Nobody Tells You How to Stop AI From Training on Your Data
In this article
Bottom line: Every major AI vendor now buries a data-training opt-out somewhere in account settings, and the defaults are almost never in your favor.
Anthropic's August 2025 policy change flipped Claude.ai's consumer default to "train on your chats for 5 years" unless you opted out by September 28, 2025 — a pattern OpenAI, Google, Meta, and Adobe have all run in some form.
API and enterprise plans are usually excluded from training by contract; consumer apps almost never are unless you dig into settings and flip a toggle yourself.
And even when you find it, opting out is rarely retroactive — it stops future collection, not the training runs that already happened.
I found out about Anthropic's policy change from a Hacker News thread, not from Anthropic.
Somebody had gone digging through the Claude.ai settings menu and found a new toggle buried under "Privacy," quietly defaulted to on, that let the company keep your conversations for five years and use them to train future models.
I'd been using Claude daily for months. I had no idea the terms under which I was doing that had just changed underneath me.
That's the whole problem in one sentence: the opt-out exists, but nobody tells you it exists.
The Setup: A Policy Change Nobody Announced Loudly
Here's what actually happened, because the details matter more than the outrage. In August 2025, Anthropic updated its consumer terms for Claude.ai, Claude Code, and the Chrome/desktop apps.
Previously, consumer chats were deleted within 30 days and not used for training by default.
The new policy flipped that: unless you actively opted out, your conversations and coding sessions could be used to train future models, with retention extended to five years.
Existing users got a pop-up with an "Accept" button and a much less visible toggle for whether to allow training — and reports at the time noted the toggle defaulted to on, with the deadline to opt out landing September 28, 2025.
If you clicked through the way most people click through terms updates, you'd opted in without meaning to.
This is not an Anthropic-specific story, and I want to be fair to them for a second: their API terms were never affected.
If you're building on Claude through the API, Bedrock, or Vertex AI, Anthropic's commercial terms say your prompts and completions aren't used for training, full stop.
The consumer product is where the ambiguity lives, and that's true across the entire industry, not just one vendor.
The Core Insight: There Are Three Different Contracts, and You're Probably Reading the Wrong One
Every AI company is running at least two, sometimes three, separate data policies simultaneously, and conflating them is how people get surprised.
Consumer apps default toward training
The free or Plus-tier chat app you use in your browser is a data-collection product first and a productivity tool second.
This is where the training defaults are usually most permissive. OpenAI's ChatGPT has a "Improve the model for everyone" toggle inside Data Controls — off means new conversations aren't used for training, but historical conversations already collected aren't clawed back.
Google's Gemini has "Gemini Apps Activity" in your Google Activity Controls; turning it off stops future review, but Google still retains conversations for up to 72 hours regardless, for safety processing, before deletion.
API and enterprise contracts default toward not training
If you're paying for API access, a Business plan, or an Enterprise seat, the contract almost always excludes your data from training by default.
This has been true of OpenAI's API since March 2023, GitHub Copilot for Business and Enterprise since launch, and Anthropic's commercial API tier throughout the 2025 consumer policy shift.
Companies protect enterprise data contractually because that's where the lawsuits and the compliance audits live — consumer users don't have that leverage, so they don't get that default.
The gray zone: free tiers of paid products
GitHub Copilot Individual (the free and Pro consumer tiers, not Business) had its own separate toggle — "Allow GitHub to use my code snippets for product improvements" — that a huge number of developers never knew existed until someone wrote a blog post about it.
Adobe ran into the same backlash in 2024 when photographers discovered a "Content analysis" setting that permitted Firefly training on their Creative Cloud files, on by default, buried three menus deep.
Where the Actual Toggles Live (September 2026)
If you want to check your own exposure right now, here's where to look — because "check your settings" is useless advice without the actual path:
- ChatGPT (OpenAI): Settings → Data Controls → "Improve the model for everyone." Also check "Temporary Chat" mode for one-off conversations you never want retained.
- Claude (Anthropic): Settings → Privacy → look for the training/data retention toggle. If you accepted the 2025 terms update without reading it, go verify this now — it likely defaulted to on.
- Gemini (Google): myactivity.google.com → Gemini Apps Activity. Turning this off also affects your ability to reference past conversations, which is the trade-off Google doesn't advertise.
- Meta AI: Meta has objection forms for EU, UK, and Brazilian users under GDPR/local law to opt out of having public posts used for training; outside those jurisdictions, there is effectively no consumer-facing opt-out.
- GitHub Copilot: Settings → Copilot → "Allow GitHub to use my code snippets for product improvements" (individual plans only; Business/Enterprise already excludes you by contract).
- LinkedIn: Settings → Data Privacy → "Data for Generative AI Improvement," added in 2024, opt-out not retroactive.
Notice the pattern: the setting is always named something bland enough to skim past, and it's always somewhere between two and four clicks deep.
The Reality Check: Opting Out Doesn't Undo Anything
Here's where I have to be blunt, because the framing of "just opt out" oversells what the toggle actually does.
Turning off training only affects data collected after you flip the switch.
If your conversations, code, or images were already ingested into a training run before you found the setting, they're baked into model weights that already exist.
There's no mechanism — not from any major vendor — to retroactively extract your specific contribution from a trained model.
Machine unlearning is an active research area, not a shipped consumer feature.
There's also a retention-versus-training distinction that trips people up.
Turning off "training" doesn't always mean your data is deleted; Google's 72-hour retention window for safety review happens regardless of your Activity setting.
And enterprise contracts that exclude your data from training often still permit retention for abuse monitoring, which is a separate clause most people never read.
If you're in the EU, you actually have more leverage than the toggle gives you: GDPR Article 21 grants a right to object to processing, including AI training, and companies are legally obligated to honor it — which is exactly the mechanism behind Meta's EU-specific opt-out forms that don't exist for US users.
Everywhere else, you're relying entirely on the vendor's voluntary settings page, and that page can change its defaults again with the next terms update, the way Anthropic's did in 2025.
The Practical Takeaway
If you actually care about this, here's the workflow I've settled into after going through every major vendor's settings this year:
1. Audit your toggles quarterly, not once. Terms updates ship silently and defaults reset. Put a recurring calendar reminder — I check mine the first week of every quarter.
2. Assume consumer apps train on you unless proven otherwise. Treat the free or Plus tier of any AI tool as a data-collection product.
If you're pasting proprietary code or sensitive client data into a consumer chat window, you've already made a decision you might regret.
3.
Use API access for anything sensitive. If your work involves confidential material, route it through the API or a Business/Enterprise tier where the contract — not a settings toggle you might miss — protects you by default.
4. Read the terms-update email once, actually. I know nobody does this. But the Anthropic case proves that the one email you skim is exactly the one that changes your default.
5. Don't expect deletion to mean erasure. Opting out protects your future inputs.
It does not undo training that's already happened, so the highest-leverage move is catching the setting before you generate the data, not after.
I still use Claude every day. I didn't cancel anything over this, and I'm not telling you to either.
But I did go back and turn off training on every account I have, and it took me twenty minutes across six different settings menus to do something that should've taken one honest email from each company.
Have you actually gone and checked your AI data settings, or are you assuming the defaults are fine? What did you find when you looked?
