The One-Hundred-Year Study on Artificial Intelligence (AI100) is an effort to understand and anticipate how the effects of AI will impact how people work, live and play. The study grew out of the 2008–09 Asilomar gathering convened by AI researcher Eric Horvitz. Eric and Mary Horvitz endowed it on a simple premise: you can’t understand what a technology does to people’s lives by watching it for one quarter. You need distance.
Peter Stone, now head of The University of Texas at Austin’s School of Computing, was recruited in 2015 to chair the very first study panel for AI100. The brief left plenty of room. “They wanted it to be relevant for our time and to set a precedent for future studies,” Stone says. “So, it was very open-ended.”
The report that emerged in the fall of 2016 set the template for what followed. Stone later chaired AI100’s standing committee through the 2021 report. With the launch of the third, AI Goes Mainstream: The Scaling Era and the Rise of Generative AI, the series now offers a decade of evidence on both technical progress and the choices it forces.
A Forecast That Holds Up, and a Clock That Runs Slowly
Reading the 2016 report against today’s world, Stone says much of its framework has held. The panel described a path from accumulating data, to predictive models, to systems that make or advise on decisions. It expected trust, privacy and regulation to define the adoption of AI in healthcare, and it recognized the economic stakes of automation early.
The area where reality ran at a different pace was transportation. The panel anticipated that by 2020, automated and semi-autonomous vehicles would be widely adopted across personal transport, delivery and public transit. Stone sees that as part of a very long pattern in the field.
“Throughout my entire career, people have been saying that all cars on the road will be autonomous within five years,” he says. “And it’s always five years. Five years always seems infinitely far away. I think that one did move more slowly than we stated.”
Today, the picture looks different. Autonomous vehicles are on Austin’s roads, and the public is being asked to trust them. For Stone, that is exactly the gating factor the panel pointed to a decade ago. “I’ve always said those were more the bottlenecks than the technology,” he says, referring to regulation, insurance and human trust. “I think they’re being worked through now. There are liability frameworks for Waymo and Tesla’s Cybercab, and those needed to be built up over time.”
Even when the engineering works, accountability is a hard problem. Stone frames it as a thought experiment. Human drivers are involved in some 40,000 fatal crashes a year in the US; imagine autonomous vehicles cut that to 5,000.
“I’ve always said we should take the utilitarian perspective, but that’s tricky legally,” he says. “Now one company is responsible for 5,000 deaths instead of 40,000 different people. And those 5,000 aren’t a subset of the 40,000, so every one of those families has a legitimate claim to say, ‘If a human had been driving, my relative would not have been killed.’ How does society deal with that? Is the utilitarian argument convincing, or is it just hard to get there from here?”
The Turn Few Expected: Language
If transportation moved on a longer timeline than anticipated, language moved much faster than almost anyone anticipated. That included the panel, and, Stone suggests, the field at large.
“We could already see in 2016 the power of data and computation for vision,” he says. “Computer vision had changed in a lot of ways by 2016, but what happened with natural language processing and language was very surprising.”
The 2026 report charts the result: general-purpose assistants, AI coding tools and agentic systems that execute multi-step tasks, powered by what the report calls the scaling era. Yet Stone, who works in robotics and reinforcement learning, resists the narrative of sudden revolutions.
“To the outside world, it sometimes looks like nothing happens for a while, then there’s a big step change, and then nothing happens again for a while,” he says. “I see it more as a gradual, maybe bumpy process, sometimes high slope, sometimes low slope, but all through the scientific process.”
New tools, he says, mean “we can ask different questions and try different methods.” But “a lot of the research that we were doing back then is still relevant today.”
Work, Education and Who Shares the Gains
Several questions from 2016 remain open, beginning with work. The first panel considered the possibility that automation would displace workers, and it warned that the gap between rich and poor could widen. It argued it was “not too soon for social debate on how the economic fruits of AI technologies should be shared.” That, Stone says, is “still very salient.”
Education has followed a different course. Massive open online courses, which many expected to blur the line between classroom and self-paced learning, “didn’t really take off in the way that they might have,” he says. He attributes that to the technology itself rather than to the forecast. Other conclusions from the first report still stand: “We did say that we’d still need engaged human teachers for quality education, and I think that’s still true.”
The Question That Shouldn’t Be a Question
Where Stone’s tone sharpens is regulation.
“It’s ridiculous that people are still asking whether AI should be regulated,” he says. “We regulate food, we regulate transportation. People say AI could be among the most powerful technologies ever built, and in the same breath ask whether it should be regulated at all.”
For him, the live questions are who regulates and how. “Should there be an external regulatory body? And how do you incentivize companies to be more cautious and responsible in how they deploy and roll out their models?”
Behind that sits a research problem the 2026 report also highlights: the limits of the guardrails developers use to constrain model behavior.
“If one decides that an AI model should never share someone’s private information, is it technically possible to build one that follows that rule? I think that’s still an unanswered question.”
The liability question follows directly. “What should the liability be if that isn’t done? If a model does something that society has determined isn’t okay, how does that get punished, and how do you incentivize the companies? Those are really big, burning questions right now.”
The geography of governance adds a new wrinkle. State-level rules, such as those on AI companions, make some sense, since that is often how legislation develops in the US. But models aren’t deployed state by state. “Can a model behave differently in Texas than in California?” Stone asks. “That’s another technical question.”
Fluent in Both
That entanglement of engineering and policy is why Stone keeps returning to a recommendation from the 2016 report: put real AI expertise inside government. Progress has been made, he says, but the need remains. “It’s still a really important recommendation: we need more people in government who are both AI experts and experts in policy and regulation.”
A decade on, those people have more powerful systems to work with and far more consequential decisions to make. The new AI100 report brings those choices into focus, from the changing nature of work and education to who has a voice in AI’s future. Read alongside its predecessors, it’s a chance to ask whether our institutions can evolve as quickly as the technology.