And if you're publishing something in general, someone feeding it to a chat bot is really low on the scale of things to be concerned about.
Same issue with phones they transcibe and allow summarizing calls who knows which models they use.
And if anyone actually sold such phones, I'd be first to buy it. LLMs are godsend for dealing with daily communications, but this process is still mostly manual on end-user devices.
[ ] Check this box and we'll *totally* not use your data for training
They might as well have this box as well [ ] Check this box if you'd like to preemptively be paid $25 as the result of a future class-action lawsuit because we just used the data anywayTangent: I personally literally don't care that all these LLMs trained on a ton of copyrighted material. It is in my humble opinion a long-awaited dose of karma for the copyright lobby, whose own abuses over the decades are numerous. It has basically restored balance to a playing field where the public has had no rights and (mostly) publishing conglomerates have held all the rights and privileges.
That means stealing leading researchers info is probably great for them, whether it's intentional or not.
They have all the best ideas flowing through there servers in real time as they are hatched around the world. It would be tempting to have top scientists accounts posted to internal slack channels to monitor new discoveries.
Don't post confidential information where there's a risk of confidentiality being broken. That includes posting to people whom you aren't 100% sure of know and uphold your confidentiality requirements. Adults usually handle this by contracts with penalties that include financial damages, loss of employment, or prison time.
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.
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.