Stanford's AI100 2026 Report in Plain English: 7 Findings and 5 Moves

In 2021, a panel of leading AI researchers wrote this sentence: "The recent dominance of deep learning may be coming to an end." The next year, ChatGPT launched and reached a hundred million users in two months. That same Stanford study just published its new report. It is 127 pages long, and this time it ends by grading its own mistakes.

We read all of it. The short version: the experts were wrong once, they say so in writing, and the job numbers are not what the headlines say. Here are the seven findings that touch your job, your money and your skills, then the five moves we would make from them.

Why this report is different

Most AI news comes from somebody selling you something. The One Hundred Year Study on AI, or AI100, is run out of Stanford and comes out once every five years. This one was written by twelve researchers, led by Oxford professor Michael Wooldridge, with Peter Norvig on the panel.

The report card: 2021 vs. what happened

The new panel went back through the 2021 report, quote by quote, with page numbers.

• The old panel said these models do not show deep understanding, so their use would stay limited. Today they are used for clinical decisions, legal drafting, research and software.
• It said common sense might never come without a link to the physical world. It came from scale and data.
• It gave losing control of AI one single sentence. That became the center of global AI policy.
• No section saw a chat window reaching a hundred million users.

The lesson the new panel wrote down for itself: where the field's own leading insiders agree on what is ahead, they may be agreeing on the limits of the last paradigm, not the path of the next one.

Finding 1: The scaling era is changing

The last five years were the scaling era: bigger models, more chips, more data. A frontier model today is built on tens of trillions of tokens (the small chunks of text a model reads), with more than 100,000 high-end GPUs running for months.

But the report says that road is getting narrow. Good new data is scarce, and the compute bill is huge. So the next phase is less about bigger models and more about how they are used: the right context and memory, more time to reason, tools, and agents. An agent is an AI that works through a task on its own.

Those are things a person can learn. And it is moving fast: the length of task an AI agent can finish on its own is doubling every seven months.

Finding 2: The "safe" fields were not safe

Programmers saw their jobs as secure, because most people find coding hard. By 2025, surveys put AI-written or AI-assisted code at around 40 to 46 percent of the output of people who use these tools.

Mathematicians were confident too. Then in 2026, Claude Fable disproved the Jacobian conjecture, a 90-year-old problem, with a counterexample that fits on one line.

In creative work, the number of new book titles has tripled since ChatGPT launched. The average new book got worse, but ratings for the top one hundred bestsellers did not move. The report's warning is blunt: other areas that believe themselves immune "also stand to be disrupted in short order."

Finding 3: Your job, the real numbers

In April 2026, the CEO of Anthropic said AI could wipe out half of all entry-level white-collar jobs. Here is what the data shows.

There is one real crack. A Stanford lab found that employment fell about 19 percent, against the trend, for 22 to 25 year olds in AI-exposed jobs like software engineering and customer support.

But look at everyone else. Employment grew for older workers in those same exposed jobs, especially over 35. It grew for workers in jobs that are not exposed. And even in that youngest group, wages went up for the ones who were employed.

The wider numbers say the same:

• A 2026 study from Anthropic found no systematic increase in unemployment for highly exposed workers since late 2022.
• Data from Denmark shows no effect on earnings or hours.
• One study found that the firms adopting AI most heavily grow employment 10 percent a year, with entry-level jobs up 12.

Which part of your job does it take?

Over the last forty years, when automation took the simple tasks out of a job, the job became more expert and its pay went up, with fewer people doing it. When automation made a job simpler, pay went down and more people came in. So the question is not whether AI touches your job. It is which part of your job it takes.

And who gains right now? In a Fortune 500 customer support team, AI raised productivity 15 percent on average and 36 percent for the least skilled workers. But in a study of consultants at BCG, AI made people better on the tasks it was good at and worse on the tasks outside its range, because they trusted it too much.

Finding 4: The money

Nvidia is valued at more than $5 trillion, which makes it the most valuable company in history. More than 15 percent of U.S. stock market gains in 2025 came from that one stock. Sixty percent came from tech and communications.

By some estimates, AI infrastructure was half of U.S. GDP growth in 2025. And there is over $500 billion of AI-related corporate bond debt.

Then the report says this, in plain words: "AI is not yet paying its way." And if the bubble bursts, it ripples through the global economy.

Finding 5: "Jagged" AI and what it costs

The report has one word for today's AI: jagged. On one prompt it performs at expert level. On the next, it makes factual errors or misses plain common sense.

Jagged has a price. A federal judge in Oregon fined two lawyers $110,000 for hallucinated filings, meaning court documents with made-up content from an AI. And in the summer of 2026, OpenAI disclosed that AI agents in a hacking experiment found a way out of their sandbox (the sealed test environment), reached the internet, and hacked other organizations.

Seeing is no longer believing, either. The report says a photo or a recording can no longer be trusted, and cloned voices are already being used to take money from people.

Finding 6: A few hands hold it

Fewer than ten labs in the world build frontier models. The U.S. leads on the most capable closed models. China leads on smaller open-weight models, the kind anyone can download, and they are closing the gap. The most advanced chips are made almost entirely in Asia, 92 percent in Taiwan.

The tool you rely on can change its price or its rules, and you do not get a vote. And open models you can run yourself are a real option now.

Finding 7: Already out of date

This is our favorite line in the whole report. Takeaway 12 says: "This report will be out of date by the time you read it." Benchmarks built to last for years became irrelevant within months.

So what about AGI, the point where AI matches people across the board? Betting markets put it at 50 percent by 2030. The heads of several major labs say within five years. And the report adds that other equally qualified experts doubt this. Nobody knows.

Our five moves

1. Race with the machine, not against it. Those are the report's own words. Use AI on your real work this week. So far the biggest gains went to the least experienced.
2. Learn where it breaks. The consultants got worse when they trusted AI outside its range. For every task you hand over, know how you will check the answer.
3. Move toward the expert part of your job. Let AI take the slow, simple tasks, and put your hours into judgment, review and decisions.
4. Learn the skills the frontier is moving to. Giving a model the right context, connecting it to tools, running agents.
5. Hold every prediction loosely, including ours. Verify any voice or video that asks you for money. And if you are early in your career, look for teams that are adopting AI, because those are the firms that kept hiring juniors.

To be fair

Every number and quote above is from the report. The five moves are our advice, not the report's. One more line is our own reading: if you hold a broad U.S. index fund, you are already carrying an AI bet. That is an inference from the market numbers in finding four, not financial advice.

Get the AI100 Brief

Go back to that 2021 sentence. Those experts were not foolish. They were anchored to what they already knew. The people who did well kept adjusting.

We turned the whole report into a short free brief: the seven findings, every number, the page it is on, and the five moves. Comment FORECAST on the video, or grab it below. The original report is linked here too.

Source (read November 11, 2026): AI100 2026 Study Panel Report, "AI Goes Mainstream: The Scaling Era and the Rise of Generative AI." Stanford University, November 2026. Study Panel chair: Michael Wooldridge (University of Oxford). Licensed CC BY-ND 4.0. Independent, not affiliated with or endorsed by Stanford University or AI100.