QWE AI Academy The switch myth is comforting – and incomplete
Most advice on “is my data used for model training” treats it like a light switch: flip one toggle and your chats are safe forever. That framing fails. Training is only one pipe. Retention windows, safety review, feedback loops, and which account you’re logged into drive most of the real exposure – and defaults still move.
Anthropic’s 2025 consumer shift (plus the Reddit threads that followed) forced a plain gap into view: “business plans don’t train by default” is not what Free/Plus/Pro accounts do. As of early 2026, here’s what the controls actually do – one setting never finishes the job.
Think of the training toggle as locking the front door while the windows stay open and the mailbox is unlocked. Door first. Then the rest.
Training is not retention is not memory
Three different things get mashed into one sentence in most writeups.
Training – your prompts and outputs can help build or fine-tune future public models.
Retention – how long a copy sits on company servers for safety, abuse detection, or legal reasons, even when training is off.
Memory / personalization – facts the assistant keeps about you for later chats. Kill training and memory usually still runs.
Temporary Chat on ChatGPT skips training, history, and new memories while it stays temporary; safety retention can still run up to about 30 days (OpenAI Data controls). Claude Incognito stays out of model improvement even if the main toggle is on. Gemini temporary chats plus Keep Activity off both block training use unless you send feedback – data can still sit roughly 72 hours. Safety-flagged content can still get analyzed across providers.
Pro tip: Flip only the training toggle, then paste client code or medical notes into a normal chat, and you’ve fixed the marketing claim – not the exposure. Temporary or incognito for anything you wouldn’t email a stranger.
Step-by-step: lock down the big three
Do this on the accounts you actually open every day. Settings ride the signed-in account across devices; signed-out is browser-local.
ChatGPT (personal Free / Plus / Pro)
- Profile menu → Settings → Data controls.
- Turn off Improve the model for everyone and confirm. Same outcome via privacy.openai.com → “Do not train on my content” for chats and Codex-style tasks on personal plans.
- Sensitive one-offs: Temporary Chat from the new-chat dropdown. No training, no history/memory while temporary; save it later and normal rules apply.
Voice mode ships its own sharing control – the global chat opt-out does not swallow it. Business / Enterprise / Edu / API workspaces already default to no training on inputs and outputs, per OpenAI’s how-your-data-is-used page.
Claude (Free / Pro / Max)
- Name menu → Settings → Privacy (or claude.ai/settings/data-privacy-controls).
- Toggle off Help improve our AI models. New chats and coding sessions under those consumer accounts stop feeding future runs; previously stored chats also drop out of future training. Models already trained stay as they are.
- Incognito when the main setting should not matter at all.
Allow improvement on consumer plans and retention of new/resumed chats can stretch to up to five years in de-identified training pipelines; leave it off and you’re closer to a ~30-day window (as of Anthropic’s 2025 consumer terms update). Commercial Claude for Work / API stays off training by default unless you explicitly feedback or join a partner program. Official split: Is my data used for model training? and the consumer terms update.
Gemini (personal)
- Gemini app or gemini.google.com → Activity / Gemini Apps Activity (myactivity.google.com/product/gemini works too).
- Keep Activity off (or off + delete existing). Future chats are not used to train generative models unless you submit feedback.
- Temporary Chat for one-off sensitive sessions – they skip training.
Off still means ~72 hours on hand for response and safety. A subset that went to human review can sit up to three years, disconnected from your account. Workspace/business Gemini is not trained on without org permission. Audio and Gemini Live? Separate improve toggle, off by default, and it wants Keep Activity on – global chat opt-out does not auto-cover them (Gemini Apps Privacy Hub).
Common pitfalls that still bite
Feedback is the one that keeps burning people.
On personal ChatGPT, thumbs up/down after opt-out can still put the whole conversation into training. Anthropic may keep the related conversation for feedback up to five years and use it for improvement (de-linked from your ID). Don’t rate anything you care about keeping private.
- Past data already in model weights does not come back out. No toggle rewinds a finished run.
- Archive or lock a chat – training eligibility still follows the account setting.
- Personal Claude or ChatGPT login on work docs bypasses every enterprise “we don’t train” promise.
- Safety classifiers can still pull flagged chats for internal enforcement.
Same pattern shows up in community threads again and again: main switch off, casual thumbs-down on a half-baked code review, then surprise that the conversation still lingered in a pipeline somewhere.
How the major options actually stack up
| Service | Personal default (as of early 2026) | Main control | Business/API default |
|---|---|---|---|
| ChatGPT | May train | Improve the model for everyone → off | No training |
| Claude | Only if you allow | Help improve our AI models | No training |
| Gemini | Trains while Keep Activity on | Keep Activity off | No without permission |
None of these promise zero retention or zero human review for safety. Data that absolutely cannot leave your environment needs self-hosted or strictly private endpoints. Consumer toggles were never built for that bar.
FAQ
Does turning off training delete my old chats from the models?
No. Future runs stop picking up new data (and on Claude, stored-but-not-yet-used chats too). Weights already trained keep what they absorbed.
I use ChatGPT at work on a Plus plan – is that safe?
Only if the workspace is Business or Enterprise. A personal Plus login still follows personal defaults – may train unless you opted out. Log into the company workspace or the policy deck is just slides. Temporary Chat patches one conversation; it does not change the account type.
Why do people still freak out if the switches exist?
Defaults on consumer plans still lean toward collection. The feedback loophole is easy to miss. Retention gets long when improvement is allowed (Claude’s five-year de-identified path is the clearest example). “Not used for training” still leaves logs, safety review, and possible human eyes. New surfaces – voice, agents, code environments – keep shipping with their own toggles. Checking once is not a permanent fix.
Open ChatGPT, Claude, and Gemini now and verify the three settings above. Next sensitive chat: temporary or incognito first. That habit beats another summary.