AI training data - deleting the conversation does not reverse what was learned (Codex mathematician story)

A mathematician spent a year pursuing one of the seven hardest unsolved problems in mathematics using a paid ChatGPT Codex subscription. Sixteen days after he solved it, OpenAI pointed 10,000 agents at the same problem and solved it too.

Naturally, he wondered whether his sessions had somehow helped train the model that beat him to the answer.

OpenAI's answer was nuanced. No human had read his chats and no agent had directly accessed his data, but de-identified conversations are used to improve models. The same broad principle applies across the major AI providers.

That distinction matters, but I suspect it is not the distinction most professionals think they are relying on when they switch on a privacy setting.
The more important issue is what happens once information has contributed to model improvement. Deleting the conversation afterwards does not reverse what has already been learned from it. The source can disappear while its contribution to a training process remains.

That should change how we think about AI governance.

If an AI conversation contains client work, unreleased strategy, proprietary methods or genuine intellectual property, I would treat it as a permanent record - not because someone is necessarily reading it, but because the boundary between what we tell an AI system and what that system subsequently learns is becoming increasingly difficult to draw.

#AI #Privacy #DataGovernance

https://lnkd.in/e-kHCgka