By default, your conversations can train the models, and deletion is conditional by design: safety flags override it and copies travel past the company you typed to. What was absorbed into a trained model does not come back out.
Your most candid disclosures, kept to train. Keep the record.
Prompts and conversations are commonly kept to "improve" the model, with no clear end date.
Sign-up asks for little; some add age or identity checks.
Industry profile reviewed 23 August 2026. Also machine-readable via the free API.
Your chats are the leak. People type medical worries, legal trouble and work secrets into these tools in the first person, and a transcript in your own words is far harder to explain away than a stolen password.
Your conversations train the models
The product runs on what you type. On personal plans your chats train the models unless you find the switch, and the company judges opt-outs itself. Humans read a sample of chats at several services, and a reviewed copy can stay for years after you delete.
Deletion is conditional by design
A safety or legal flag you never see means deleted data is kept anyway, with no notice and often no end date. And once data is relabelled de-identified, no promise in the policy applies to it any more.
What you typed travels past the company
Every policy read routes conversations and account data onward to contractors: hosting, moderation, support, and in places marketing. The list is given by category and left open-ended, the countries are usually unstated, and you hold no agreement with any of them.
Worked out from the industry, not from any one company. What you actually handed over is yours to record.
Built only from what this kind of service actually collects. A dimension that the data does not support is not listed.
People ask AI about symptoms and mental health frankly.
Conversations often reveal emotional state.
What you ask implies your views.
Two kinds of hold. LAW SETS IT: a statute makes them keep it. THEY SET IT: a ground the company grants itself.
They treat it as no longer being about you, though such data can sometimes be re-identified.
To prove they checked your age or identity and to block known fraud, sometimes held by a separate verification company.
To enforce bans and stop blocked or abusive users coming back.
Tax and company law makes them keep billing and payment records.
They can keep records to defend a live or possible legal claim.
Your chats are the value, feeding the company's own models: training is on by default, opted-in chats are kept for years, and at the platform giants, assistant conversations now shape the ads you see. No documented sales of chat logs to brokers so far. The demand is in-house, and huge.
Conversations are valuable behavioural and intent data.
Your prompts and uploads train the model itself unless you opt out, and opt-out works only going forward.
Anonymised is their word, and language models are the exact tool now shown to infer who wrote a text, at scale (Staab et al., ICLR 2024).
Fifteen demographic attributes re-identify 99.98% of Americans in a released dataset (Rocher, Hendrickx and de Montjoye, Nature Communications, 2019); date of birth, postcode, and sex alone did it for most people in the first study of the problem (Sweeney, 2000).
Four card transactions identify 90% of people in payment data (de Montjoye et al., Science, 2015).
Language models infer where a person lives, their income, and their sex from their writing alone, at near-human accuracy and at scale (Staab et al., ICLR 2024).
A face, voice, or fingerprint template identifies a person directly; there is nothing left to anonymise, and it cannot be reissued like a password.
Browser and device fingerprints were unique for 84% of visitors in the first large study (Eckersley, 2010), and sparse histories of what people viewed re-identified them against public reviews (Narayanan and Shmatikov, 2008).
The shape of who a person connects with re-identifies accounts across networks with no other data (Narayanan and Shmatikov, 2009).
THE STUDIES Estimating the success of re-identifications in incomplete datasets using generative models (Nature Communications 10, 3069, 2019)·Simple Demographics Often Identify People Uniquely (Carnegie Mellon University, Data Privacy Working Paper 3, 2000)·Unique in the shopping mall: On the reidentifiability of credit card metadata (Science 347 (6221), 2015)·Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024, 2024)·How Unique Is Your Web Browser? (Privacy Enhancing Technologies Symposium (PETS 2010), 2010)·Robust De-anonymization of Large Sparse Datasets (IEEE Symposium on Security and Privacy, 2008)·De-anonymizing Social Networks (IEEE Symposium on Security and Privacy, 2009)
Clauses that recur across this industry, and what each one actually permits.
“to provide and improve our services”
The catch-all purpose. Analytics, profiling, personalisation, and increasingly AI training can all ride under it. When they want to do something new with your data, this sentence usually already covers it.
THE MOVE An objection draws the line: use your data to run the service, not to improve, target, or train on it.
“we do not sell your personal information”
Often technically true, and still misleading. It usually means no cash changes hands. Data can still flow to ad networks, analytics firms, and partners: on their definition, sharing is not selling.
THE MOVE Flip the do-not-sell switch where one exists. The written objection on top of it goes on your record.
“service providers, partners, and affiliates”
How data leaves the building with no name attached. Recipients are listed by what they do, never who they are. You cannot send a request to a company you cannot name, which is the point.
THE MOVE An access request can ask for the recipients by name, not just the categories. UK and EU law put that choice with you. The reply, or the silence, goes on your record.
“aggregated or de-identified information”
Stripping the name does not strip the pattern, and the pattern often still points at you. Policies grant themselves free, indefinite use of this data because in their telling it is no longer about you.
THE MOVE If a deletion comes back as 'anonymised', keep the reply. It is their claim, not a fact you can check.
“retained as long as necessary, or as required by law”
They can keep it for: legal duties, tax rules, fraud prevention, possible lawsuits, their own business reasons. None of them carries a firm end date. Deletion becomes a negotiation, not an event.
THE MOVE Which reasons apply to you, and how long each runs, is a request of its own. Their reply goes on your record.
“you grant us a licence to use your content”
A contract term, not a data setting, so a privacy request cannot undo it. The careful version ends with your account. The broad version can be passed on, never expires, and survives deletion.
THE MOVE Whether the licence ends with the account is written in their terms, not yours. Closing the account goes on your record.
“we may use content you provide to improve our services, for example to train the models”
Your chats train the models by default, and improve covers research and new products. Opting out is on you, and it stops only future use, not what already went in.
THE MOVE Flip the training switch first, then put the objection in writing. Both go on your record.
“we will delete the data within 30 days unless it is necessary to retain it for legal, compliance, or safety reasons”
Every deletion has an exception. A flag you cannot see means your data stays, with no notice and no stated end date.
THE MOVE The deletion request can carry a question with it: is anything being kept. Their reply goes on your record.
“even if you opt out, we will use your conversations for model improvement when they are flagged for safety review”
A classifier you cannot see or contest overrides the setting you chose. A flag turns an opted-out conversation back into training material, and it is the same route that puts a human reader on it.
THE MOVE A request covers whether your conversations were flagged and what followed. Their reply goes on your record.
Their own policy is the one that binds them. Pin it down with a request, and keep the reply. Start your record →