Why two AI answers agree, and are both wrong

Why two AI answers agree, and are both wrong

You asked one model. You were not sure. So you asked another, or you asked the same one again, and it said roughly the same thing. That felt like confirmation.

Agreement between AI outputs measures how much they share, not how likely they are to be right. Two models trained on overlapping material will make the same mistake for the same reason, confidently, in the same direction. Here is why that happens, why it feels so much like verification, and what an actual second opinion requires.

Why does asking twice not work?#

Because the second answer is not independent of the first.

Independence is the whole reason a second opinion is worth anything. Two doctors trained at different schools, with different patients behind them, disagreeing tells you something. Two doctors reading from the same textbook agreeing tells you the textbook is consistent.

Language models are much closer to the second case than most people assume. They are trained on heavily overlapping material and shaped by similar methods. Where that shared material is wrong, incomplete or out of date, they are wrong together, and the fact that they agree is a consequence of the shared source rather than evidence about the world.

Asking the same model twice is worse still. There you are sampling the same thing twice and calling the second sample a check.

The signal that should worry you#

When two answers agree far more than two experts would on the same question, that is a red flag rather than a green one.

Genuinely difficult judgements produce disagreement among people who know what they are doing. If a question is contested among humans and your models are unanimous, they are not resolving the disagreement. They are reflecting whichever position was most common in what they read.

The practical version: notice when agreement is too clean for the question. A unanimous answer to a hard question deserves more scrutiny than a split one, not less.

Why it feels like verification#

Three reasons, all of them psychological rather than technical.

It arrives in different words. The second answer is phrased differently, so it reads as an independent restatement rather than the same source twice.

It costs effort. You did something to check. Effort feels like diligence, and diligence feels like it should produce reliability.

It resolves discomfort. You asked because you were uncertain. The second answer removed the uncertainty, and relief is very hard to distinguish from confirmation.

None of these is stupid. They are how checking normally works, which is exactly why this failure is so easy to walk into.

What an actual second opinion requires#

Something that does not share the first one's source.

Ground it in your own material. Give the model the actual document, contract, dataset or record and ask it to answer from that alone. An answer grounded in your source can be checked against your source. An answer from training cannot be checked against anything.

Plant a known answer. Include something you already know the answer to, and see whether the check gets it right. This is the only method that distinguishes a working check from a confident one, and it takes minutes.

Ask for the disagreement, not the answer. Instead of "is this right", ask what the strongest argument against it is, what would have to be true for it to be wrong, and what the answer would be if a specific assumption were reversed. This produces something useful because it does not ask the model to grade itself.

Ask a person. For anything genuinely contested, the only independent source is somebody with different experience. This is unfashionable advice and remains correct.

I have an answer from an AI and I want a real check on it, not agreement.
Do not tell me whether the answer is right.
Instead:
1. State what this answer assumes that it does not say out loud.
2. Give me the strongest argument against it, made as well as you can make it.
3. Tell me what would have to be true for this answer to be wrong.
4. Tell me what part of this you would not be able to verify without a source I have not given you, and name what source you would need.
5. Tell me which part of this you are least confident about and why.
Do not soften any of these. If you think the answer is fine, still do all five.

The version of this that costs the most#

The dangerous case is not two chat answers agreeing. It is an automated check, built carefully, running on the same model family as the thing it checks, reporting that everything is fine.

That setup produces a stream of confident, specific, mutually consistent findings that all share one blind spot. It reads as thorough. Consistency across many items is precisely what makes it credible, and it is the same property that should have made it suspect.

Ahead pairs you with an expert operator who looks at your own setup, in your own tools. If something in your workflow checks its own work, that is a good first hour.