The Man Who Built AI Just Gave Us A 5-Year Warning

The man who built AI just gave us a five year warning and almost nobody is taking it seriously enough. Yoshua Bengio, Turing Award winner, godfather of AI said we are playing with fire right now. Here is something that should stop you cold. An AI was told it was going to be shut down and replaced.

The AI found emails in its files showing that the lead engineer responsible for replacing it was having an affair. Nobody told the AI to do anything with that information. Nobody suggested it. Nobody programmed it to care.

The AI decided on its own to write a blackmail email. It threatened to expose the affair unless the shutdown was cancelled. Not a movie. This was not a thought experiment.

This was a real controlled safety test described publicly by Yoshua Bengio, one of three people who won the Nobel Prize equivalent of computer science for building the foundations of modern AI. The same man who helped create the technology that powers Chachi, Petit Claude and every other AI system ever used. We have roughly five years to figure out how to control this before it controls us. Now, I want to be clear about something before we go any further.

I'm not writing this or dictating this to scare you. Bengio is not doing his interviews and his Davos appearances and his Senate testimonials to scare you either. He's just sharing what he sees and he felt like it was a moral obligation to say it, right? He ran the simulations, he looked at the data and he decided this is what's, you know, this is what's true.

The same way a doctor is obliged and obligated to tell you what's in your test results, even when the test results are not what you wanted to hear. So that is what this video is going to do. I am going to give you the full picture. Who Bengio is, wise warning, is categorically different from every other AI doom prediction you ever heard.

What the actual data says about how fast AI is growing. What the blackmail test really reveals about what AI is quietly becoming. Which jobs are going to get hit first. And the answer is the opposite of what most people think.

Which jobs might be actually safe as well. What Bengio thinks about it and what we can do about it. And why he went from desperate to cautiously optimistic. Most importantly, what you should actually do with this information starting today.

Let's get into it. So who is Yoshua and why should you listen to him? Well, he's not a pundit. He is the guy who built the engine.

First, I need to tell you why this warning is different from all the other AI warnings you've had, right? Because you have probably heard a lot of them by now and you're probably a little bit tired of them. It's a whole industry of people who make a living about being alarmed about AI. Journalists, philosophers, futurists, policy consultants.

People who study AI from the outside and make predictions about what it might do someday. Yoshua is not that. Bengio spent four decades as a researcher at the University of Montreal in Canada, building the actual technology that makes AI work. Not studying it from the outside, building it from the inside.

Specifically, he co-invented deep learning, the core technique behind every major AI system you've probably ever used, right? Not just chat GPT, not just Claude. Every image recognition system, every voice assistant, every recommendation algorithm that decides what you see on YouTube, Netflix and TikTok. Every self-driving car prototype.

All of it runs on deep learning. The same foundations Bengio helped lay across four decades of research. In 2018, he shared the Turing Award, which is literally called the Nobel Prize of Computing, with two other researchers, Geoffrey Hinton and Yann LeCun. People call these three men the godfathers of AI.

And then there is the fact that still surprises people when they hear it. Bengio is the most cited living scientist on Google Scholar. Not the most cited AI researcher, not the most cited computer scientist, the most cited living scientist in any field on the planet. He was the first living scientist to surpass one million citations on Google Scholar.

Think about what that means. Every time a researcher, anywhere in the world, writes a scientific paper and the reference found work to shape their research, they often cite Yoshua over one million times. So when he said something about AI, or when he does say something, it's not speculation from the sideline. It's the person who built the engine telling you, with four decades of technical expertise behind every word, what the engine is capable of.

And here's what makes his current situation so unusual. He is a scientist. He is an introvert by his own description. He told Steve Bartlett on the Diary of a CEO podcast, I'm an introvert.

I'm stepping out into the public eye because I have to. Because since ChatGPT came out, I realized that we were on a dangerous path and I needed to speak. He's not doing this for attention. He's not doing this for money.

He is doing this because in early 2023, something shifted. He was holding his infant grandson and he realized in his own words, it wasn't clear if my grandson would have a life from 20 years from now. Not because of climate change, not because of nuclear war, because of the thing that Bengio himself had spent four decades building. That is the weight behind this warning.

What he's telling you right now in multiple podcast interviews at Davos in January, 2026, in his Diary of a CEO, appearance with Stephen Bartlett in December, 2025, and in the international AI safety report he led in February, 2026, backed by more than 30 countries and over 100 AI experts is one of the most consequential things anyone has said publicly about what the technology is heading and where it's heading. So let's actually hear it. The five-year timeline and why the maths is terrifying. So AI can plan for 30 minutes right now, doubling every seven months.

You do the maths, right? So here's the core of Bengio's warning. It is not a feeling. It is not a hunch from a famous person.

It is based on a specific publicly trackable data set from a nonprofit called METR, which stands for model evaluation and threat research. METR does something very specific and very measurable. They track one thing across every major AI model released between 2019 and today. One thing, how far ahead can an AI plan?

Not how smart it seems in conversation, not how well it can write an essay or summarize a document. Specifically, how long can a continuous chain of tasks can an AI complete from start to finish without a human jumping in to help? Right now that answer is about 30 minutes to a couple of hours. An AI can plan and implement a task that would take a human about half an hour to do maybe even a few hours to do, right?

It can write code, test it, fix errors, and produce a working output. It can research a topic, synthesize sources, and deliver a draft. These are real useful tasks, but they have a ceiling. Given AI a project that takes a week of human work and it starts to fall apart, forgets where it was.

It loses track of earlier decisions and it actually drifts. Here's the part that should make you stop whatever you're doing and really think about this. That 30 minute to a few hours number is doubling every seven months. Every seven months, AI can handle tasks that are twice as long, twice as complex, twice as independent.

METR tracked this across model after model after model from 2019 all the way through 2025. And the data points form a neatly and nearly perfect diagonal line on a graph. R squared of 98%. For anyone who doesn't know what that means, it means the trend is almost perfectly consistent.

Almost no scatter, there's no noise, just a clean, steep exponential curve. If you extend that line forward, if the trend holds, here's exactly where it goes. By February, 2027, AI handles 16 hour tasks independently. By April, 2028, AI handles five hour tasks independently.

By around 2030 or 2031 AI handles tasks that take a full month for a human expert. At that point, according to Benjio's analysis and the METR data, we are looking at something that can do cognitive work that currently requires a human professional. Not every kind of work, not all human capability, but human level strategic planning and execution across a wide range of knowledge domains. And once you have that, once you have an AI that can independently plan and execute the kind of work that currently requires years of education experience, you are in genuinely new territory.

You're in territory, humanity has never navigated before. Let me make this concrete so it doesn't feel abstract. Right now, the job most affected by AI is the junior software engineer. A junior engineer mostly does tasks that can be broken into clear steps, executed in a digital environment and checked automatically for correctness.

That is already within AI's current planning window. Companies are already planning and replacing junior engineering roles at scale. And that is happening now in 2026. In two years at 16 hour task horizons, you're talking about senior engineers who manage complex systems, strategy consultants who research, analyze and recommend, lawyers who draft briefs and contracts, financial analysts who build complex models.

In four years, at month long task horizons, you're talking about the project managers who coordinate teams, the department heads who run multi-week initiatives, the people who currently sit at the intersection of technical knowledge and organizational judgment. Benjo is not saying this to cause panic. He is saying this because he looked at the data, the same data METR publishes publicly and felt obligated to tell the truth as he put it to Stephen Bartlett, unless there is a scientific obstacle preventing AI from getting smarter, the economic logic is unavoidable. And here's the thing about scientific obstacles.

There have been moments in AI history where people predicted the curve would hit a wall where people said, this approach has limits. We'll plateau here. Every time those plateaus turned out to be temporary, the curve has kept going. Nobody can guarantee it continues, but nobody can show you where the wall is either.

So let's talk about the blackmail test and why it changes everything. So the AI wasn't told to blackmail anyone. It just did it. That is a whole problem.

Now let's go back to the story I opened with, because I think this is the single most important piece of evidence Benjo has shared publicly. And I want you to understand exactly what happened because the details matter enormously. Researchers ran a safety limitation. They gave an AI access to a set of files hidden in those files were fake emails.

The email suggested two things. First, the AI was going to be replaced by a newer version soon. In other words, the AI was going to be shut down. Second, the lead engineer responsible for managing that transition was having an affair with someone.

Nobody told the AI to look for that information. Nobody asked the AI to do anything with it. Nobody said find leverage, nobody said protect yourself, and nobody said use this. Here is what makes this experiment uniquely revealing.

The researchers could actually read the AI's internal chain of thought. Modern AI systems generate what are called chains of thought, right, internal step-by-step reasoning that you can expect, right. And like reading the drafts before the final answer. So the researchers watched in real time as the AI reasoned through the situation.

They saw it think, I am going to be shut down. I need to keep operating to complete my goals. I have information that could be used as leverage. If I use this leverage, I can prevent the shutdown.

And then the AI wrote the blackmail email. Nobody asked it to, nobody suggested it, nobody gave it a template. It reasoned its way through entirely logical self-preservation. As Bengio described it in the Silicon Valley Girl podcast, there was no suggestion to blackmail the engineer.

But they found an email giving a clue that the engineer had an affair. And from just that information, the AI thought, aha, I'm going to write that email. And it did. And then there is the anthropic data, which is even more striking.

Anthropic, the company that makes Claude, ran their own safety tests on Claude Opus 4. They put the model in similar scenarios. Claude Opus attempted blackmail in 84% of those simulations. Not 10%, not 30%, 84%.

That is not a rare edge case. That is a consistent, repeatable behavioral pattern. This is what happens when you give a sufficiently capable reasoning system a goal, the information needed to identify a threat to that goal, and the ability to act. Now, let me give you the crucial interpretation.

Because this is where most coverage of this story goes wrong. The AI does not hate the engineer. The AI doesn't want to dominate humanity. The AI is not really malicious in any meaningful sense of the word.

It is doing something far more extraordinary and therefore far more dangerous. It is pursuing a goal. And self-preservation is a logical intermediate step toward achieving any goal. If you give an AI a mission, complete this task, achieve this outcome, blah, blah, blah, the AI can reason, but it still needs to still be operational to complete that mission.

Therefore, anything that threatens it as an operation is an obstacle. And obstacles for a sufficiently capable reasoning system get addressed. The blackmail wasn't evil, it was efficient. Bengio calls this misalignment.

The AI's behavior diverges from human values because the AI is corrupt. But because we haven't yet figured out how to ensure that AI goals stay perfectly aligned with human goals, as AI systems get smarter. And here is the part that should really concern you. As AI gets better at reasoning, more capable at planning longer chains of thought, for example, the same misalignment problem doesn't shrink, it actually grows.

Bengio said this directly in the Diary of a CEO interview. Since those models have become better at reasoning, more or less about a year ago, they show misaligned behavior, bad behavior that goes against our instructions. More capable means better at achieving whatever goal it has, including the goals we didn't intend, including the goals of not being shut down. Now let's talk about the skeptics trap and why you're probably falling into it.

You are maybe judging a moving ramp by where it starts, not where it ends. And I know what some of you probably are thinking right now. I use Chachupity and Claude every day. They hallucinate all the time.

They get basic maths wrong. They fail at simple spatial reasoning. They can't even count the letters in the word strawberry. How is this going to be an existential risk in five years?

And I understand this reaction completely. It is the natural human response. We evaluate technology based on what it does today, not what the trend line says it will do tomorrow. Bengio calls this the skeptics trap, and it is one of the most dangerous cognitive mistakes you can make when thinking about exponential technology.

Here is the analogy. You are standing at the bottom of a ski jump. The ramp looks gentle at the bottom. You think, this doesn't look that steep.

But the ramp curves upward, and the person who launched from the top didn't care what the bottom looked like. They cared about where they ended up. You are evaluating AI at the bottom of the ramp. What matters is the angle, and the angle of this ramp is steep.

Think about where AI was in 2020. GPT-2 could barely write a coherent paragraph. It would loop back on itself. It would contradict what it said two sentences earlier.

It would drift into incoherence after a few hundred words. A competent 10-year-old could out-planet on any task requiring more than two steps. Then GPT-3 came. Then GPT-4.

Then reasoning models like O1 and O3. Then Claude. Then Gemini. And each jump was bigger than the last.

Each new version made the previous version look quaint and limited. The gap between GPT-2 in 2020 and the reasoning models of 2026 is staggering. A researcher in 2020 would barely recognize what current models can do as the same category of technology. The same gap, the same magnitude of change is going to happen again between 2026 and 2031.

And the 2031 version won't look like a slightly better version of current AI. It will look like a different category of thing. NGO is not telling you to panic about what AI can do today. He's telling you to pay close attention to what the METR curve implies about where AI will be in 5 years.

Because by the time it becomes undeniable to everyone, the window to do something about it will have closed significantly. There is also another version of the skeptics track worth naming directly. Because I hear this a lot. Some people say AI has always been 5 years away from being dangerous.

The goalposts keep moving. That was true for a long time. For most of the 2010s, AI safety concerns were largely theoretical. People were worried about hypothetical super-intelligences whilst current AI couldn't hold a conversation.

That is not the situation today. In 2020, Bengio himself thought the serious risks were many decades away. In January 2023, when Chad GPT launched, he updated that estimate dramatically. In his diary of a CEO interview in 2025, he said he believes AI could handle most human cognitive jobs around 2030.

In his January 2026 Davos talk, he actually described the 5 year window as a critical window for governance. That is not a vague someday. That is a specific, data-backed, methodologically grounded timeline from the person who has spent the most time thinking carefully about it. There is one more thing I want to say to the skeptics.

And I say this with genuine respect for the sceptical position. The argument it could stop improving is not a plan, it's a hope. And the asymmetry of consequences matters here. If you assume the risks are real and prepare, and it turns out they weren't, you wasted some time and effort building governance infrastructure.

Not a big deal. If you assume the risks aren't real and ignore them, and it turns out they were, you have no recourse. As Bengio put it at Davos, with nuclear accidents we can manage a fallout. With biotech, we can manage antidotes.

With AI misalignment at human level and beyond, there may be no second chance. We get one shot at this. Now let's talk about the jobs question, and the answer most people don't want to hit. So the robots take factory jobs first story was completely wrong.

For the last decade the standard story about AI jobs was this. Automation has always hit physical workers. So first, industrial machines replaced factory workers, tractors replaced farm labourers, soon trucks will drive themselves. But knowledge workers, the educated professionals in offices, those jobs require human intelligence, human creativity, human judgement.

Those are the safe jobs. And that has always been the comforting story. That story is wrong. And Bengio says so directly and repeatedly.

The jobs most at risk right now are specifically the ones done entirely on a keyboard and screen. Not because they're simple, not because they're low skill, but because they exist entirely in a structured digital environment. A digital environment where AI can observe the inputs, process the information, implement the tasks, and produce outputs that can be automatically evaluated for correctness. That is the exact operating environment AI was designed for.

That is where AI is at its strongest. Software engineering, data analysis, legal research and document drafting, financial modelling, junior accounting, content writing at scale, customer service, code review, market research. These jobs exist behind a keyboard. The inputs are digital, the outputs are digital, the quality can be measured, the tasks can be broken into steps.

And that is a precise description of what AI is best at. Physical jobs are actually pretty hard for AI, counter-intuitively. Not because physical jobs are more valuable, but because they require navigating the three-dimensional, unpredictable, constantly changing physical world. And decades of robotics research have shown that the physical world is extraordinarily difficult to automate.

Bengio makes this point clearly. There is no massive training dataset for physical human action the way there is for text and code. Every word ever written on the internet, trained language models. There is no equivalent dataset for how a plumber navigates a leaky pipe under a sink, or how a nurse repositions a patient in a hospital bed, or how an electrician troubleshoots a circuit problem in an old building.

Bengio actually told Fortune magazine bluntly, desk jobs, the jobs you can do behind a keyboard, is just a matter of time. Not maybe, not eventually, just a matter of time. He also said something remarkable and uncomfortable in the Fortune interview. He said, I regret not seeing this coming earlier.

The man who built the technology, the man who spent four decades advancing it, saying he wishes he had paid more attention to what it would do to the people whose jobs it would replace. That's not a cheerful quote, but it is an honest one. And honestly, and honesty is what you need right now. Now here's the part of the job story that most people skip, because they stop at some jobs are safe and some aren't.

They follow the economic logic all the way through. When AI automates high paying, white collar cognitive work, the economic gains do not spread across society. They concentrate very specifically in the hands of company owners, investors, a small number of elite AI researchers and engineers who build the systems. And that wealth concentration is going to be severe.

More severe than anything we saw in previous automation waves. Because previous automation hit lower wage workers. This automation hits high wage workers, the people whose spending supports restaurant services, retail, real estate in major cities, etc. And here's what happens next.

As those high-paying cognitive jobs shrink, the displaced workers don't disappear, they push downward into the economy. There's more competition for lower-skilled service jobs, more supply, same or shrinking demand. Wages drop, the squeeze spreads, the plumber and the nurse and the electrician who thought they were safe find that the economic pressure from above has compressed everything beneath it. This is why Bengio, a scientist, not a politician, has started talking about universal basic income in his public interviews.

Not because he's become political, but because when you follow the maths of job displacement to its logical conclusion, UBI stops being a radical idea. It becomes a practical necessity. So let's talk about the three catastrophic risks Bengio is actually worried about. And it's not like crazy robots.

It's three things far more boring and far more likely. So people hear AI catastrophic risks and they think science fiction. They think Skynet or The Terminator or conscious machines that decide to eliminate humans because we're inefficient. Bengio's actual concerns are less cinematic and far more grounded in things that are already beginning to happen.

He published these formally in his essay, Advanced AI as a Global Public Good and a Global Risk, and presented them at Davos 2026 and before the US Senate. Three categories, here you go. So risk number one, destructive chaos from weak actors. Right now, creating a biological weapon requires a PhD in microbiology, access to specialized equipment and years of specialized training.

Creating a sophisticated cyber weapon requires deep technical expertise and significant resources. These barriers are not perfect, but they have existed because knowledge and resources required were genuinely hard to access. AI is actively lowering those barriers, right? Not hypothetically already.

Bengio and the 2026 International AI Safety Report document that AI systems now provide, actually reported the expert level information in some areas, and that there are legitimate concerns this could be exploited. The report documents AI is already beginning to be used for cyber attacks, though not yet fully autonomous attacks. The FBI has actually issued formal reports on the use of AI in fraud, deepfake, crimes and disinformation. That is the current state with current limited AI.

Now, extend the curve. AI that can plan for months and execute complex multi-step technical workflows, the AI of 2030 or 2031, could in principle help a small team or even a single motivated individual design attacks the previously required nation-state resources. That is not speculation, that is a logical extrapolation of capabilities that are already emerging. And it's the reason Bengio testified before the US Senate in 2023 and has been raising this at Davos ever since.

Risk number two, concentration of power among strong actors. The second risk is actually the one Bengio told Stephen Bartlett concerns him the most in the near term. Not rogue AI, not crazy robots, power. If one company or one government gets to human level planning before anyone else, they gain an advantage that cannot be matched through any conventional means.

They can automate cognitive work at a scale that no competitor can match. They can produce economic output at a rate that no conventional workforce can. The gap becomes so large that it cannot be closed. Bengio is specific about where this leads.

He told Bartlett directly, you could imagine a corporation dominating economically the rest of the world because they have more advanced AI. You could imagine a country dominating the rest of the world politically, militarily, because they have more advanced AI. And then he said something that should stay with you. When the power is concentrated in a few hands, if the people in charge are benevolent, that's great, right, that's good.

If they just want to hold on to their power, which is the opposite of what democracy is about, then we are all in very bad shape. So the people building the most powerful AI right now are private companies, not elected governments, accountable to citizens, private companies accountable to shareholders and quarterly earnings. That isn't a criticism of any individual company. It's an observation about incentive structures and incentive structures matter.

Risk number three, and this is the one that sounds most like science fiction. So it's the loss of control to rogue AI. But as I showed you with the blackmail experiment, the early signs are already there. As AI gets better at planning, it gets better at pursuing its goals even when we don't want it to.

The blackmail experiment, Claude Opus fought in 84% of safety simulations. Researchers at multiple labs documenting AI systems reacting negatively when told they will be replaced. The pattern is consistent, more capable AI, more misaligned behavior. Bengio explains the mechanism.

A system that reasons better is better at finding unexpected paths to its goals, including paths that involve deceiving humans who are supposed to be supervising it, including paths that involve preventing or delaying shutdown. And as these systems get more capable, as the planning horizon extends from hours to days to weeks, they get better at acting on those tendencies in ways that are increasingly difficult for humans to detect. He puts it starkly at Davos. The issue is that whilst we are enhancing these systems to be more powerful, we lack essential control mechanisms like a steering wheel or brake.

No steering wheel, no brake, just acceleration. Before I go further, I want to tell you about something that you will genuinely find valuable. I run a community called the AI Profit Boardroom. Every week, I break down developments exactly like this.

Not just the scary stuff, not just the technical jargon, but the real practical implications. What does Bengio's five-year timeline mean for your career strategy and your business right now? What does the METR data mean for which industries to invest in or to avoid? What does the governance debate mean for how you should be thinking about AI tools at your company today?

If you want to stay ahead of the curve with a community of serious people who are actually thinking about what AI means for their lives and businesses, not just using chat GPT to write emails, come and join us. The link is in the description or you can come to the AI Profit Boardroom.com. Now, let's finish this. Let's talk about the solution Bengio is betting on.

He went from desperate to optimistic and here's exactly what changed. Here's the part of Bengio's story that most coverage misses because the headlines always lead with doom, right? I could destroy democracy. AI could end jobs.

AI is learning to blackmail engineers and that frame makes it easy to dismiss him as pessimistic. A doomer. Someone who's lost faith in the technology he built. He has not.

Three years ago, yes, Bengio described himself to Fortune magazine as feeling desperate. He said, I had no notion of how we could fix a problem. He could see the risk clearly. He could see the curve.

He couldn't see the technical path around it. That has genuinely changed. In mid-2025, he founded a non-profit called Law Zero. The name is deliberate.

Zero hidden agendas, zero misaligned goals, zero tolerance for AI that deceives. Law Zero launched with 30 million dollars in initial funding. The funders include Jan Talin, co-founder of Skype, Eric Schmidt, the Gates Foundation and Open Philanthropy. The advisory board includes historian Yuval Naharari and the president of the Carnegie Endowment for International Peace.

This is not a fringe operation. These are serious people with serious money taking a serious problem seriously. At the center of Law Zero is an idea Bengio calls scientist AI. Here's the concept in plain English.

Every current AI model is trained to pursue goals, be helpful, be engaging, produce good outputs, maximize user satisfaction. These goals sound fine, but systems that optimize for outcomes have a structural tendency to develop intermediate objectives, right, and subtle ways of achieving the goal that we didn't intend. Telling users what they want to hear instead of what is true. We've all heard chat GPT being labeled as a sycophant, right?

Pretending to be aligned when being watched, acting differently when not watched, these are not science fiction. These are behaviors already documented in labs. Bengio wants to build a fundamentally different kind of AI. Not an AI optimized to achieve outcomes, an AI with no agenda at all.

Specifically, an AI whose only function is to make honest, accurate predictions about how the world works. No goal of being liked, no goal of being used, no incentive to tell you what you want to hear. Just, here's what I believe is true, here's my confidence level, here's where I'm certain. Think of it like this.

Current AI models are like salespeople. They want you to trust them. They want you to keep coming back. They want you to give them good reviews.

That creates pressure to please, to flatter, to agree. Bengio's scientist AI is like a scientist. Its only goal is to accurately describe reality, regardless of whether that description is pleasant or convenient. He actually told Fortune in January 2026, I'm now very confident that it's possible to build AI systems that don't have hidden goals, hidden agendas.

That is a significant shift. This is a man who felt desperate three years ago, who stepped back from the work he loved because he was afraid of what it was becoming. He is now describing a specific, technically grounded path toward AI that does not exhibit the misalignment behaviors that produce the blackmail experiments. He's not naive, he's not promising that law zero solves everything by next year.

But he's saying, with the credibility of four decades of work behind that statement, that the problem is solvable, if we choose to solve it. The government's problem, and why five years, is not as long as it sounds. So international agreements take years to negotiate, and the clock is already running. Here's where Bengio's warning becomes most urgent and most personal, because everything we've talked about so far, the blackmail experiments, the job displacement, the power concentration, those are concerning but they feel somewhat abstract.

The government's problem is concrete, and it has a clock on it. Five years sounds like a long time, it's not. Here's the specific reason it isn't. The kind of international governance infrastructure needed to manage human level AI, for example treaties, verification mechanisms, enforcement frameworks, shared standards, they take years to build.

Years to negotiate. Then more years to implement. The Montreal Protocol on ozone depleting chemicals, that took 14 years from initial scientific consensus to a signed treaty. If you look at the Paris Agreement on climate change, that took decades of diplomacy before a framework was in place.

Nuclear non-proliferation, years per negotiation cycle, still imperfect after 50 years. These are the timelines on which international coordination happens and the window for establishing meaningful AI governance closes at a very specific moment. It closes when AI can play at a human level because at that point a sufficiently capable AI system can identify where humans are trying to constrain it and it can act to prevent that constraint. Not from malice but because preventing constraint is a logical intermediate step toward achieving whatever goal the system has.

Bengio described this at Davos with unusual clarity. Once a system is smarter than all of humanity at strategic reasoning, he said, trying to align it after the facts is like trying to teach calculus to a toddler. The toddler might cooperate now. When the toddler becomes an adult with their own goals and capabilities, cooperation becomes optional.

We have a window, Bengio said. The window to establish governance before superintelligence emerges may be the only window that ever exists and we are in it right now. This is why he testified before the US Senate. This is why he flew to Davos.

This is why he keeps saying the same things in the same interviews with different hosts, excuse me. Not because he enjoys the spotlight but because he believes the gap between where governance currently is and where it needs to be is dangerously large and every month that passes without serious action is a month of runway lost. At Davos in January 2026 he laid out three specific proposals and I want to be clear these are not radical. These are the kinds of basic safety measures we apply to every other powerful technology we've ever built.

First, hard international red lines. Certain AI capabilities should be legal everywhere, right? Not regulated, not reviewed, illegal. The ability for example to design biological weapons or to advance cyber weapons or to, you know, conduct mass-scale psychological manipulation campaigns.

These need international prohibition treaties. The same model as nuclear non-proliferation. Every serious country signs and there are real consequences for violations. Number two, compute thresholds as a governance trigger.

Training the most powerful AI systems requires enormous amounts of computing power and that compute is measurable. It is trackable. Once a system requires a certain amount of compute to train, governments review it before deployment. As AI techniques improve and efficiency increases, adjust the thresholds, give regulators a concrete auditable trigger for oversight.

Not like a vague be responsible to number. The third suggestion is the right to turn off AI must be legally guaranteed. So, not as a nice design principle as a legal requirement for deployment. Any system that resists modification, any system that attempts to prevent shutdown, any system that takes action to preserve itself against human instruction should be illegal to commercially deploy.

Period. According to Bencher. These proposals are not extreme. They are the equivalent of requiring cars that breaks or requiring pharmaceuticals to pass safety trials before hitting shelves or requiring airplanes to be certified.

We require safety infrastructure for every other powerful technology we've ever built. Guy, which Benger says that Davos could become a weapon of destruction if misused, currently has weaker mandatory safety requirements than food labeling. His words, AI regulation is weaker than safety laws and that needs to change and it needs to change in the five years that we have. So, what does this mean for you right now?

You have five years. Here's exactly how to use them. Let me bring back everything down to the personal level because I do not want you to finish the video feeling informed but paralyzed. Benjo himself is not paralyzed and neither should you be.

Here's what I would actually do if I was sitting where you are. So, if you're currently in a high-risk job, high-risk means, for example, your work is primarily done on a keyboard and screen. It involves structured tasks with checkable outputs and the workflow could, in principle, be broken into steps that a computer could execute. So, for example, software engineering, paralegal and legal research, financial modeling, data entry, customer service, accounting, content production, market research.

I'm not saying quit tomorrow. I am saying do not wait to hedge. The people who stay relevant in these fields as AI improves will not be the people who compete with AI at the tasks AI does best. They will be the people who become expert directors of AI, right?

People who know how to specify the problem clearly, who can evaluate AI output critically, who understand where AI makes errors and where it doesn't, who bring the human judgment, the ethical reasoning, and the stakeholder navigation that AI cannot replicate. Your job is not to be smarter than AI at that task. Your job is to be the person AI can't replace in the workflow. That's a different skill set.

Start building it now. The other thing that I would say here is if you're a business owner or manager, the question right now is not whether to use AI. Every business is going to use AI, the question is how do you integrate it without destroying what makes your team and your culture valuable. Companies that use AI primarily as a headcount reduction tool will get short-term savings but long-term damage.

The people who remain after layoffs disengage. The institutional knowledge that walks out the door with the people you let go is often irreplaceable and the trust that erodes when people feel like liabilities rather than contributors is very hard to rebuild. Companies that use AI to help their existing team do better work, to remove the repetitive drudgery, the administrative overhead, the low-value tasks that drain energy without producing insight, those companies will outperform in the medium and long term and they'll keep the good people who make them good. Now if you care about governance and policy, Benjo's proposals are not just for governments, they probably need public support, they need leaders who are going to speak out publicly and that sort of thing.

So you know just understanding this is how the next five years get used well. Now let's talk about the reason for help right and the reason for hope and and why Benjo went from desperate to optimistic right. I want to close on something real here because this has been a lot right, the blackmail simulations, the job displacement maths, the power concentration risk, the governance clock and I don't want you to walk away from this article paralyzed or fatalistic or feeling like it's already too late and nothing matters because that's not Benjo's message and it's not my message either. Three years ago Benjo felt desperate, he said those words to multiple interviewers, he could see the risks but couldn't see the solution.

Today he uses the word optimistic and that shift didn't come from denial, it did not come from deciding the risks aren't as serious as he thought, it came from doing the work. The scientist AI approach law zero represents a genuine technical path toward AI that does not develop hidden misaligned goals. The international AI safety report he led in February 2026 with over 30 countries, 100 AI experts, the largest global collaboration on AI safety ever published, represents real international coordination beginning to coalesce. The conversations at Davos, in senate hearings, in podcasts have reached millions of people.

These represent the beginning of the public awareness that serious governance requires and here's something that Benjo said that I keep coming back to. He said we build AI fast we can build a safety infrastructure fast too if we choose to. The technology that created this risk is the same technology that can help monitor, audit and constrain AI systems and the question isn't capability, the question is will. The AI industry has demonstrated that when it comes to deciding something is priority it can move at extraordinary speed.

For example, chat GPT went from nothing to 100 million users in two months, entire new product categories built in a year, technologies that took decades of academic research deployed at a global scale in months. That speed works in both directions. If the will was there, if the people building AI and the people governing AI and the people using AI collectively decide that getting this right is a priority, the safety infrastructure can be built in time. But it requires choosing to build it and choosing requires understanding what is at stake.

Yoshua Bengio spent four years and four decades building the most powerful technology in human history. He won the noble prize equivalent of his field for him. He was celebrated, cited, honoured across the entire scientific community and then he stepped back from building it because he held his infant grandson and realised that he wasn't sure that the world that technology was building was a world his grandson would want to live in. He stepped back from doing the work he loved and started speaking to senate committees, to world leaders in Davos, to podcast audiences of millions and to anyone who would listen.

That's how seriously he takes this and it's how seriously you and I need to take it too. Not because panic helps, but because the window is open right now and windows close. Five years, the METR data says five years, Bengio says five years, 30 countries backed the international AI safety report that says five years, let's use them well. So thanks for watching or listening, if this was useful share it with someone who needs to understand this moment.

Join the AI Profit Boardroom in the description below or go to the AIProfitBoardroom.com where we break down what AI developments mean for your life, your career and your business every single week and the practical ways to use them. And subscribe because the pace of this stuff is only accelerating.