x.com/natolambert/status/2097320558559895862
1 correction found
0% chance any researchers at OpenAI or Anthropic look directly at specific user's data
Anthropic has publicly said its researchers have analyzed de-identified responses from specific Claude users, and its support docs say a small number of personnel involved in model training may review chats users share for improvement. So the chance is not literally 0%.
Full reasoning
This claim is too absolute.
Anthropic has publicly described at least two situations that contradict a literal "0% chance":
-
In Anthropic's own "What 81,000 people want from AI" writeup, the company says users were told their responses would be used for research, and then states: "All responses were de-identified before being analyzed by a small team of researchers at Anthropic." That means Anthropic researchers did in fact analyze data originating from specific users' conversations, even if identifying details were removed first.
-
In Anthropic's help article "I would like to input sensitive data into my chats with Claude. Who can view my conversations?", Anthropic says that when users allow chats to be used to improve Claude, "Access is limited to a small number of personnel involved in model training" and that safety-flagged conversations "may still be used to improve our internal trust and safety models ... or advance our safety research."
Those disclosures show that, under some circumstances, Anthropic personnel conducting research or model-improvement work can review user-originated conversation data. The data may be de-identified and access-controlled, but that is different from saying there is a 0% chance researchers ever look directly at specific users' data.
Because the post makes a categorical claim about OpenAI or Anthropic, a documented Anthropic counterexample is enough to make the statement inaccurate.
2 sources
- What 81,000 people want from AI | Anthropic
Before choosing to participate, users were informed their responses would be used for research ... All responses were de-identified before being analyzed by a small team of researchers at Anthropic.
- I would like to input sensitive data into my chats with Claude. Who can view my conversations? | Claude Help Center
Access is limited to a small number of personnel involved in model training ... If our safety classifiers flag your conversations, they may still be used to improve our internal trust and safety models ... or advance our safety research.