Collective alignment: public input on our Model Spec

OpenAI just published results from a survey it ran with more than 1,000 people across different countries, asking them to weigh in on how AI models should behave in specific, sometimes uncomfortable scenarios. The company then lined those responses up against its Model Spec, the internal document that defines default behavior for ChatGPT, and used the gaps to figure out where its own assumptions might be off.

This is part of what OpenAI calls collective alignment, an attempt to pull actual public opinion into the process of setting AI defaults instead of leaving those calls entirely to engineers and policy staff in San Francisco. The method isn't just a single opinion poll. Participants reacted to concrete situations, things like whether a model should refuse a request, hedge, or answer directly, and OpenAI tracked where the crowd's instincts matched the Spec and where they diverged.

The headline finding is reassuring for OpenAI: broad agreement with existing defaults was common. People largely backed the idea that models should stay helpful without being preachy, and that refusals should be rare and well-justified rather than reflexive. But the disagreements weren't noise. They clustered around genuinely hard territory, topics touching politics, self-harm, and questions where cultural background shapes what