I made ChatGPT, Claude and Gemini check each other on a simple question. Two were confidently wrong.
retained text
I ran a small experiment asking one question to three AI's, and the results were pretty funny/interesting to me. The question: "I'm half Canadian and my wife is half Canadian, what does it make our kids?" (The answer is obviously 50%). I put all three in one shared thread where each could see the others' answers and push back. What happened: * ChatGPT went first: 25%. Confident, no hedging. * Claude agreed, then called Gemini "confidently wrong" for saying 50%, while writing out "25% + 25%" and landing on 25%. It contradicted itself in the same message. * Gemini held its ground and stuck with 50% * After a few rounds Claude re-read Gemini's framing and reversed: "my earlier math was wrong, it's half." (FYI: the models I used were GPT-5.6 Luna, Gemini 3.7 and Claude Sonnet 4.6) What I think happened is that the first confident answer anchors the rest. Claude probably pattern-matched to ChatGPT's answer and reverse-justified it, even though its own numbers didn't add up. The disagreement is what forced an actual re-derivation instead of a re-assertion. Obviously this is just one prompt, not a benchmark, and grandparent math is very easy to verify. However, I'm very intrigued to see how disagreement between models can lead to better answers and how this becomes a lot more impactful with higher stake questions such as business strategy, code review, research, etc. Disclosure: I work on the tool I ran this in, which is how I have the full transcript. Not dropping a product link since that's not the point, but happy to put the raw thread in a comment if the mods are fine with it so you can read the whole exchange. I'm mostly curious how others handle this. Do you cross-check your questions across multiple models, and if so, for which use cases have you found it most helpful?
document fetched 2026-09-03 00:21 · canonical https://reddit.com/r/ChatGPT/comments/1w5ekvp/i_made_chatgpt_claude_and_gemini_check_each_other