A look at the privacy concerns raised by OpenAI's GPT-3, Google's LaMDA, Meta's OPT-175B, and other large language models trained on troves of personal data
Large language models are trained on troves of personal data hoovered from the internet. So I wanted to know: What does it have on me? Tweets: @garymarcus , @hypervisible , @eileenguo , and @zcobb Tweets: Gary Marcus / @garymarcus : LLMs are an incorrigible mix of misinformation and privacy invasion, yielding an untraceable stew of half-truth. great article by @Melissahei @techreview https://twitter.com/... @hypervisible : “If you've posted anything even remotely personal in English on the internet, chances are your data might be part of some of the world's most popular LLMs.” “...the problem is only going to get worse and worse over time.” https://www.technologyreview.com/ ... Eileen Guo / @eileenguo : Among the other impt (and sometimes funny) points this story raises, it also highlights for me the real difference that privacy law makes. https://twitter.com/... @zcobb : Speaking of problems that are “only going to get worse over time” without a massive shift in ethics and resources ... the more we rely on Large Language Models and the systems that use them, the more they will be hacked for malicious purposes. https://twitter.com/...
Context & Ripple Effects
This investigation picks up a thread that started with the draft paper co-authored by Timnit Gebru, which catalogued the risks of large language models in late 2020 and made training-data practices a subject of mainstream debate. Two years on, Eileen Guo's reporting turns that critique into a concrete question — what do GPT-3, Google's LaMDA, and Meta's OPT-175B actually hold about any given individual scraped off the open web?
First-order effects
- Anyone who has posted personal content in English online may find it embedded in these models' training corpora, with no practical way to inspect or remove it.
- OpenAI, Google, and Meta now face direct questioning about what personal data their flagship models contain and whether individuals can exercise any control over it.
Second-order effects
- Regulators have a concrete hook: as later coverage of OpenAI's exposure under GDPR shows, 'right to be forgotten' requests are nearly unanswerable when personal data is baked into model weights rather than stored in a database.
- Dataset publishers feel the same squeeze downstream — LAION-5B was pulled from download after researchers found thousands of CSAM instances in it, showing that unaudited web-scale datasets carry legal as well as ethical risk.
Third-order effects
- If regulators treat training corpora as personal-data processing, web-scraping stops being a free input and becomes a compliance cost that favors incumbents who can afford audit trails over smaller labs.
- The industry-wide question shifts from 'can we scrape the internet?' to 'under what permission boundary?' — forcing explicit licensing or consent mechanisms where none existed.
The trend: The boundary between publicly posted data and data usable for AI training is being redrawn by regulators and dataset audits, turning training-data provenance into a core compliance function.