Stanford's Big Yearly AI Report Card Is Out — and the Trust Gap Is Widening

Once a year, Stanford’s Institute for Human-Centered AI publishes the closest thing the AI world has to a report card: the AI Index. The 2026 edition landed this week, and the headline isn’t that AI got better — it clearly did — it’s that the public’s trust in it didn’t keep pace, and in some ways went backward.

Start with the good news: models keep getting more capable, cheaper to run, and more widely deployed across industries. But a few numbers stand out on the other side of the ledger. Documented AI incidents — real-world harms or failures tied to AI systems — rose to 362 in 2025, up from 233 the year before. On a new, tougher accuracy benchmark testing 26 top models, hallucination rates (when a model confidently states something false) ranged from 22% up to a striking 94% depending on the model and task. Some previously strong performers dropped sharply: one widely-used model’s accuracy fell from over 90% to just 14.4% under harder, more realistic conditions.

Transparency — how much labs actually disclose about their training data, compute, and how a model behaves once deployed — got worse too. After two years of improvement, the average score on the Foundation Model Transparency Index dropped from 58 to 40 in 2025. The report notes an uncomfortable pattern: the most capable models often disclose the least.

Then there’s the public. In the US, only 31% of people say they trust their own government to regulate AI well — the lowest of any country surveyed, with the EU trusted more than either the US or China. And the gap between experts and everyone else is stark: 73% of US AI experts think AI’s effect on jobs will be positive, versus just 23% of the general public.

What’s actually going on: capability and trust have decoupled. Models are objectively better at benchmarks every year, but benchmarks measure narrow, well-defined tasks — real-world use is messier, and that’s exactly where hallucination and reliability problems still show up. Meanwhile, as AI companies race to ship faster, disclosure about how these systems actually work has quietly slipped down the priority list, even as public skepticism grows for good reason.

What this means for you: if you’re new to AI tools, this is a useful reality check against the hype — even top models get things confidently wrong more often than most marketing suggests, especially outside narrow, tested use cases. If you already lean on AI for real work, treat outputs in open-ended, consequential situations (medical, legal, financial) with real skepticism and verify independently — the report’s own numbers back up double-checking rather than trusting by default.

Sources

• The 2026 AI Index Report (Stanford HAI)
• Inside the AI Index: 12 Takeaways from the 2026 Report (Stanford HAI)
• Stanford’s AI Index 2026 shows rapid progress, growing safety concerns, and declining public trust (The Decoder)