I didn’t set out to write this article. I set out to settle a small argument I was having with myself.

I asked ChatGPT to imagine a line with the AI doomers on one end and the AI advocates on the other, and then to tell me where I landed on it. It put me at about 72 percent toward the optimistic end. I’ll let you decide whether that’s reassuring or alarming. Being a curious sort, I then asked it to place a bunch of other people I follow in this field on the same line, and it produced a tidy little graphic with everyone sorted from dystopia to utopia.

The trouble is, the more I looked at that line, the less it actually told me. A single axis forces two very different questions into one answer, and once I noticed that, I couldn’t un-notice it. So I asked for a different graph. One axis for how likely someone thinks a genuinely catastrophic AI outcome is, and a second axis for how big they think the upside could eventually become. That version told a much richer story, and it’s the one that became this article.

Here is what the line graph was hiding. You can believe AI has the potential to cure diseases, accelerate scientific discovery, and create abundance we have never experienced before, and you can believe, at the same time, that there is a real chance advanced AI becomes hard to control. Those two beliefs are not in conflict. They are answers to two separate questions, and most of the interesting people in this field hold both at once.

That is what the two-axis version makes visible. The horizontal axis asks how worried a person is about the probability of a catastrophic outcome, something the researchers like to write as P(catastrophe), meaning the perceived probability of catastrophe. The vertical axis asks how large that person believes the ultimate payoff could be. Those two things move independently. Somebody can score high on both. Somebody else can think catastrophe is nearly impossible while also thinking the wilder promises about superintelligence are overblown.

There is one more wrinkle, and it matters. Ask someone whether they are optimistic or pessimistic about AI and you may be asking the wrong question. Not everyone means the same thing by “AI risk.” Some people are worried about a future superintelligence slipping out of human control. Others are far more concerned with what is already happening today: bias, surveillance, misinformation, job loss, environmental cost, and a great deal of power ending up in very few hands. Those are not the same fear, and lumping them together is how you get a graph that misleads you. So treat everything here as an estimate, not a measurement. There is no dataset that tells us what percentage of civilizations survive building something smarter than themselves. What follows is my read, informed by what these people have actually said and built, of where they sit and why.

The people who take both seriously

Start with the cluster that surprises people, the group that sits high on catastrophe concern and high on upside at the same time.

Geoffrey Hinton is the obvious anchor. He helped invent the neural network ideas that everything else is built on, which makes his warnings hard to wave away. He genuinely believes increasingly capable AI could deliver enormous gains in medicine, education, and research. He also worries that machines could eventually become smarter than we are and that we may not know how to keep control once they do. He has floated some uncomfortable probabilities for a bad outcome while being honest that these are judgment calls, not calculations. That honesty is worth respecting. Nobody has a spreadsheet for this.

Stuart Russell has spent years on what’s called the control problem. Ordinary software follows instructions. A capable autonomous system pursues goals, and the worry is what happens when the goal we hand it doesn’t quite match what we actually meant. The more capable the system, the more expensive that gap becomes. Russell isn’t arguing that advanced AI shouldn’t exist. Most of his work is about how to build it so it stays compatible with what humans want, which is exactly why he sits high on both axes.

Ilya Sutskever might be the cleanest example of the whole pattern. He co-founded OpenAI, then left to start a company he named Safe Superintelligence, and now runs it as CEO. The name is the thesis. He clearly thinks superintelligence is possible and worth building, or there’d be no point in trying, and he just as clearly thinks safety cannot be bolted on afterward. He isn’t afraid of intelligence. He’s afraid of building enormous capability without building the ability to steer it at the same time.

Nick Bostrom put this risk on the map with his book Superintelligence, which forced a lot of people to sit with an uncomfortable question about what happens if we create something far smarter than ourselves whose goals drift from ours. But his later work runs the other direction and asks what happens if it all goes right, if scarcity and disease and drudgery become far less central to human life. He takes both the enormous downside and the enormous upside seriously, which is the honest place to be.

Dario Amodei, who runs Anthropic, has one of the more interesting positions on the board. He spends a great deal of time talking about safety: biological threats, cyberattacks, authoritarian misuse, and the behavior of increasingly autonomous systems. Then you read his description of what powerful AI could do for medical research, biology, neuroscience, and human health, and it’s about as optimistic as anything on here. The concern and the optimism come from the same root. He worries precisely because he thinks the technology is going to become extraordinarily powerful. If it were just a better spreadsheet, none of the worry would be necessary.

Elon Musk is harder to pin down, mostly because he says both extremes on a regular basis. He has warned for years that AI could pose an existential threat, and in nearly the same breath he predicts that AI combined with humanoid robots could produce staggering abundance. Those aren’t necessarily contradictory. The fair summary is that he thinks the destination could be extraordinary and he’s not at all sure we’ll manage the trip without hurting ourselves.

A different kind of worry

Three people on this graph don’t fit the superintelligence framing at all, and pretending they do would misrepresent them.

Timnit Gebru isn’t mainly warning that a machine wakes up one day and ends civilization. She questions whether all the attention on hypothetical future machines is a convenient distraction from harm that’s already here. Her work centers on algorithmic bias, environmental cost, exploited labor, surveillance, concentrated corporate power, and inflated claims about what today’s systems actually understand. On the catastrophe axis she reads low. On the “humans using this irresponsibly right now” axis she reads much higher. That distinction is the entire point.

Joy Buolamwini comes at it from a similar angle. Her research showed that automated systems can inherit and amplify the inequities baked into their training data, which quietly demolishes the assumption that an automated decision is a neutral one. A model doesn’t become objective just because no human is visibly making the call. Humans picked the data, set the objectives, and chose where to point the thing. Her concern lives in the present, not in some post-human future.

Melanie Mitchell occupies a spot I find genuinely useful. She isn’t primarily arguing that AI is too dangerous. She’s questioning whether today’s systems actually possess the understanding we casually credit them with. A model can ace a benchmark without grasping much of anything the way a person would, and that gap matters enormously the moment you start extrapolating from today’s chatbots to tomorrow’s superintelligence. Her caution is less fear and more intellectual discipline, which this conversation could use a lot more of.

The builders and the practical optimists

Now the larger group, the people who put less weight on existential catastrophe and see somewhere between significant and enormous upside.

Yann LeCun has been one of the loudest critics of AI doomerism, but calling him an AI skeptic would be flat wrong. He thinks artificial intelligence will become extraordinarily powerful. What he doubts is that today’s large language models, the LLMs behind the chatbots, are anywhere close to the robust intelligence humans have. He argues the current approach is missing key pieces, including a real understanding of the physical world, planning, and persistent world models. He put his money where his mouth is, too. At the end of 2025 he announced he was leaving his role as Meta’s chief AI scientist to start his own company built around exactly that idea, so his low-catastrophe, high-upside position now comes with a startup attached to it.

Andrew Ng likes to call AI “the new electricity,” and that line tells you where he stands. Electricity didn’t become one industry, it became the thing sitting underneath thousands of them. Ng sees AI the same way and stays relentlessly practical about it: what can businesses build, what can workers do with it, how do we get these capabilities into more hands. He has little patience for extinction arguments and a lot of enthusiasm for what you can actually do with the technology today.

Daphne Koller is a useful reality check on the single biggest AI promise, curing disease. Her company uses machine learning and large biological datasets to speed up drug discovery, so she plainly believes AI can transform our grasp of biology. But she also knows biology isn’t just a software problem. AI can surface patterns a human would never find by hand, and then you still have to run the experiments, validate the results, and prove it in the clinic. That lands her firmly optimistic without requiring the fantasy that superintelligence solves medicine overnight.

Fei-Fei Li helped build the foundations of modern computer vision and now champions what she calls human-centered AI. The phrase carries weight. The goal isn’t to replace as much human activity as possible, it’s to expand what people can do, with people kept at the center. She sees enormous potential in medicine, science, and education, but the technology is meant to serve human aims rather than force humans to rearrange themselves around whatever the machine happens to be able to do.

Bill Gates describes his own outlook as “optimism with footnotes,” which might be the most accurate self-label on the whole graph. He expects real gains in health, education, science, and productivity, and he expects real disruption right alongside them. Jobs change. Some work disappears. Bad actors get new tools. Institutions scramble to keep up. He doesn’t think either outcome is automatic, and that’s the footnote: human choices decide most of what happens between the good ending and the bad one.

Jensen Huang, who runs NVIDIA, sees AI mostly through an industrial lens, and his logic is a straight line. AI becomes infrastructure, infrastructure creates productivity, productivity creates industries, industries create wealth. He also draws a distinction I keep coming back to, between the task and the purpose. AI may automate the task of writing code, but innovation is still the purpose. It may read the scan, but caring for the patient is still the purpose. That puts him very high on upside and largely uninterested in the AGI philosophy debates.

Sam Altman, who runs OpenAI, builds his vision around abundant intelligence, the idea that if intelligence becomes cheap and widely available, things that used to require scarce expertise could reach billions of people, across education, medicine, research, and entrepreneurship. He sits near the top of the upside axis. He is not, however, at zero on concern. The heavy investment his company pours into safety and alignment says as much. His position, more or less, is that the upside is big enough to be worth the trouble of solving the risks.

Ray Kurzweil is the purest optimist on the board. He has spent decades arguing that progress compounds exponentially, and his predictions run well past better software into radical medical advances, dramatically longer lifespans, human-machine integration, and eventually a technological singularity, the point where progress accelerates beyond our ability to keep up with it. He grants that powerful technology carries risk, but he fundamentally expects us to come through it far more capable than we are now. Of everyone here, his expected upside is probably the largest.

Then there are the newer voices, the ones actually building businesses on this stuff rather than theorizing about it, and their vantage point is worth having.

Mira Murati sits between the big camps. She was OpenAI’s chief technology officer through the launches that made the company famous, and she left in 2024 to found Thinking Machines Lab, where she’s now co-founder and CEO. Her stated philosophy is that AI should extend human agency rather than take it over, that people should be able to steer and customize these systems rather than simply hand their decisions off to a machine. That keeps her high on upside while holding on to more concern than most of the pure optimists.

Julia McCoy makes one of the boldest claims on the graph. She talks about a possible post-labor economy, where AI handles a growing share of economic production and traditional employment stops being the center of how people survive. That’s a very large claim, because it says AI doesn’t just improve work, it changes why we work at all. She tends to frame that shift as an opening, a chance to spend more time on creativity, relationships, and the distinctly human things, which puts her very high on upside and low on catastrophe.

Dustin Stout, Joel Comm, and Austin Armstrong all land in roughly the same neighborhood, the builders who see AI mostly as leverage. Stout’s point is that using AI just to replace people usually misses the bigger prize, which is making an already capable person dramatically more capable by pairing machine speed with human judgment. Comm looks at it from the perspective of entrepreneurs and workers trying to adapt, and lands on empowering people rather than clearing payroll. Armstrong builds directly with these tools for content and marketing and treats AI as a thinking partner, with the human still supplying strategy, taste, and knowledge of the audience. None of them are waiting around for a post-human future. They’re low on catastrophe and high on what a person can get done today.

Which brings me, a little awkwardly, to my own spot on the graph. I land low-to-moderate on catastrophe and high on upside, and my reasoning is simple enough. If you look at AI as a way to replace human effort, you’ll be afraid of it. If you look at it as a way to amplify human potential, you’ll be inspired by it. I think it can absorb an enormous amount of repetitive research, administration, analysis, and production, and hand people back the leverage to spend their time on judgment, creativity, relationships, and leadership. Done right, it moves us from being human doings back toward being human beings.

I’m also skeptical of confident predictions about when AGI, Artificial General Intelligence, actually shows up, and I don’t buy the assumption that more intelligence automatically drags human motives like greed or self-preservation along with it. My worries sit somewhere else entirely, on what people choose to do with these tools: privacy, surveillance, manipulation, misinformation, environmental cost, intellectual property, and, yes, the concentration of power. Which, you’ll notice, puts me closer to Gebru and Buolamwini on that particular question than to the doomers, even though I’m an optimist. That is the whole argument for two axes instead of one.

What the two axes actually show

The interesting part of this graph was never where any one person landed. It’s the clusters.

Hinton, Russell, Sutskever, Bostrom, Amodei, and Musk all put real weight on catastrophic risk while most of them simultaneously expect extraordinary benefits. LeCun, Ng, Koller, Li, Huang, Altman, Kurzweil, McCoy, the builders, and I generally assign lower odds to existential catastrophe while still seeing large upside. Gebru, Buolamwini, and Mitchell introduce a different dimension altogether, less “will a superintelligence kill us” and more “is today’s technology being oversold, deployed carelessly, and governed unfairly.” Gates and Murati sit in the middle, optimistic about the payoff while holding on to real concern about disruption and control. And Altman occupies his own odd corner, expecting about as much upside as anyone while carrying enough safety concern that you can’t honestly call him unconcerned.

That is why “are you an optimist or a pessimist about AI” is close to a useless question. A better conversation starts with two: how powerful do you think this technology can ultimately become, and how much danger comes with creating that power. The answers tell you far more than any single label ever could.

The part the graph is really about

There’s one more question hiding under all of this, and it’s the one I keep landing on. How much of the outcome is actually up to the AI?

The upside depends partly on what the technology turns out to be capable of. The risk depends heavily on us. What do we build, what do we connect these systems to, how much authority do we hand autonomous software, who controls the infrastructure, who profits, what safeguards we put in place, how we protect privacy, how we handle misinformation, and who stays accountable when a system makes an expensive mistake. None of those are questions the AI gets to answer for us. They’re ours.

AI doesn’t have to be our replacement, and it doesn’t have to be our savior either. It can be something much more useful: a tool that makes people capable of more than they were before. And maybe that’s what this graph is really measuring. Not whether AI will be good or bad, but whether we can pursue extraordinary capability without becoming blind to extraordinary risk. The real question isn’t how intelligent the machines become. It’s whether we become wise enough to use that intelligence well.