I haven’t read this yet but if the title is accurate, I’m not surprised. Honestly, it’s felt like it would only be a matter of time before you can google your name and get back profiles compiled by LLMs - sure there will be inaccuracies but anytime you’ve posted something, used your name on a Facebook post, sent an email that got logged somewhere public, etc, it will be sucked up and pieced together at astonishing rates. The time for safeguarding privacy in regards to LLMs was yesterday.

> If LLMs continue to get better and also become more lax with their RLHF rules, it’s possible that future LLMs could start to identify nonpublic figures, and that will cause issues without sufficient awareness and preparation.

You can already do this extremely well with PimEyss and (especially) FaceCheck (which has indexed a lot of Instagram and LinkedIn); make a tool out of them for an LLM to use, and it could likely automate all the boring parts of cross-referencing.

Seems like something they've been capable of since image support in ChatGPT?

I really doubt that Gemini was using the Fantastic Four poster itself when it said it recognized Ebon Moss-Bacharach as The Thing.

I noted nearly a year ago that Claude's system prompt specifies that it is face blind.

https://news.ycombinator.com/item?id=41369641