A group of Stanford researchers got together a year ago and made their predictions. Not the hyped-up kind you see in tech news stories, but the careful, slightly doubtful kind that academics make when they’re keeping a close eye on an industry and starting to see cracks. It’s time to see how that prediction holds up now that 2026 has arrived. It turns out pretty well. Sometimes too well for comfort.
Stanford HAI’s main idea for its 2026 predictions was simple: the time of AI promotion is ending and a time of evaluation is beginning. Rather than “look what it can do” instead of “does it actually work, for real people, in real settings?” That change is taking place. If you walk through any big city right now, you’ll still see AI billboards—the slightly crazy ones that say AI will change every industry—but the talk in boardrooms has become much more cautious. Angèle Christin, a communication professor at Stanford, saw this coming a long time ago and predicted that AI would deliver more realistic results. She was right.
Perhaps the most realistic prediction of the group came from James Landay, co-director of HAI: there will be no AGI this year. He put that one with something more interesting even though it was safe. He said that businesses would finally say out loud that AI hasn’t made them more productive, except in certain areas like customer service and programming. Slowly, companies that spent a lot of money and now have investors asking where the returns are are admitting this.
Landay also brought up an issue that does not get enough attention: the rule of law for AI. As a way to keep data within their own borders, countries are building their own infrastructure, buying GPUs, and hiring local models. It’s not so much an ideological issue as a matter of practicality. No one wants to depend on a foreign provider for something so vital to the functioning of the country. He said that the investments in data centers in South Korea and the UAE were not unusual. Those were signs.
When it comes to science, Russ Altman’s predictions are more important. He was interested in what he called the “archeology” of high-performing neural networks. This is the work that is being done to open the black box and figure out not only what AI predicts but also how it does it. It’s not good enough to have the right answer in medicine without being able to back it up. A lot of AI startup pitches are coming to hospitals, and Stanford’s team is working on ways to properly evaluate them. Even though it sounds quiet, that work on building the framework could be the most important thing going on right now.
Erik Brynjolfsson’s guess seemed almost right on the mark. He wanted “AI economic dashboards” that would show in real time, down to the task level, which jobs are being lost and which ones are being gained. Early data from his work with ADP showed that people just starting out in fields that are affected by AI were making less money and finding fewer job opportunities. Not nearly as much attention is being paid to that finding as it should be.
The interesting thing about these predictions is not that Stanford was right all the time. It’s that in 2025, those who knew this technology best were already talking about rules, responsibility, and limits, while most people were still talking about the magic. There is a space that is slowly getting filled.
That gap will soon be closed, but it’s still not clear whether most people will see it as progress or disappointment. Most likely both. There is still a bubble. But, as Christin said, it’s not getting much bigger. Maybe that’s the most honest way to sum up how things are now.
