Artificial intelligence has moved faster than both the expectations set for it and the rules written to govern it, according to the third report from AI100, the One Hundred Year Study on Artificial Intelligence.
The report, released by Stanford's Institute for Human-Centered AI (HAI), is the project's five-yearly check-up. AI100 was launched in 2014 with an endowment from Eric and Mary Horvitz and is administered out of Stanford HAI by an international standing committee chaired by Sheila McIlraith of the University of Toronto. The 2026 study panel was led by Michael Wooldridge of the University of Oxford and draws on computer science, philosophy and sociology.
Its headline finding is that the era of pure scale is ending. High-quality training data is increasingly scarce and the cost of training frontier models has become prohibitive. The next phase, the panel argues, will be defined less by bigger systems than by smarter algorithms, more efficient models and deeper reasoning.
Disruption is spreading to fields long considered immune: creative industries, software development and even pure mathematics -- a parallel the authors draw to what technology did to journalism, music and retail. In science, AI is compressing research timelines from years to months, and the panel expects AI-powered laboratories to become the norm within a decade, though it warns that validating machine-generated discoveries will require new methods.
The report also documents an investment boom of unprecedented size. Data-center construction is straining energy supplies, pushing up carbon emissions and reshaping national energy policy; chipmaker Nvidia briefly became the most valuable company in history. Yet the economic payoff is not yet visible in aggregate data or in companies' returns on investment -- a gap the panel flags as a possible financial bubble.
Geopolitically, frontier development remains concentrated among a handful of U.S. and Chinese actors. American companies lead on the most powerful proprietary models, while Chinese open-weight models are rapidly narrowing the gap, leaving other countries to choose between American systems, Chinese models or sovereign capability of their own.
Two warnings stand out. In cybersecurity, AI now cuts both ways: systems have identified software vulnerabilities and carried out attacks using teams of agents. And in warfare, the report says lethal targeting decisions have already been made without human oversight, calling for new international rules of engagement.
The panel closes with the public it says has been left out. Billions of people are affected by AI yet have had little say in how it is built, and international governance -- framed increasingly as national security rather than global cooperation -- has produced statements of principle but few enforceable commitments.

