Abstract

This disclosure describes a method of polling AI models alongside people, so that the answers of AI systems to questions that matter to the public can be compared openly with the answers of verified people, without ever being mistaken for or mixed with them. Each poll of verified people runs with a twin AI panel: a declared roster of credentialed AI models, each identified by provider, model, version and configuration, answers the same question with the same wording and options. The question is sealed until the poll opens; each model is queried many times, so that the spread of its answers is recorded rather than a single answer, by an independent runner working under a general method published in advance and exact settings fixed before the poll opens but sealed until the result, so that no provider can recognize the poll's queries and serve them differently; and each response is signed and committed to a registry before the human poll closes, while its content stays sealed. Each model also forecasts the human result before the close, and these forecasts are committed in the same way. No interim result of either panel is revealed. At the point of announcement, a single joint release publishes the human result, each model's responses, the models' forecasts and their scores, and the gap between each model and the people, and anyone can check that every model response matches a commitment made before the human result was known. Model responses are never combined with human results, never used to weight, adjust or fill in human results, and are always labelled as model responses. People remain free to consult AI before answering, but may not delegate their ballot to an AI, and each may report privately whether they consulted one, so that the influence of AI on people's answers can be measured. The official result is always that of all verified people counted equally, and every other figure, including the split between people who did and did not consult AI and each model's responses, is published beside it as reference only, never as a weight. Each model may also be asked again after the human poll closes, and any change from its opening answers is published as drift, revealing adjustment while the question was public. The gap between each model and verified people, the measured influence of AI on people's answers and each model's drift are accumulated across polls into running public records.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS