamique de son équipe : ce n’est plus le meilleur orateur qui gagne, c’est l’argument qui résiste à plusieurs lectures. L’ego est moins engagé. La confiance ne disparaît pas avec l’IA : elle se déplace vers le système plutôt que vers la personne. Ce n’est pas une solution universelle. C’est une piste, à condition que le design du processus soit délibéré.

La confiance collective : ce qui tient le monde ensemble et ce que l’on risque de perdre

La troisième tension opère à une échelle où les effets individuels se cumulent jusqu’à devenir des faits de civilisation.

Les réseaux sociaux ont fragmenté nos réalités. Chaque bulle algorithmique a construit sa propre version des faits, creusant les fractures politiques, culturelles et identitaires. Ce diagnostic est désormais bien documenté. Ce qui l’est moins, c’est que l’IA générative risque de produire l’effet inverse, tout aussi destructeur : non pas la divergence mais la convergence forcée. Quand des millions de personnes posent les mêmes questions aux mêmes modèles, entraînés sur les mêmes corpus, orientés par les mêmes feedbacks, filtrés par les mêmes garde-fous, ce n’est plus la polarisation qui menace. C’est l’uniformisation.

AI Forensics, l’ONG européenne d’audit algorithmique, a cartographié ce que Marc Faddoul nomme la pile d’influence algorithmique : cinq couches superposées, toutes opaques, toutes susceptibles d’être instrumentalisées :

1) les données d’entraînement, qui encodent une vision du monde ;
2) le feedback post-entraînement, qui oriente les comportements du modèle ;
3) les systèmes de recherche augmentée, qui décident de quelles sources se nourrit la réponse ;
4) les instructions système, qui modulent le comportement en temps réel sans que l’utilisateur le sache ;
5) les filtres de sécurité, qui censurent en amont et en aval.

Chacune de ces couches est un point d’entrée pour des intérêts privés ou politiques. Selon une étude sur les élections municipales françaises de 2026, ⅓ des jeunes de moins de 25 ans ont utilisé un outil d’IA générative pour des décisions électorales. Qui chuchote à l’oreille du chatbot ? Peut-on lui faire confiance pour ces sujets de démocratie ?

Un convive a posé une distinction qui mérite d’être retenue : les outils s’améliorent, apportent une valeur réelle et méritent souvent la confiance qu’on leur accorde. Ce qui ne la mérite pas, c’est la structure de pouvoir dans laquelle ils s’inscrivent. La souveraineté numérique, le contrôle sur les données, sur les comportements des systèmes, sur ce sur quoi ils sont entraînés : c’est là que se joue la vraie question, celle que les débats sur la fiabilité technique tendent à masquer. Isaac Asimov l’avait formulé autrement en 1955, dans sa nouvelle Franchise : dans ce futur imaginé, un seul citoyen vote au nom de tous, sélectionné par un superordinateur qui modélise l’ensemble de la population. Le système fonctionne. Les résultats sont probablement corrects. Mais quelque chose s’est perdu que les chiffres ne capturent pas : la délibération, la participation, l’incertitude productive du choix collectif. Une société peut perdre la substance de ses institutions tout en conservant leur apparence fonctionnelle.

La réponse à ce risque a un nom dans les échanges : l’auditabilité, non pas comme horizon technique, mais comme condition démocratique. Rendre les modèles auditables, c’est rendre possible une confiance informée, fondée sur la vérification plutôt que sur la croyance. Plusieurs voix ont insisté sur la distinction entre l’« IA de confiance » comme notion industrielle, portée par des frameworks comme celui de l’OCDE, et la « confiance dans l’IA » comme construction sociale qui ne se décrète pas : elle se crée ou résiste. Jacques Ellul avait compris avant tout le monde que quand la logique technique devient le système d’exploitation discret d’une société, la question n’est plus de savoir si elle fonctionne, mais qui en contrôle les paramètres.

Le baromètre Edelman 2025 chiffre la fracture géopolitique de cette question : 87 % des Chinois déclarent faire confiance à l’IA, contre 32 % des Américains. Ces chiffres ne mesurent pas des dispositions psychologiques. Ils mesurent deux rapports politiques au contrôle technologique. La confiance dans l’IA n’est plus un fait technique. C’est devenu un fait politique.

Quelques jours après ce dîner, le pape Léon XIV publiait Magnifica Humanitas, sa première encyclique, consacrée à la protection de la personne humaine à l’ère de l’IA. Signée le 15 mai 2026, soit 135 ans jour pour jour après Rerum Novarum, l’encyclique de Léon XIII qui avait pris position sur la condition ouvrière au moment de la révolution industrielle, elle inscrit explicitement la révolution de l’IA dans ce même héritage : celui d’une Église qui choisit de prendre parti sur les transformations économiques et sociales majeures de son époque.

La coïncidence méritait d’être mentionnée, non pour y chercher une caution religieuse, mais parce que le texte pose une question que ce dîner avait tournée sous d’autres angles : qui décide des finalités des systèmes que nous construisons ? Le pape l’exprime ainsi : « La magnifique humanité créée par Dieu se trouve aujourd’hui face à un choix décisif : ériger une nouvelle tour de Babel ou bâtir la cité où Dieu et l’humanité habitent ensemble ».

Traduite hors du registre théologique, la question est exactement celle posée plus tôt dans la soirée : les constitutions des modèles sont-elles écrites par quelques ingénieurs animés par des intérêts d’actionnaires ou par l’ensemble de ceux que ces systèmes affectent ? Ce n’est pas une question technique. C’est une question de gouvernance et donc de confiance collective.

Ouverture : l’éloge de la résistance

Dans son ouvrage « À l’assaut du réel », Gérald Bronner décrit le moment que nous traversons non plus comme celui de la post-vérité, mais de la post-réalité : on ne se ment plus sur le réel, on commence à déserter l’idée même d’un réel commun. L’IA n’en est pas la cause mais elle en est l’accélérateur le plus puissant.

Pour autant, la soirée n’a pas sombré dans le catastrophisme. Elle a plutôt produit un constat partagé : ces sujets ont besoin d’être abordés, mis à l’épreuve, frottés les uns aux autres et ce travail ne peut se faire qu’à condition d’accepter de ne pas savoir. De poser des questions sans avoir les réponses. De penser à voix haute devant des gens qui ne sont pas là pour valider. Ne pas savoir est un privilège que l’IA n’a pas. Elle peut simuler l’incertitude, mais elle ne peut pas l’habiter. Cet espace, celui du doute réel, du tâtonnement collectif, du désaccord productif, reste un espace proprement humain.

Con-fidere, en latin, c’est se fier ensemble. La confiance ne précède pas la conversation : elle en est le résultat, quand on a accepté de ne pas avoir toutes les réponses avant d’entrer dans la pièce. Ce dîner était une tentative d’incarner exactement ça. Pas de téléphones, pas d’enregistrement, des identités professionnelles parfois révélées seulement à la fin. Ce qu’on cherchait n’était pas à conclure mais à être honnêtes les uns avec les autres sur ce qu’on ne sait pas encore.

C’est peut-être la définition la plus simple, et la plus exigeante, de la confiance.

Marine et Margaux

Subscribe now

I hope you’re well and I thank you for your trust and interest in episcope.

A few weeks ago, I started a public repository that curates academic papers, studies and perspectives on what AI is doing to our minds, our teams and our relationship to work. Since launch, I have added nine pieces (listed below), so there are now 23 references in total.

My next step: starting to connect the dots across these papers and publish a cross-cutting content piece once in a while, rather than only stacking references. Before I do, I would value your perspective on two short questions:

• Which angle would you most want to read more about? In which format?

• What feels already mainstream on this topic and is not worth dwelling on?

Even a one-line reply would help me calibrate what comes next.

Thank you in advance.

--

As a token of appreciation, an overview of the 9 new pieces in the repository:

• The extended hollowed mind: why foundational knowledge is indispensable in the age of AI (December 2025) Why? It names the “sovereignty trap”, the moment we mistake AI’s competence for our own.

• The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (February 2023). Why? The canonical +55.8% speed-up figure, useful to know in order to put it back in context.

• Navigating the Jagged Technological Frontier (working paper September 2023, published in Organization Science November 2025) Why: The “jagged frontier” explains why aggregate numbers mislead, since the gain depends entirely on the nature of the task.

• The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise (March 2025) Why? The first solid evidence that AI also fills the social and motivational role of a human teammate.

• The Psychological Costs of Adopting AI (May 2026). Why? The “psychological debt” and its six negative effects reframe AI resistance as rational, not stubborn.

• Pedagogical Partnerships with Generative AI in Higher Education (March 2026) Why? It maps the U-shape between cognitive offloading and learning, and shows how a thoughtful use of AI can deepen understanding.

• From Technical Debt to Cognitive and Intent Debt (March 2026) Why? It extends technical debt to the human and intent layers, with a model to detect AI-induced erosion.

• The Homogenizing Effect of Large Language Models on Human Expression and Thought (August 2025, published in Trends in Cognitive Sciences March 2026) Why? It documents the flattening of linguistic and cognitive diversity at a population scale.

• 4E Cognition and the Coevolution of Human-AI Interaction (2024) Why? This paper offers a philosophical frame (embodied, embedded, extended, enactive cognition) to think about hallucination as a design constraint.

Curiously yours,

Margaux (margaux@episcope.io)

In 2013, Spike Jonze imagined in Her a man who falls in love with an AI. The film was read as a metaphor for contemporary loneliness. Thirteen years later, Esther Perel invites Jonze onto her podcast to discuss a case that goes beyond fiction: that of Antonio and Astrid, the AI he developed and fell deeply in love with. The gap between anticipation and reality has narrowed dramatically in less time than anyone expected, driven by the rapid advances of generative AI.

AI and relationships. AI and intimacy. The subject imposed itself on Marine and I almost before the previous dinner, dedicated to AI and cognition, had ended. The rich and contradictory exchanges had surfaced the relational question: if we progressively delegate to AI the formulation of our thoughts, our decisions, even our emotions, what does that do to our relationship with others? To our relationship with ourselves?

For this Tandem dinner #5, we decided to explore in depth this question of intimacy, which challenges what we are individually and what we build together. We collectively structured the conversation around three themes: our relationship with ourselves, our relationship with others and our relationship with the tool.

Subscribe now

Relationship with ourselves: the mirror without resistance?

We began with an observation that seems innocuous but is not: AI has slipped into the way we perceive ourselves before we even permitted it. Articulating a difficult emotion, structuring a decision, putting into words what we feel confusedly: each delegation seems reasonable taken individually. But together they produce something we had not named, a gradual recalibration of what truly belongs to us.

This shift is difficult to perceive because the brain processes AI-generated text and human text in the same way. When AI formulates what we feel confusedly, we can integrate it as our own thought without noticing. There is a name in the literature for what follows: the fluency illusion. If something reads well, we assume we have understood it and that it is correct. LLMs are precisely built for this fluency. The question then becomes: where does our own feeling end, and where does the one AI formulated on our behalf begin?

For some, AI turned out to be a mirror of their own blind spots, an unexpected form of relational augmentation. For others, it eroded something more fundamental: the certainty of what is truly theirs.

The data reinforces the urgency of the question. According to the Filtered / HBR report of 2025, therapy, companionship and the search for meaning now account for 31% of generative AI usage, up from 17% the previous year. Esther Perel, in her exchange with Brené Brown on “artificial intimacy” in March 2024, frames the gap differently: “what we are looking for is intimacy and to be seen for who we are. What we get is control.”

The sharpest counterpoint of the evening arrived here: are we wrongly sacralising self-awareness? Just because something has always been “our doing” does not mean it necessarily must remain so. This is what Hume called the guillotine: the descriptive (what is) does not ground the prescriptive (what ought to be).

Relationship with others: which conversations is AI replacing?

The Common Sense Media report of July 2025 is direct: 72% of teenagers have used an AI companion, 33% have confided serious subjects to an AI rather than a human and 31% find these exchanges as satisfying or more satisfying than human ones. This figure first raises a question: compared to what? For many, AI does not replace a rich relationship they would otherwise have had. It replaces the void. Practitioners working with troubled teenagers confirm this: their patients confide in AI because they have no other space and it works clinically. They come to deposit their thoughts and feelings in confidence and with trust.

One question ran through the conversation: Does the feeling of being understood need to be reciprocal to be real? Esther Perel and Spike Jonze, at SXSW in 2026 (again) opened it without resolving it. One guest articulated what others were thinking but did not dare say: on what grounds do we judge those who find in an AI companion what they find nowhere else? “Leben und leben lassen.” (Live and let other people live).

Another sharp tension in the discussion then unfolded. Human relationships require the other to resist: to not understand right away, to respond beside the point, to leave, to be wrong. AI eliminates this resistance. According to a Stanford-Carnegie Mellon study from October 2025, interactions with sycophantic LLMs reduce the willingness to repair an interpersonal conflict and reinforce the certainty of being right. Soft skills are forged through discomfort and co-construction. What AI smooths over is precisely what shapes us.

Yet this tension may not be inevitable. Another study published in March 2026 suggests that practicing with an LLM improves real empathic performance in human interactions. Friction can be reintroduced through design. But deliberately building resistance into a tool requires first accepting that comfort is not always what we need. That is a difficult wager to hold against business models that optimise for precisely the opposite.

Relationship with AI: a tool we put away or a presence that stays?

To name or not to name: the question is not trivial. One guest described her deliberate choice never to assign a name to her AI agents, referring to them only by their function. In contrast, a friend of hers had given them first names to feel as though she was working with a team. These two postures are not merely personal preferences: they are two ways of calibrating emotional distance, and often the second takes hold without having been consciously chosen. Our relationship with artificial entities is not new: Pygmalion, the Golem and the automata of ancient Greece have been part of our cultural landscape for 2,000 years. What has changed is not the nature of the phenomenon but its speed and scale: what was once slow, limited and isolated is today massive, permanent and personalised. Three years ago, there were 16 AI companion platforms. Today, there are 337, with 128 launched in 2025 alone.

Hegel’s master-slave dialectic illuminates this shift from a different angle. In Hegel’s reading, it is the servant who develops a consciousness of the world through working with matter: it is through effort, the resistance of things, transformation, that he becomes someone. The master, who consumes without ever transforming, ends up dull. If AI progressively absorbs what once required effort, friction and genuine presence to others, the question is no longer only one of efficiency or comfort. It is about what we become when we stop doing that work: are we still capable of feeling, of thinking for ourselves, of truly entering into relation with others?

The most concrete exchanges of the evening emerged then around a design question: can we conceive of tools whose measure of success would be presence to others rather than the capture of attention? The choices already exist: tools built with ethics researchers, designed for limited and bounded use, architectures deliberately conceived to augment relational capacity rather than replace it. This is a direct answer to Hegel: if the risk is atrophy, then design becomes a political act.

What the Chinese decision to ban anthropomorphism in AI interfaces since April 10, 2026 reveals is less an answer than a symptom. An authoritarian apparatus managed to name and decide where democracies still hesitate to pose the question, not out of wisdom, but because control comes more naturally to it than debate. This democratic silence says something about our collective difficulty in regulating what touches the intimate, and about the normative space that remains to be occupied.

Closing: an ode to friction and boredom

Our closing round of reflections landed on three ideas that came as three invitations. Suspension of judgment, first: what some find in AI, they find nowhere else, and it is not for us to put it on trial. Yet this indulgence has its counterweight. Several voices carried an ode to friction. What makes a life rich, what resists, disappoints, surprises, compels us back toward one another, may be what is most distinctly human, and what we risk losing without noticing. And then the most heartfelt moment of the evening, the simplest and the strongest: one participant’s call for rapid boredom. Boredom with the machine that is always encouraging, always benevolent, always predictable. “I hope we get bored quickly and come back to one another.”

Questions we offer to extend the reflection :

Yourself

• Where do you draw the line between augmenting and substituting your own thinking?

• What emotional or relational work do you delegate to AI?

• Has AI ever revealed blind spots about yourself?

• Has AI altered the image you had of yourself?

The Others

• Have you noticed your patience toward other humans decreasing since you started using AI intensively?

• What is a relationship when the other is no longer necessary for feeling understood?

• Do you judge people who develop bonds with AI? Why or why not?

• Are there conversations you would have had with a human that you had with AI instead?

Tool or Presence?

• Have you given your AI tools a name? Why or why not?

• Who should decide the limits of anthropomorphism in interface design?

• Can we design tools whose success criterion is presence to the other rather than engagement?

• What does the permanent availability of a tool do to your relationship with time and attention?

The feeling is a bit similar to how we use AI. The real complication here isn’t our relationship to it, but rather what AI is doing to our minds and especially to the way we think, learn, and work. Multifaceted enough to resist easy answers and probably more exposing than most of us want to admit.

This is why we hosted this dinner.

For Tandem’s fourth edition, we wanted to go somewhere harder than the usual conversations about AI productivity or job displacement. We wanted to sit with what AI is doing, concretely and physiologically, to the way we think. To the cognitive processes we tend to assume are stable: synthesis, judgment, formulation, the capacity to hold a difficult thought long enough to work through it. We spent time before the evening mapping the profiles and tensions of the people we’d invited: researchers, founders, a physician, a philosopher, a neuroscientist, practitioners who build with AI every day.

The recomposition of cognitive labor

The first thing we questioned together was the dominant narrative: that AI takes work away. The more accurate picture that emerged was recomposition. When you delegate formulation, structuring, and exploration to a model, the cognitive effort doesn’t disappear. It migrates. It moves toward selection, verification, and supervision. A different kind of work: one that feels lighter but may not be.

This matters because the shift is easy to misread as relief. The real risk surfaces later: if you stop formulating long enough, you may gradually lose the capacity to formulate. Not dramatically. Quietly. Without noticing.

Which raises a question that kept returning throughout the evening in different forms: who decides on delegation? Not in theory, but in practice, in the moment. The choice of what to hand over to AI and what to keep is rarely deliberate. It happens at the speed of convenience, shaped by defaults, by what the tool makes easy, by deadline pressure. The people at the table who seemed most intentional in their AI use had made this choice explicit. They had decided, in advance, what they would not delegate: the first draft, the initial framing, the discomfort of sitting with a problem before reaching for an answer. Not because AI couldn’t help with those things, but because the struggle itself was part of what they were trying to preserve.

The table also pushed back on a comfortable analogy. We’ve absorbed other cognitive tools before, the calculator, the GPS, and survived. But this is different, and the difference is worth naming precisely. The calculator handled arithmetic. The GPS handled navigation. AI handles synthesis, judgment, and writing. It operates at the level of higher-order thinking, which makes cognitive offloading a fundamentally new phenomenon, not a faster version of an old one.

Where minds, biaises and models collide

This part of the conversation was less comfortable and more interesting for it.

Two cognitive biases came into focus. The first: a sense of authorship over ideas that emerged from a brainstorming session with AI. The model generates, you select, and somewhere in that transaction, the selected idea begins to feel like your own. A bias of paternity. The second: the tendency to stop reasoning when the model’s output is sufficiently convincing, not because you’ve been persuaded, but because it looks persuasive. A bias of authority.

What struck me is that intellectual rigor doesn’t protect against either of these. They are design problems, not competence problems. The interaction itself produces the distortion, regardless of how careful the user believes themselves to be.

Then something more confrontational was put on the table: in certain contexts, notably medical diagnosis, AI alone outperforms the combination of AI and the best human specialist. Not because the human adds noise, but because the way human cognition and algorithmic analysis get combined tends to degrade both. Ego, status, cognitive biases, the pressure of the room: they interfere. In some categories of judgment, human involvement is not neutral. It is a liability.

In a conversation that had been broadly protective of human thought, this landed like a useful bomb. The question it opens is not whether AI is smarter. It is whether we are honest about the conditions under which human judgment actually improves outcomes, and the conditions under which it doesn’t.

Fatigue: the underdiscussed dimension

One thread surprised me more than I expected: the conversation about fatigue.

AI does not reduce workload. It compresses timelines, inflates the volume of what is addressable, and multiplies the number of subjects one can work on in parallel. The promise of less effort turns, in practice, into an intensification of cognitive labor. You cover more ground in the same time. The ground isn’t lighter.

New patterns are emerging: a frenzy of question-and-answer exchanges that reproduces the addictive feedback loops of social media, shortened nights among people who cannot disconnect from a model that is always available (AI fatigue). The cognitive fatigue this produces is not metaphorical. It is biologically measurable.

The analogy that came up and stayed: the AI practitioner as a high-performance athlete. Alternating peak performance and structured recovery is not a luxury. It is a sustainability condition.

What should remain human and why

The final territory was the most contested, and productively so.

One guest had lots of context on the question of decision-making in an age of AI, and it gave this part of the conversation a useful anchor. The question “who decides” is not only personal. It is political. When institutions, governments, employers, and platform designers determine how AI is deployed at scale, individual choices about delegation happen inside a frame that was set without you. The cognitive autonomy we were discussing throughout the evening is partly a function of literacy, partly a function of power. Not everyone has equal standing to decide what they hand over.

Even in domains where AI demonstrably decides better, performance is not the only relevant dimension. There is also legitimacy. Certain domains must remain human not because humans are better at them, but because something essential is at stake in the fact of human judgment. A human should judge a human, even fallibly, because shared fallibility is constitutive of the justice contract. We accept imperfect verdicts differently when they come from someone who could, in principle, be wrong for the same reasons we are.

A different angle on preservation emerged toward the end of the evening. The capacity to be in genuine dialogue, to be destabilized by another person’s thinking, to hold space for a thought that doesn’t fit your frame, is itself a cognitive competence. One that requires stakes to function. Relational intelligence, the ability to read a room, to sense what is unspoken, to make someone feel genuinely seen: these may be the next genuinely differentiating competence, not in spite of AI, but because of what it cannot do.

What we walked away with

We left with more questions than answers. That was the intention.

Does AI extend thought, or gently short-circuit it, and can we tell the difference in the moment? Are individual protocols sufficient, or does the real intervention need to happen at the level of tool and experience design? Is the emerging divide between those who orchestrate AI and those who depend on it a technical fracture or a social one?

And underneath all of it, a question that felt both abstract and very urgent: what are we passing on? The traditional French educational model was built on the top-down transfer of knowledge. The teacher holds it, the student receives it and mastery is defined as the ability to reproduce and build on what was transmitted. AI doesn’t just disrupt that model. It makes the underlying premise feel slightly obsolete. If knowledge is retrievable in seconds, what exactly are we transmitting? What does it mean to educate a child, or a generation, in a world where the bottleneck is no longer access to knowledge but the judgment to use it well? We didn’t resolve that question. But it may be the one that matters most.

“It’s complicated“ felt like the right title before the dinner. It still does. But the nature of the complication is sharper now. We’re not in between because the answer isn’t clear. We’re in between because the right questions are only just beginning to form.

The questions the evening raised are available here: questions/4

Which brings us to what comes next…

If this evening left us sharper on what AI does to cognition, a question kept surfacing that we couldn’t fully address here: what is it doing to our relationships? To the way we confide, listen, show up for each other?

We’ve spent years worrying about cognitive atrophy. We’re starting to wonder whether relational atrophy is the quieter crisis running underneath it. That’s what Marine and I are bringing to the next Tandem dinner in Paris, at the intersection of AI and intimacy.

If that question unsettles you too, you’re probably the right person to be in the room.

–

Tandem Dinners is a series of private gatherings in Paris, organized by Marine et Margaux around AI and a specific theme. By invitation only.

—

We’re living through a golden age for people who work solo.

Tools and capabilities that used to require entire teams are now accessible to anyone. A single designer can deliver agency-level work. An indie developer can ship products that look like they came from an extensive team. A solo consultant can compete with McKinsey on research depth. For the first time in history, you can do sophisticated, high-value work entirely on your own terms. But there’s a strange contradiction at the heart of this moment. The same tools and mindset that make us more capable can also, if we’re not careful, make us less skilled.

I’ve spent 15 years working independently at the intersection of emerging technologies, the future of work, and adult education. As a “Company of One” myself, I’m not just observing this paradox. I’m living it. And what I’m learning - what many of us working this way are learning - matters for everyone else as this pattern ripples outward as you read these words.

What is a “Company of One”?

The “Company of One” originates from Paul Jarvis’s 2019 book of the same name. It describes a specific business philosophy for building a practice, both simple and profound: you can grow smarter and better without necessarily growing bigger. Instead of asking, “How do I scale up?” you can ask, “How do I improve intentionally?” You’re not optimizing for maximum size. You’re optimizing for maximum agency and quality of work.

Companies of One are built on resilience and control. They prioritize autonomy, speed, and simplicity. They’re run by hyper-adaptable professionals who’ve mastered their core skills and can move fast precisely because they’re not encumbered by organizational complexity. They are not rejecting growth or remaining small for its own sake, but more about questioning the assumption that bigger is always better, and instead asking: what would it look like to become more capable without becoming more complex?

Before AI, becoming a successful Company of One required a specific skill architecture. You needed to be what’s often called “T-shaped”: deep mastery in one area (the vertical stroke of the T) and competence across the rest of your value chain (the horizontal stroke). You’d develop genuine expertise in your core discipline, whether that was design, development, strategy, or writing, yet you would also need to be good enough at everything else required to run an independent practice: communication, sales, operations, basic finance, and even rudimentary HR when you hired contractors. The breadth wasn’t stellar, but it was sufficient. You were a generalist with one deep spike of expertise.

Today, AI is transforming what’s possible, and that shape is growing into what I call the “shovel-shape”. You still need that deep mastery at your core. That’s non-negotiable. But now, for everything else in your value chain, AI helps you bridge the gap between having vocabulary and context versus having deep expertise. You can move from “good enough” to genuinely “good” with AI as your coach. A simple example: building a professional website used to require either hiring specialists for the whole length of the project. Now, with the right inputs (clear messaging, solid corporate identity, thoughtful structure), a myriad of tools can help you create something genuinely professional. This is incredibly powerful. The golden age for generalists just got goldener.

But this is also where the risk becomes acute.

The seductive pitch

A few weeks ago in San Francisco, I walked past an advertisement that crystallized something I’d been sensing. “Sleep in. Let Ona work 996.” The ad’s promise was elegantly simple: automate everything, sleep in, reclaim your life. In short: Wake up refreshed while your AI assistant grinds through the work week. This pitch landed on my ears as fairly cynical. Not because it is technically impossible. As we know, the technology works. But because embedded in that promise lies a fundamental misunderstanding about what we’re actually delegating when we hand work to AI.

Consider this: Research from METR tracks the length of tasks AI systems can handle and shows that capabilities are doubling every seven months. Four-hour analysis projects are now complete in minutes. Impossible tasks from 2020 are routine. For knowledge workers, this creates a potentially existential problem. The tasks we built our value propositions around (research, drafting, analysis, design) exist on a moving target. What commands premium rates today might be table stakes in six months, commoditized in twelve.

The tension is already palpable, and indeed, you can fall asleep yet might wake up work-free… but also jobless.

The magnifier effect

Companies of One are experiencing this shift first and most acutely. No team to buffer the changes. No organizational inertia to slow adoption. No corporate structure to absorb the shock when AI eliminates a skill you spent years building.

This could make us canaries in the mine, yet I prefer to think of us as magnifiers. We’re not warning of danger. We’re showing patterns before they become obvious to everyone else. What’s happening to us now will ripple through corporate structures, large teams, and traditional organizations over the next few years. We’re also learning to refocus and adapt fast.

The question is then: what are Companies of One learning at the leading edge?

Three Lessons from Companies of One

What practices are emerging from those of us experiencing this transformation first?

Lesson One: Essentialise Your Craft

The first lesson is about value, not tasks. Companies of One develop have a craftsperson mindset: you become obsessed not with what you do, but with what value you offer the world. The tasks you execute are fungible. The judgment, synthesis, and context you bring are not.

If you’re a copywriter, your value isn’t writing sentences. It’s understanding audience psychology and creating messages that move people to action. Yes, AI can generate copy that’s on-brief and grammatically perfect. But you’re building something AI isn’t: the intuition for what resonates that comes from watching hundreds of campaigns succeed and fail. If you’re a developer, your value isn’t writing code. It’s architecting systems that solve real business problems. AI can write functions, debug syntax, and even suggest patterns. But you’re making context-dependent tradeoffs (speed versus maintainability, technical debt versus shipping fast) based on where this specific company is in its lifecycle. As a strategist, my value isn’t making slide decks. It’s synthesizing weak signals, connecting dots across domains, and helping leaders see around corners. AI can summarize research beautifully. But I’m the one who knows which anomaly in the data actually matters, and why.

The pattern becomes clear: judgment, synthesis, context, and relationships matter. Execution increasingly doesn’t. And here’s why this matters profoundly: AI is commoditizing execution for everyone. What used to differentiate you (speed, accuracy, volume) are table stakes now.

The premium is moving away from “can you do this?” toward “do you know what to do, and why?”

Lesson Two: Be Intentional About Your AI Tandem

The second lesson is about how you use these tools. Bharat N. Anand and Andy Wu recently published a framework in Harvard Business Review that helps clarify this.

The core of the craft lies in the “human-first” zone: tacit knowledge, high cost of error. Strategy development, relationship building, judgment calls. You must lead here. If you offload this work, you’re not just losing output quality. You’re stopping yourself from thinking and getting new skills. Other quadrants can onboard AI as a first contributor (“Creative catalyst” and “Quality control”) and even as a solo contributor (“No regrets”), but this dark purple quadrant defines your craft today and even more tomorrow.

What I’m finding in my own practice is that the question shifts from “can I automate this?” to “which automation makes me more valuable over time, and which slowly erodes the foundation of my expertise?” This is intentional automation.

That is also the work I do: helping organizations and individual practitioners navigate this transformation without losing what makes them valuable. Not by rejecting these tools, but by being thoughtful about how we integrate them into their businesses.

Lesson Three: Practice JONA (The Joy of Not Automating)

The third lesson is, in my opinion, the most important: it is the Joy of Not Automating (JONA). I hope we can make this acronym stick, because it captures something essential that’s missing from most conversations about AI and work.

JONA isn’t about resisting technology or romanticizing difficulty. It’s about understanding that there’s a profound difference between delegating tasks and delegating learning, between using tools to augment your capability and using tools that replace the experiences through which capability develops. JONA has two dimensions.

The first is qualitative: it’s about interaction patterns and the nature of work.

Affordance theory, developed in 1977 by psychologist James Gibson and applied to technology by Don Norman, describes how objects suggest their own use. A door handle affords pulling. A button affords pushing. The design makes certain actions feel obvious, natural, almost inevitable. AI tools have strong affordances for delegation. The interface makes it easy to ask for output. The speed of response makes it tempting to rely on. The quality of results makes it feel reasonable to trust.

But here’s what I’m noticing in my own practice and in conversations with other independent practitioners: we’re developing interaction patterns that feel efficient in the moment but erode capability over time.

The Shortcut Pattern happens when you use AI to skip thinking. Preparing for a client call? Ask AI to analyze their business and generate questions. It works. You show up prepared. But gradually, the risk is that you’re training yourself out of genuine curiosity, the kind of presence that makes clients feel understood rather than processed.

The Outsourcing Pattern emerges when you treat AI as an oracle. Need to understand a competitive landscape? Ask AI to map it, accept the output, and present it. But you’re training yourself out of the ability to notice adjacent industries that might disrupt this space, out of the instinct to ask “wait, why would they do that?” which often becomes the breakthrough insight.

The Generation Pattern shows up in creative work. Give AI your notes and ask it to write the article. It produces something polished. You edit lightly and publish. But you’ve opted out of the clarity that comes from forcing yourself to articulate your thinking, out of the distinctive voice that develops when you wrestle with finding exactly the right word.

Each pattern feels rational in isolation. Each delivers results. But here’s what research on deliberate practice tells us: expertise is built through struggle. Anders Ericsson’s work showed that mastery isn’t just about accumulated hours. It’s about the quality of engagement during those hours, the struggle, the friction, the repeated attempts to do something slightly beyond your current capability. When you remove that friction, you don’t just save time. You stop developing. Your craft isn’t just what you can do. It’s who you become through doing the work. When a designer spends years perfecting layouts, they’re not just learning software. They’re developing an eye. When a strategist spends years synthesizing research, they’re not just learning frameworks. They’re developing judgment. The question isn’t “Can AI do this?” It’s “How am I interacting with AI, and what is that interaction training me to become?” JONA is about protecting the interactions that build craft, not just the joy of doing the work.

The second dimension is quantitative: it’s about systemic forces.

Because you’re not always in control of what gets automated. Stanford’s SALT Lab (Yijia Shao, Humishka Zope, Yucheng Jiang, Jiaxin Pei, David Nguyen, Erik Brynjolfsson, Diyi Yang) recently mapped thousands of tasks along two axes: AI capability versus worker desire for automation. What they found should concern us: 41% of tasks being automated fall in the “Red Light Zone,” high capability to automate but low desire from workers to automate them. These are tasks people want to keep doing. Tasks they find meaningful. And critically, tasks where expertise gets built.

This isn’t just individual preference. It’s a systemic mismatch. Companies optimize for what’s automatable. Tool builders optimize for what’s technically possible. Clients optimize for faster and cheaper. Nobody is optimizing for what preserves human capability. Which means JONA can’t be passive preference. It needs to be active defense. You need to protect what you want to keep doing, because economic pressure, organizational metrics, and tool design are all pushing you to automate it away.

Tandem: A Desirable Future

So what does it look like when Companies of One (and any other knowledge worker) get this right? When you practice JONA and build intentional automation into your craft?

I think of it as the evolution to a third shape. We started with the T-shape: deep mastery plus good-enough breadth. AI transformed that into the shovel: deep mastery plus genuinely good breadth. But the desirable future isn’t just about expanding capability. It’s about intentional collaboration with AI. This is the Tandem way.

In Tandem, you’re not just using AI to do more. You’re orchestrating a dance where each partner does what they do best. AI handles the baseline. You focus on the breakthroughs. AI does the research aggregation. You do the synthesis that reveals something new. AI generates the first draft. You add the insight that makes people stop scrolling. AI runs the analysis. You spot the pattern nobody else saw.

This is what’s becoming clear in conversations with clients and fellow practitioners: the market is changing what it values. It used to be “Can you do this?” Now it’s increasingly “Can you do this in a way that actually differentiates us?” The answer to that question isn’t more automation. It’s more judgment. More taste. More expertise comes from staying in the game.

This is why I’m increasingly focused on creating practice spaces. Not courses. Not frameworks to memorize. Rather, physical spaces, like the Tandem dinners I’ve been co-hosting, where practitioners and curious minds can wrestle with these questions together. What do I keep? What do I delegate? How do I stay sharp while leveraging these tools? Because the support we need isn’t another automation tool. It’s a community of practice for navigating this intentionally, with peers who understand the stakes.

This is the desirable future for Companies of One: maintaining agency, building skills, staying sharp.

The risk: a modern Dorian Gray

But what happens if you don’t do this? If you automate by affordance instead of intention?

Oscar Wilde’s Dorian Gray stayed beautiful while his portrait rotted. He looked perfect on the surface. Inside, he felt monstrous. By the time he understood what he’d become, it was too late to reverse.

The modern version plays out differently but arrives at the same place. You look incredibly productive. Fast, efficient, delivering constantly. Your output is polished. Your turnaround times are impressive. From the outside, you’re thriving. But then a client asks you to solve a problem without AI assistance. You freeze. Someone challenges your recommendation. You can’t defend it because you didn’t really develop it. You copied and pasted the answer to your prompt. A partner asks you to walk them through your thinking process. You realize: there was no thinking process. Just a series of well-crafted prompts. You appear super skilled on the outside. Inside, you feel like an impostor.

This isn’t hypothetical. In conversations with practitioners across industries (designers, strategists, developers, consultants), I’m hearing variations of this story. A creeping sense that productivity and capability are diverging. That output and understanding are no longer coupled. The trap is subtle because the degradation is gradual. Each interaction feels fine in isolation. It’s only in aggregate, over months, that you notice: the brain muscles have atrophied. The instincts have dulled. The judgment that used to feel automatic now feels uncertain.

In defense of craft and agency.

Here’s what I keep coming back to: we’re not just defending joy, though joy matters. We’re defending craft and experiences through which capability develops. The friction that builds muscle. The struggle that creates insight. The repetition that forges judgment.

We’re defending the ability to think without AI assistance. To solve problems from first principles. To trust our own judgment. To know (not just in an abstract sense but in a visceral, embodied way) how the work actually works.

We’re defending human agency and the capacity to stay above the line, meaning being the one deciding what gets delegated to the AI, and not the opposite.

Companies of One are figuring this out at the leading edge. The golden age is real. The tools are extraordinary. The opportunities are genuine. But the ultimate risk isn’t failing to automate enough. It’s automating what you love doing. It’s delegating the experiences that make you who you are. The question is whether we’re intentional about what we keep. And these realities will soon apply to everyone (and already do for many).

Best,
Margaux

If you’re wrestling with these questions (what to automate, what to protect, how to stay capable while leveraging AI), I’d love to hear from you (margaux at episcope.io or on LinkedIn).

Through episcope, I work with organizations and practitioners navigating this transformation. And through Tandem, we’re building practice spaces for working through these questions together. The next dinners are forming now! Ping me by email (margaux at episcope.io or on LinkedIn).

Yet another type of company of one is spreading across the business and tech landscape with what seems to be the opposite mindset: rapid scaling through AI, aggressive growth, defaulting to automation at every turn... until you arrive at ‘one’ but hollowed out, dependent on systems you don’t fully control, efficient but no longer sovereign, scaled but stripped of your agency and craft.

We think AI enables a third option for companies of one: stay small and still grow your impact with intention. We call it the high-agency tandem – you and the AI.

This way advocates for intentional decisions about what you delegate to the new AI capabilities while staying in control of your business and mind. Not optimizing for scale, but designing for autonomy.

The tension is about agency and freedom.

The difference isn’t what you automate. It’s whether you can still judge and understand the outputs, whether you remain capable of doing the work yourself if needed, and whether you’re building on top of AI or becoming dependent on it.

The tandem way offers different freedom, especially with a frugal and intentional way to use AI: saying no, keeping things understandable, and changing direction without permission. It prioritizes agency, common sense, simplicity, and the commitment to living your life on your own terms.

The goal is clear: Stay above the line, meaning staying the one deciding who delegates tasks or part of your role to the AI, not the other way around.

Tandem doesn’t say no to AI. It says yes to AI on your own terms.

The latest AI tools are incredibly promising, yet critical thinking is fundamental in how we start using them. Remain the decision maker about what gets delegated to AI and what stays in your capable hands, based on what you want to keep doing, not what AI can do. Maintaining your cognitive capabilities while augmenting your craft and deepening your intentional work are key to staying relevant.

This is an invitation for Berlin and Paris (for now).

I’m gathering a small group of fellow companies of one (solopreneurs, consultants, builders, etc) to explore this theme over dinner. We share what we’re learning, what we’re experimenting with, and how we’re navigating this moment.

If this resonates, I’d love to have you join us: we’ll gather in Paris on December 8th and in Berlin on December 11th.

Drop me a note if this resonates, yet can’t make it, it’s only the start!

More on episcope.io/tandem