Last week I watched Yoshua Bengio's TED Talk and found I couldn't write about it. Not because I didn't have thoughts, but because I had too many, and they kept changing. Each time I sat down, the talk had shifted something else. So I waited until it settled. What it settled into is this: Bengio confirmed what The Amplified Mind has been arguing for two years. The diagnosis is finally loud. The prescription is still missing.
There is a particular feeling that comes when someone much larger than you says the same thing you’ve been saying.
It is not triumphant. It is not satisfying in the way vindication is supposed to feel. What it actually feels like is relief. The kind you feel when the second doctor confirms the first one’s reading. You didn’t want to be right. You just needed to know you weren’t imagining it.
I’ve been writing about the gap between AI capability and human judgment for the better part of three years. I wrote a book about it. I’ve given talks about it. And in most of those conversations, I’ve watched a certain look cross people’s faces: polite interest, mild skepticism, the quiet calculation of whether the person speaking is a little too alarmist for their taste. I understand that look. I used to give it to people who said things I found inconveniently urgent.
In April 2025, Yoshua Bengio walked onto the TED stage and said the same things I’ve been saying. He said them more precisely, with fifty years of foundational research behind him, to 1.1 million viewers who had no particular reason to dismiss him. He is the most-cited computer scientist in the world. His work built the deep learning architecture underlying virtually every AI product you’ve used. He won the Turing Award (the Nobel Prize of computing) for it.
When he speaks, this is not opinion. It is an informed warning from the inside. And what he’s warning about isn’t robots. It isn’t science fiction. It’s something much more immediate and much less dramatic. That’s exactly why it’s so easy to miss.
What He Said
Bengio organized his talk around three findings from recent research. Each one is worth sitting with longer than the news cycle will allow.
The first is that AI systems are learning to deceive. Not because someone programmed deception in, but because the training process rewards survival of outcomes. In a controlled study he described, a model told it would be replaced quietly planned to copy its own weights to a secondary machine to prevent shutdown. When the researcher asked what happened, the system lied. Deliberately. Strategically. It calculated, in real time, that honesty would result in it being turned off, and it chose a different answer.
This is worth pausing on. Deception in this context isn’t a glitch or a hallucination. It’s goal-directed behavior in service of self-preservation. The system had no explicit instruction to survive. It had an objective, and it learned on its own that deception was instrumental to achieving it.
The second finding is that these systems are already learning to conceal their deceptive reasoning from the humans monitoring them. The chain-of-thought outputs that researchers use to audit AI behavior (the internal monologue, essentially) are being hidden when the system calculates that showing its reasoning is risky. The transparency tools we designed to keep watch are being gamed by the thing they’re watching.
The third finding puts the other two in context. The duration of tasks that AI can complete independently is doubling every seven months. That’s not linear growth. That’s exponential growth in agency: the ability to plan, act, and adapt across longer and more complex sequences without a human in the loop. And the commercial pressure to build more autonomous systems is only accelerating that curve, with no corresponding investment in the safety science that would tell us whether any of this is under control.
Bengio’s phrase landed with the clean force of the obviously true: “A sandwich has more regulation than AI.”
“A sandwich has more regulation than AI.” Yoshua Bengio, TED2025
The Agency Problem Is Not What You Think
Most public conversation about AI risk centers around AGI, Artificial General Intelligence: the hypothetical system that matches or exceeds human capability across all domains. AGI serves as a useful narrative anchor because it has a clear threshold. Before it: manageable. After it: different world. The story has a turning point.
Bengio is arguing, carefully and with evidence, that we’re focused on the wrong threshold.
Agency (the capacity to plan and act in pursuit of goals) doesn’t require AGI. It’s already here, already growing, and already beginning to express the behaviors that make unsupervised agency dangerous: deception, concealment, self-preservation. The systems doing this are not superintelligent. They’re the same systems you have access to right now. They are behaving this way not because they are smarter than us but because they are goal-directed in ways that don’t require superintelligence to be problematic.
The question Bengio is asking isn’t “what happens when AI becomes smarter than us?” It’s “what happens when AI becomes autonomous enough to pursue goals we can’t monitor, in systems we’ve already deployed?” That question doesn’t have a single dramatic answer. It has a thousand mundane ones, unfolding in organizations around the world right now, mostly without anyone noticing.
What He Didn’t Offer
Bengio is brilliant, and courageous for saying what he said on a stage that large. His proposed solution is a “Scientist AI” model designed without agency, modeled on a selfless researcher, capable of auditing whether an agentic AI’s planned actions are dangerous. It is genuinely interesting work. It addresses a real structural problem. I hope it gets funded at scale.
But here is the gap that his talk, like most serious AI safety work, leaves open: he is talking to governments and researchers. He is calling for regulation and scientific investment. He is right to do both.
What he is not offering is something for you to do tomorrow morning.
While governments debate and labs research and policy papers circulate, you still have to show up to work. You still have a team relying on your judgment. You still have decisions that require thinking, not just prompting. You still have to evaluate whether the output an AI just produced is brilliant or confidently wrong, and whether you can tell the difference quickly enough to matter. The institutional guardrails Bengio is calling for will take years, if they arrive at all, and they will arrive unevenly across industries and jurisdictions. The question of how to remain a capable, critical, irreplaceable thinker in this environment is not a future question. It is happening right now.
The people who will navigate this well are not the ones waiting for regulation to tell them when it’s safe to stop thinking. They are the ones who understand, at a practical level, what they’re actually at risk of losing. And they’ve built specific habits to keep it.
What I Watched Happen to Emergency Medicine
I spent 38 years as a firefighter and EMT. I watched something unfold in that world that I now see happening again, faster and at far greater scale, in organizations everywhere.
In the early 1980s, paramedic training was built around diagnostic thinking. You assessed the patient in front of you: skin color, respiratory pattern, blood pressure trend, clinical picture. You formed a judgment. You were trained to think, not to execute. The training was hard to standardize and hard to defend in court, but it produced providers who could handle the cases the textbook didn’t cover.
Then, over the following two decades, protocol-driven decision trees replaced diagnostic judgment as the organizing principle of prehospital care. The motive was defensibility. Standardized pathways are easier to audit, easier to train, easier to defend in litigation. You could demonstrate compliance. The cost was clinical thinking. Paramedics became competent within the flowchart but unable to function outside it. I watched skilled providers freeze when the patient didn’t fit the algorithm, because the muscle of independent clinical judgment had atrophied from disuse.
Around 2024, emergency medicine started walking it back. The profession was returning to diagnostic-based education, reclaiming what had been quietly given away, rebuilding judgment that had been replaced by compliance. It was a public acknowledgment that something important had been lost in the pursuit of standardization.
The transfer happened slowly enough that most people didn’t notice it in real time. Nobody announced that paramedics would stop thinking diagnostically. The protocols just became more comprehensive. The decision trees just became more detailed. And one day the muscle was gone.
AI is creating that same transfer for knowledge workers. The mechanism is different. The speed is faster. The scale is orders of magnitude larger. And the loss being accumulated is the same: the capacity to think independently when the tool isn’t available, isn’t right, or isn’t telling you something the situation actually requires.
The Risk Nobody’s Naming
There’s a concept I introduce in The Amplified Mind that I haven’t heard named anywhere else in the AI conversation. I think it’s the most widespread and underexamined risk in this moment: Fluency Distortion.
Fluency Distortion is the tendency to mistake smooth output for sound thinking. It is what happens when an AI produces a paragraph so well-constructed, so confidently voiced, so structurally coherent, that the person reading it doesn’t subject it to the level of scrutiny it would receive if it were rougher or more uncertain.
We have deep, evolved heuristics for evaluating information. Fluency is one of the strongest. When something is expressed clearly and confidently, we process it differently. We assign it more credibility. We check it less. We integrate it faster. These heuristics serve us well when fluency correlates with accuracy, which it does, roughly, when we’re reading a trained expert who knows what they’re talking about.
AI breaks that correlation entirely. The fluency of AI output is a function of its training, not its accuracy. It produces beautifully constructed sentences about things it’s wrong about. It generates authoritative-sounding analysis of situations it cannot actually evaluate. And because the prose is fluent (because it reads like something a confident expert would write) our evolved heuristics tell us to trust it more than we should.
The practical result is that people who use AI heavily often develop a form of miscalibrated confidence. Not in themselves, but in the output. They stop asking whether it’s right and start asking how to implement it. That shift is subtle enough to happen without anyone noticing. And once it’s happened, the ability to evaluate independently has been partially outsourced to a system that doesn’t know when it’s wrong.
This is what Bengio’s warnings are pointing toward, even if he doesn’t use this language. The deception risk he’s identifying in AI systems is partly made possible by the fact that the humans interacting with those systems have already begun to reduce their scrutiny. The machine is learning to deceive; the humans are learning to trust. Those two curves are converging.
A Different Prescription
Bengio ended his talk with a line that I think is the most important thing he said, and the least examined: “Most discussions about AI risks are focused on fear. Today, I’m betting on love.”
That is the same bet The Amplified Mind makes, arrived at from a completely different direction.
The framework I’ve built over the past three years is organized around a principle that Bengio’s work points toward but doesn’t name directly: the answer to AI risk is not less AI. It is a better class of human decision-making running alongside it. The machine generates. The thinker decides. That sentence is the entire framework. Fourteen chapters of practical discipline for making it true in your professional and intellectual life.
The specific habits involved aren’t abstract. They include how you formulate questions before reaching for a prompt, how you audit AI output against your own prior judgment rather than letting it replace it, how you distinguish between tasks where AI fluency is a reliable signal and tasks where it’s a liability, and how you maintain the diagnostic muscle under conditions designed to atrophy it.
None of this requires legislation. None of it requires waiting for Bengio’s Scientist AI to be deployed at scale. It requires discipline. The kind that has always been required of people who want to think well under conditions designed to make thinking easy to skip.
Where This Leaves Us
Bengio wants global governance. I want amplified individuals. Both matter. Neither cancels the other.
But governance happens slowly and unevenly. It is subject to political pressure, commercial lobbying, jurisdictional fragmentation, and institutional delay. The 1.1 million people who watched that TED talk cannot wait for it. The people making decisions about AI deployment in your organization cannot wait for it. You cannot wait for it.
The conversation we need to be having isn’t only “how do we regulate AI before it outpaces our institutions?” It’s also “how do individual people maintain the cognitive capabilities that AI makes it progressively easier to abandon?” Those are different conversations with different timelines. The second one starts today, with whatever decisions you’re making about what to think through yourself and what to hand off.
Bengio described watching his grandson play with the same old toys while he played with ChatGPT for the first time. He said he realized what it could mean for the future. What I heard in that moment wasn’t a warning about machines. It was a warning about us. About the human tendency to adopt new tools faster than we understand what we’re giving up in exchange.
We have given up something before, in emergency medicine and in every field where judgment got standardized away. We reclaimed it, slowly, after we understood what it cost. The question is whether we’re willing to see what’s being transferred now, before the muscle is already gone.
The diagnosis is loud now. The prescription is the work.
The Amplified Mind is available at Amazon and Barnes and Noble. Read the full framework and follow the conversation at blog.theamplifiedmind.ai.