26.7%. That's the share of people who say their biggest concern about AI is that it makes things up.
Anthropic surveyed 80,508 people across 159 countries in 70 languages, and the results flip the usual AI anxiety narrative on its head. The top fear isn't mass unemployment. It's hallucinations - the AI industry's polite term for when a model confidently presents false information as fact.
The Numbers Tell a Different Story Than the Headlines
Here's how the concerns ranked:
- AI unreliability (hallucinations, fake citations, inaccuracies): 26.7%
- Economic impact (job displacement, wage stagnation, inequality): 22.3%
- Loss of human agency (maintaining autonomy and control): 21.9%
The gap between "it lies to me" and "it'll take my job" is 4.4 percentage points. That's not a rounding error. People who actually use AI tools daily have apparently concluded that an unreliable assistant is a more immediate problem than a future where AI replaces them.
There's an interesting tension in the data, though. Anthropic notes that "concern about jobs and the economy was the strongest predictor of overall AI sentiment." So while fewer people rank job loss as their top worry, the people who do worry about it tend to be the most negative about AI overall.
81% Say AI Is Actually Helping
The survey wasn't all anxiety. A full 81% of respondents said AI is helping them make progress toward their goals. Only 18.9% said it failed entirely. Among those getting value from AI:
- 32% cited productivity improvements
- 17.2% use it as a thinking partner or collaborator
- 10% said they learned new skills through AI tools
Smaller groups mentioned emotional support and research synthesis as primary benefits.
These numbers suggest most people have found a practical use for AI despite its reliability problems. They're not rejecting the technology. They're frustrated that it keeps getting basic facts wrong.
The Verification Tax Is Real
Separate research puts a dollar figure on this frustration. Employees spend an average of 4.3 hours per week checking whether AI-generated content is accurate. That works out to roughly $14,200 per employee per year in verification overhead. For a 500-person company, that's $7.1 million annually spent on the question "is this actually true?"
This is the real cost of hallucinations, and it helps explain why reliability outranks job loss as a concern. People aren't using AI in the abstract. They're using ChatGPT, Claude, and Gemini to draft emails, summarize documents, and write code. When the output requires constant fact-checking, the productivity gains shrink fast.
For AI tool makers, the message is clear: users will tolerate imperfect AI, but they're tired of being lied to with confidence. The companies that solve reliability first will win the users who are currently spending a full workday each week on verification.