Every conversation you have with AI is data. Some of it trains the next model. Some of it sits in logs. All of it deserves more attention than you're giving it.
What gets collected
All the major providers store your prompts, responses, and associated metadata — device, approximate location, timing, and account identifiers. Whether that content is also used to train future models depends on your tier and your settings, and that is the part worth checking rather than assuming.
The pattern that holds across providers
Specific defaults have changed at every major provider at least once, sometimes in ways that flipped an opt-out into an opt-in or vice versa. Rather than memorise a table that will be wrong in six months, learn the shape, which has been stable:
- Consumer tiers are where training happens. Free and individual paid plans are generally the ones with a training setting to find. Whether it defaults on or off varies by provider and has changed over time.
- Business, Team, and Enterprise tiers contractually exclude your content from training. This is the single clearest line in the whole landscape. If your work genuinely cannot be trained on, that tier is the answer, not a toggle.
- API access is typically excluded from training by default. Developers are usually on stricter terms than consumers of the same company's chat product.
- Retention is separate from training. Even when content is not used for training, providers keep logs for a period — commonly around 30 days, longer for anything flagged by safety systems. Deleting a conversation in the interface does not necessarily purge those logs on the same schedule.
"Treat every AI prompt like an email to a stranger. Because it kind of is."
What to change today
Fifteen minutes, once, for each tool you actually use:
- Find the data or privacy section in settings and read what the training toggle currently says. Do not rely on what it said when you set it up — several providers have revised this, and a policy change can reset or reinterpret an old preference.
- Turn on auto-delete if the provider offers it. Google's activity controls support a rolling auto-delete window; use the shortest one you can tolerate.
- Check whether your employer has an enterprise tier. If it does, use that account for work rather than your personal one. This is the highest-value single change most people can make, and it costs nothing.
- Audit what you have already sent. Most providers let you export or review your history. It is usually more revealing than people expect.
What to never paste
- Passwords, API keys, full account or card numbers.
- Other people's personal data without consent.
- Confidential work documents (use your employer's enterprise tier instead).
- Medical records or anything you'd want HIPAA-protected.
Local AI: the privacy answer
Capable open-weight models now run on consumer laptops through tools like Ollama and LM Studio. They are a step behind the frontier hosted models on hard reasoning, and for a great deal of everyday work — summarising, drafting, reformatting, answering questions about a document you supply — that gap does not matter.
The advantage is categorical rather than incremental: data that never leaves your machine cannot be retained, trained on, subpoenaed from a provider, or exposed in someone else's breach. No policy change can reach it. For genuinely sensitive material, this is the only option that does not require trusting a company's current terms to stay current.
The cost is setup effort and hardware. A machine with enough memory to hold a mid-sized model is the practical requirement; below that, it works but slowly.
FAQ
Does deleting a conversation actually delete it?
It removes it from your view and starts the clock on the provider's deletion process. It does not instantly purge backups or safety-review copies, and content already used in a completed training run cannot be retroactively removed from a model. Assume deletion is a forward-looking control, not an undo.
Is the paid tier more private than the free one?
Sometimes, but not reliably — the meaningful line is usually consumer versus business/enterprise, not free versus paid. A personal paid subscription often sits under the same consumer terms as the free tier.
Can I use AI at work safely?
Through your employer's sanctioned enterprise account, generally yes. Through your personal account with company data, generally no — and that is a policy problem as much as a technical one. Ask before assuming; most organisations now have a written position.
What about AI features built into apps I already use?
Those inherit the host app's terms, which are often different from the model provider's. A writing tool with an AI feature may send your document to a third party under its own agreement. Worth a look at the privacy policy of the app, not just the model.
The bottom line
You cannot outsource judgement on privacy, and you cannot set it once and forget it — the defaults move. Check your settings, understand which tier you are actually on, and keep genuinely sensitive work either local or on an enterprise account. Fifteen minutes today saves a lot of regret later.
If you want the practical version of this applied to daily use rather than settings, how to use AI in everyday life covers the habits that keep sensitive material out of these tools in the first place.