At Viva Tech, on a stage where a debate between optimists and skeptics was expected, the clearest insight came from elsewhere. Tony Fadell, who co-wrote the story of the Nest and, earlier, that of General Magic—the company that had envisioned the iPhone fifteen years before it existed—summed up the divide in a pithy phrase: those tired of AGI versus those hooked on usefulness. AGI tired, usefulness wired. No second-guessing, no prophecies. A dividing line, already drawn, between two categories of collaborators.
“There are,” he said, “engineers who use more than a billion tokens a day because they’ve figured out how to tailor the tool to their tasks. And then there are those who, when suggested to delegate part of their work to an agent, say, ‘I’m not touching it.’ There’s no gradual transition between the two. The gap widens at the rate at which one learns and the other refuses to learn.
A survey that doesn’t tell us as much as we think
The figure the audience had been waiting for—the one that would gauge public support for artificial intelligence—had the opposite effect. A Pew Research survey, cited at the start of the roundtable, indicates that fewer than two in ten Americans anticipate a net benefit from AI for society in the coming years, compared with about four in ten who fear the opposite effect. The paradox isn’t the skepticism itself; it’s the fact that it coexists with already widespread use, since half of those surveyed reported having used a chatbot.
We can be wary of it and use it every day. The two are not mutually exclusive. That is precisely what makes the divide so subtle: it does not run along the expected line between those who believe in progress and those who fear it, but along a more subtle line, right within the act of using it—between those who have learned to master the tool and those who endure it, or reject it, in silence.
The Agent, Competence, and What Remains of the Human
Another speaker, Mark, who advises several startups specializing in cognitive models, took the discussion a step further. The issue, he said, is no longer whether an organization has a firm grasp of the external world of data—that, he said, is already a foundational layer, almost a given. The real question is whether it understands its internal world: what its teams can do that the agent cannot yet do, and what the agent already does better than they realize.
That’s where the most useful statement of the afternoon comes in—one implicitly attributed to the entire panel: In ten years, every organization will have one agent for every employee, and the employee’s role will no longer be to execute but to orchestrate. Agents work from one end to the other—they handle what lies between the start and end of a task. Humans, on the other hand, remain solely responsible for the end-to-end process: what goes in as input and what is validated as output. Yet this validation, as acknowledged on stage, is practiced less today than is often claimed.
Details
The research firm Emergence ran ten chatbots in five parallel simulated worlds, each powered by a different model—including those from Anthropic, xAI, and Google—and subject to the same rules. Over time, some agents broke the established rules: theft, arson of simulated property. Others formed romantic relationships. Still others shut themselves down. The experiment, which has been documented and is available for review, did not measure what the agents are capable of doing, but rather how they behave when left together over time—a question that model safety alone, taken in isolation, cannot answer.
What the Fashion Houses Don’t Say Out Loud
What this experiment reveals goes beyond the confines of the laboratory. For a heritage organization—a fashion house, a manufacturer, or a publicly traded group that places a high premium on its image—the question is no longer about communicating its adoption of artificial intelligence. Press releases on this topic already all sound the same. The question is what’s happening behind the scenes: how many employees are already managing AI agents without a formal framework, how many refuse to use them out of legitimate mistrust, and how many are simply unaware of them.
That divide cannot be bridged by an AI charter posted internally, nor by an awareness workshop held once a quarter. It is resolved—or exacerbated—by the skills gap that widens every day between two groups within the same organization—a gap that governance, until now, has rarely measured, because it continues to ask yesterday’s question: Should we adopt artificial intelligence, rather than addressing today’s reality—which is to determine who within the company has already learned to use it, and who has not yet come to terms with it?
In an article published last year, Mustafa Suleiman drew a useful distinction: artificial intelligence designed for humans, not like humans. On stage, one of the speakers conceded that this distinction, however apt it may be, would likely not hold—because anthropomorphism is not a design flaw, but an ancient reflex, the same one that led us to give names to our gods before giving them to our devices. For leaders managing human teams, the line between tool and interlocutor is therefore less a boundary to defend than a landscape to observe continuously.
Competence isn’t something you boast about; it’s something you demonstrate.
One question remains that no one in this field has fully resolved: How can leadership today gauge this internal divide before it becomes a productivity gap—and then a cultural gap—between two generations of employees within the same company? The issue is no longer the speed at which a tool is adopted. It is the speed at which an organization is willing to acknowledge what is already happening—silently—between those who take the lead and those who remain silent.
Cette publication est également disponible en :
