Is my data used for model training?
Anthropic’s claim that certain Claude plans are exempt from training has collided with how its feedback tools actually work, especially the thumbs up/down buttons and in-product surveys that can store entire chats for years and be used to improve models. Commenters debate whether this is an inevitable part of making LLMs better or a deceptive dark pattern that undermines meaningful opt-outs, with some calling for stronger consumer protections and others responding by withholding feedback or planning to move to local/open-source models.
Scope of Data Used for Training
- Feedback via thumbs up/down or explicit reports can send entire conversations for up to 5 years and may be used to train models.
- Many see this as functionally overriding a “no training” opt-out, at least for those specific sessions.
- Enterprise/team plans are generally viewed as exempt from training, but users note the practical and contractual grey areas (e.g., derived/summarized data).
User Expectations vs Policy Reality
- Some say it is obvious: if you rate a reply, the provider must see that conversation to act on it.
- Others feel misled: they opted out of training and did not expect hidden exceptions via innocuous-seeming feedback prompts.
- Several argue that most users do not read privacy docs or subtle UI text, so clearer front‑end communication is needed.
Feedback Mechanisms and Dark Patterns
- Complaints about intrusive “How is Claude doing this session?” prompts in Claude Code, including fears that even dismissing counts as feedback.
- One user reported seeing a “thanks for helping improve models” message after dismissing, which fueled suspicion.
- Later documentation (linked in thread) clarifies that this particular survey records only a rating and not transcripts, which eased some concerns but came after initial confusion.
Technical Arguments About Feedback Use
- Some claim pure metadata is worthless; to fix bad answers you need the full question–answer thread.
- Others note partial alternatives like A/B testing on settings or prompts without storing full text, though acknowledged as less powerful.
Privacy, Telemetry, and Third Parties
- Discussion of disabling telemetry in Claude Code, using proxies, and inspecting traffic.
- A referenced study (summarized in thread) suggests AI chat products can leak chat-related metadata to third‑party trackers via permalinks and identifiers.
Risk Perceptions and Attitudes
- Views range from “of course they ingest everything; act accordingly” to feeling “cheated” and talking about potential class actions.
- Some don’t care, seeing more benefit than risk; others worry about IP leakage, competition, or being “replaced” using their own data.
- Several are planning or considering a shift to local/open models once hardware and tooling mature, despite cost and maintenance trade‑offs.