ChatGPT users don’t only give feedback through thumbs up/down or explicit feature requests. A lot of useful human input — product ideas, failure modes, research hypotheses, new use cases — already surfaces naturally inside conversations. Most of it never becomes a formal proposal, so it’s effectively lost.
OpenAI already has an explicit consent mechanism for letting conversations contribute to model improvement. A contributor layer could build on that same model, making a small subset of unusually valuable input attributable instead of purely one-directional:
conversation → model improvement
could become
conversation → valuable contribution detected → validation → attribution → reward
The reward doesn’t need to be cash — subscription months, API credits, or extra compute likely matter more to contributors while costing OpenAI far less than cash. Example: “An idea from one of your conversations contributed to a feature or research direction. We’ve added 3 months of Pro to your account.”
The hard part is attribution, and I don’t think it should be solved by scoring a contribution instantly. It could instead “vest”: marked evaluation pending, left to mature for weeks or months while the system checks how often the same idea appears independently elsewhere. Rare independent rediscovery is evidence of novelty — but rarity isn’t usefulness, so reward can’t simply be 1 / number of people who had the idea; that breaks immediately and would incentivize hiding good ideas. Usefulness, novelty, and authorship are three separate problems — I’d avoid collapsing them into one global “contributor score.” Past contributions can affect how much attention a new signal gets, without making a person automatically “more correct.” Over time, a contribution history also becomes a talent signal — spotting people who repeatedly notice real problems, without needing them to be researchers or engineers first.
Even a small pilot on a tiny fraction of conversations, with no promises attached, seems worth testing. And maybe some version of this already happens internally and users just never see the attribution layer — in some regions, I suspect regulators would be curious about those footnotes. Which is really the point: if conversations can already improve the system, the next step isn’t collecting more data — it’s making the rare, genuinely useful human input visible, attributable, and worth closing the loop on