New Study: When Can You Trust AI's Answers?
Blindly trusting AI can lead to false beliefs—here's how to verify.
Generative AI is everywhere, but when should you trust what it says? A new paper argues that AI outputs are different from other sources. Unlike a calculator, which has a proven track record, AI models are general-purpose and don't have a stable reliability record. So you can't just assume they're right.
The paper calls this the 'Gettier machine' problem. If you just take AI's output as truth without checking, you might believe something that happens to be true but for the wrong reasons—like a clock running ten times too fast that shows the correct time by coincidence. The speed and fluency of AI make it seem more reliable than it is, inflating your confidence without improving accuracy.
The solution is a three-step process: acquire the information, understand it, and independently verify it. The key question to ask: 'Had this been false, would I have noticed?' If not, you need to check. This is especially important for high-stakes decisions.
The paper defends this view against objections: it's not too demanding (you only need to verify when stakes are high), it's not solved by AI getting better (the problem is structural), and it's not redundant with other checks. The bottom line: treat AI as a starting point, not the final answer.
- AI outputs lack a reliability track record, so you can't blindly trust them like a calculator.
- Using AI without verification can make you believe things that are true by coincidence, not because the AI is reliable.
- Always verify important claims—ask yourself: 'If this were false, would I have caught it?'
Why It Matters
Blind trust in AI can lead to costly mistakes; verifying key facts keeps you safe.