Yup. As someone who’s worked a little bit on GDPR compliance, it’s not some magic wand you wave at your data. Any data they receive after the request is also not covered by that request. Also, only EU citizens and residents are legally entitled to make a request. A company may choose to comply with non-EU users, but that’s purely their choice.

Comments that contain any info about where you live, your ethnicity, disabilities (cognitive or physical), gender, where you work, etc must be deleted as part of a forget request, so that might impact LLM training data.

Personally identifying information can be somewhat of a grey area in some situations as well. If I were to say I’m from New York, that’d be personally identifying. If I were to say I’m a fan of a sports team in New York, that’s not (even if that implies my location). If I were to say I’m a fan of a New York sports team, my favourite pizza place is in New York, my favourite park is in New York, etc etc, that might arguably be identifying, even if each of the pieces by itself is not.

EDIT: Oh, and I forgot one of the most important parts: it’s not like there are any spot checks or anything. You’d need someone to actually lodge a formal complaint, with some kind of evidence they haven’t done what they’re supposed to, and the procedures are different for every EU country. They are normally very involved and complex. Essentially, you’d need to lawyer up and care enough to slowly and painfully shove it through the legal system.

Well, that’d be the mechanism of how GDPR protections are actioned, yes; but leaving themselves open to these ramifications broadly would be risky. I don’t think it’d satisfy ‘compliance’ to ignore GDPR except upon request. Perhaps the issues with it are even more significant when using it as training data, given they’re investing compute and potentially needing to re-train down the track.

Based on my understanding; de-identifying the dataset wouldn’t be sufficient to be in compliance. That’s actually how it worked prior to it for the most part, but I know companies largely ended up just re-identifying data by cross-referencing multiple de-identified datasets. That nullification forming part of the basis for GDPR protections being as comprehensive as they are.

There’d almost certainly be actors who previously deleted their content that later seek to verify whether it was later used to train any public AI.

Definitely fair to say I’m making some assumptions, but essentially I think at a certain point trying to use user-deleted content as a value add just becomes riskier than it’s worth for a public company