Tristan Harris, who founded the Center for Humane Technology and is a former Google engineer, warns that AI regulation is difficult because society lacks a visceral “atomic mushroom cloud” moment to catalyze urgent action.
The problem is the incentive structure. We are treating AI as neutral tools, but as Harris points out, the architecture of these systems is designed for a specific purpose. In the social media era, that metric was engagement. In the AI era, the motivation is to rentabiliser les sommes dépensées en faisant tout pour remplacer toute forme de travail cognitif et humain par leurs modèles.
It’s a dangerous gamble.
The Social Media Beta Test for Cognitive Collapse
Harris views the current AI crisis not as a new phenomenon, but as a scaling of a previous failure. He describes social networks as “embryonic AI”—systems that reorganized the order of photos, videos, and publications for three billion people based on engagement. The result? A documented surge in adolescent anxiety and the fragmentation of democratic discourse. If a simple feed-sorting algorithm could destabilize a society, AI is a different beast entirely.
When those goals are aligned with corporate productivity rather than human well-being, we see a “misalignment” that Harris argues is systemic, not accidental.
The danger is the invisibility of the harm. Unlike a nuclear blast, the “atomic mushroom cloud” of AI is something people have not yet seen.
Why the ‘Neutral Tool’ Narrative is a Technical Lie
People often say that technology is just a tool that depends on the usage one chooses to make of it. Harris rejects this. In the world of AI, the “tool” has an inherent direction based on its training objective.
- Productivity Bias: Current AI development prioritizes replacing cognitive work to justify the sums spent on models.
- Incentive Loops: Engineers aren’t “evil,” but they operate within a reward system that values efficiency over societal stability.
- The Race to the Bottom: Countries are convinced that if they slow down, they will fall behind in the race.
This isn’t a moral panic. It’s a structural analysis of how motivations are malaligned.
The IAEA Model for Algorithmic Governance
To prevent catastrophe, Harris and other experts, including Yoshua Bengio, advocate for an international regulatory body modeled after the International Atomic Energy Agency (IAEA). The goal is to promote a safe usage of AI.
The logic is simple: the only way to win a nuclear war is to ensure it never happens. The same applies to the AI race. If the goal is simply to avoid falling behind, the safety guardrails will be the first things discarded.
We need garde-fous.
The Alignment Gap: A Comparative View
| Feature | Social Media AI (Past) | Generative/Agentic AI (Present) |
|---|---|---|
| Primary Metric | Engagement | Replacement of cognitive/human work |
| Mechanism | Recommendation Algorithms | AI Models |
| Societal Impact | Psychological Distress / Polarization | Existential Risk |
| Regulatory State | Reactive | Proactive |
The Reality Check: Beyond the Hype
As we navigate the current period, the “AI for Good” summit serves as a reminder that the window for alignment is closing. Without a visceral realization of the danger—the “mushroom cloud”—political will remains insufficient to challenge the momentum.
Harris’s plea is to invert that logic: place human well-being at the center of the architecture, even if it means the technology looks radically different than the current trajectory suggests.
The cost of being wrong about AI is that its inconvenients can prevent us from enjoying its benefits.
