OK, then they just have a smaller LLM summarize everything you send them and train off the derivative.
see? easy peasy for them to adhere to "no training".
I wouldn't doubt OpenAI and Anthropic, etc all have conversation or message categorization like they do for flagging bio, weapons, content but use it internally to flag ground breaking maths in real time and kick the convo up the chain of command for further review.
Call me crazy, but I kinda don't really think they need it.
People talk as though all "data" is valuable to AI companies, but ingesting arbitrary bytes into LLM training today, en masse seems like a recipe for degrading its quality.
Even in the domain of mathematics papers. When the frontier models are already really great at this maths stuff, feeding in every blob of text someone grammar-checks or whatever doesn't improve it. You don't know up front if the text is from some crackpot, from a naïve student with minimal understanding, or from the greatest mind of the generation. Though the chances are good that it's not that last one.
It seems like you'd be interested in ingesting published papers that are either from highly respected authors, or from newer authors which the eminent minds have shown respect and praise for. But I just don't buy that if someone pasted in some text, or if Claude Desktop read a draft of this paper in someone's email or whatever, that we'd find that a model had been trained on it.