QWE AI Academy You’ll walk away knowing exactly what leaves when you tap thumbs – and a habit so giving chat feedback gives Anthropic your data only when you mean it.
People flip off “Help improve Claude,” then 👍 a sharp reply that still held client notes, a pasted key, or half a contract. Anthropic’s Privacy Center is direct about that path: the entire related conversation can sit in a secured backend for up to five years and may be used to train, after de-linking from your user or customer ID.
Pure opt-out posts stop one door too soon. The model-improvement toggle and the feedback path are separate.
When giving chat feedback gives Anthropic your data – decide first
Not “never click.” Click only on threads you’d ship to a multi-year store without wincing.
- Scan the whole thread, not the last bubble. Feedback pulls the related conversation. Early turns with secrets still count.
- Sanitize, then rate. Good answer stuck next to a sensitive paste? Copy the useful pattern into a fresh chat with dummy data, regenerate, thumbs the clean thread.
- Incognito for throwaway sensitive asks. Docs say Incognito chats aren’t used to improve Claude even when Model Improvement is on (same training article). Whether thumbs inside Incognito still opens the five-year feedback store isn’t spelled out. I treat those threads as don’t-rate territory.
- Claude Code: read which prompt you’re on. “How is Claude doing this session?” (Dismiss included) is rating + metadata only – docs say no transcript and it cannot train models. Optional “Yes, look at transcript” can upload for up to six months and is still described as outside generative training.
/feedbackand/bugare the five-year transcript channel. Different animals – see Claude Code data-usage. - Org owners: policy forbids training-adjacent share? Disable Rate chats / Thumbs feedback once under Organization settings → Data and Privacy (or Console data retention) and tell the team why. Consumer accounts can’t hide the icons.
Breathing room: one icon can outrank a privacy toggle you set on purpose. Product incentives and privacy docs pull opposite ways. Notice which control your finger is actually on.
Why the training toggle isn’t enough
Free, Pro, and Max (and Claude Code from those accounts) let you choose whether new chats help improve models. Commercial defaults lean no-train on routine inputs/outputs. Explicit thumbs or bug reports break both patterns – official feedback wording treats that click as a hand-raise.
What leaves with the hand-raise, as of current Privacy Center language: messages, custom styles, conversation preferences, model settings. De-linked from your ID first; not stitched onto your other chats. May support analysis, research, behavior study, and training where law allows.
Raw connector payloads (Google Drive, remote/local MCP) stay out of the feedback bundle. Anything you pasted into the message does not. Paste-in content is the real leak path.
August 28, 2025 consumer terms updates also tied the longer feedback retention story more tightly to Free/Pro/Max (and Claude Code from those accounts) – worth re-checking after any terms email via Anthropic’s consumer terms note.
Real-world example: the client-notes trap
Pro account. Model Improvement off. Status update draft. You paste a customer Slack paragraph – names, dollar figures. Claude polishes it. Muscle-memory thumbs up.
Eligible bundle now includes that Slack text. Sidebar-deleting the chat later does not unwind a feedback submission that already left. No public retract control shows up in Privacy Center or settings docs.
Next time:
- Paste sanitized only (“Customer A, Q3 renewal risk, ~$40k”).
- Take the rewrite.
- Want to reward the model? Thumbs the clean thread – or skip the icon and keep the win local.
- Need real names in context? Incognito, copy output out, don’t rate.
| Action | What may be kept | Training-relevant? |
|---|---|---|
| Thumbs up/down on claude.ai | Entire related conversation, up to 5 years | May train (explicit feedback) |
| Model Improvement on (consumer) | New/resumed chats, de-identified, up to 5 years | Yes, while setting is on |
| Model Improvement off | Ordinary product retention for chats | No – except feedback, safety flags, other explicit paths |
| Claude Code session rating only | Rating + metadata | Docs: cannot train models |
| Claude Code /feedback (transcript path) | Transcript path, up to 5 years | Treat as sensitive product-feedback data |
| Usage-policy flagged chat | Inputs/outputs up to 2 years; trust/safety scores up to 7 years | Safety / policy systems |
Figures above track Anthropic Privacy Center retention language and Claude Code data-usage pages; treat UI labels as “as of late 2025-2026 docs” because copy moves.
Pro habits that change outcomes
Pro tip: Keep one “public demo” project with zero secrets. When Claude nails a pattern worth reinforcing, recreate it there with fake data and rate that thread. You still signal quality. You don’t donate the real spreadsheet.
After terms mail, open data-privacy controls and confirm nothing flipped. On Team/Enterprise, verify Rate chats is off if compliance needs it. Reddit threads rediscover the thumbs path every few months – same mechanics, new panic – so assume new teammates never opened the Privacy Center.
Incognito behavior, consumer vs commercial training defaults, and how memory/search treat deleted chats sit beside feedback. They aren’t the same door.
FAQ
Does thumbs feedback override my “Help improve Claude” off switch?
Yes. Explicit feedback is its own path. The opt-out covers routine chat improvement, not the hand-raise.
I already clicked thumbs on a sensitive thread – now what?
No documented undo. Assume the submission can remain in the feedback store for the full retention window, de-linked from your ID. Rotate any secrets that sat in the clear, scrub copies you still control, and stop rating dirty threads. Team/Enterprise: ask an owner whether org-level thumbs disable is on so the next person doesn’t repeat your click.
Are Claude Code’s session surveys the same as thumbs?
No – and a lot of fear glues two prompts into one honeypot. The numeric session rating does not collect the transcript and, per Claude Code docs (plus public team clarification), cannot train AI models. A separate optional transcript share uses a shorter window and is still framed outside generative training preferences. /feedback is the heavier multi-year transcript channel. Read the prompt text. Don’t mash on autopilot.
Open one recent thread you’d rate. Five-second scan: would I email this whole transcript to a stranger for years? If not, skip the icon – or rebuild clean, then rate that.