Why do different AI models give different answers to the same question?

Because they aren't looking up an answer — they're each generating one, and they were built differently. Different training data, different architectures, different fine-tuning, different values baked in by their developers, and randomness in the generation process itself. Variation is the expected behaviour, not a malfunction. The specific sources of difference: 1. **Training data.** Each was trained on a different corpus with a different cutoff date. A model trained through 2024 and one trained through 2026 will disagree about anything that changed in between, and neither is 'wrong' relative to what it saw. 2. **Sampling randomness.** Generation involves picking from a probability distribution. Most products don't pick the single most likely token every time — a temperature setting introduces deliberate variation, which is why *the same model* gives you different answers to the same question on Tuesday. Set temperature to zero and outputs become much more consistent, though rarely perfectly so. 3. **Fine-tuning and alignment.** After pre-training, each developer shapes behaviour differently — how cautious to be, how much to hedge, what to refuse, how long answers should run, house style. These choices produce visibly different personalities and different willingness to commit to an answer. 4. **System prompts.** Every product wraps your message in hidden instructions you never see. Those instructions materially change the response. 5. **Capability differences.** On genuinely hard reasoning, models differ in ability. Sometimes one is simply better on this class of problem. 6. **Tools.** One may be searching the web and grounding its answer; another is recalling from parameters. That alone explains many disagreements about current facts. What to do with this, practically: - **Disagreement is a useful signal.** If three models agree, the claim is probably well-represented in training data. If they diverge sharply, treat the topic as uncertain and go verify. Cross-checking is a genuinely good habit for anything that matters. - **Ask for reasoning, not just conclusions.** You can evaluate an argument; you can't evaluate a bare assertion. Often one model's reasoning is visibly weaker, which resolves the disagreement without external research. - **For anything factual and consequential, verify against a primary source** regardless of how many models agree. Consensus among models reflects consensus in training data, which is not the same as being true. The underlying reframe: these are not oracles returning the answer. They're generators producing a plausible answer, and 'plausible' has many valid instances. Treating variation as suspicious is the wrong instinct — treating uniform confidence as reliable is the dangerous one.