Which AI Chatbots Do Not Train on Your Data? (2026)

Almost none of the free consumer ones, by default. ChatGPT and Gemini both train on consumer conversations unless you change a setting. Character.AI offers a model training opt-out only in the EEA and UK, so US users cannot opt out at all. Chai's own notice says it fine tunes its models on data derived from your in-app conversations. The chatbots that genuinely do not train on your chats fall into three groups: models you run locally, business and enterprise API tiers with contractual no-training terms, and a small number of paid consumer products that say so explicitly.

We are one of that last group, so treat this as a comparison written by an interested party and check every line. Each claim below is drawn from the company's own documentation, read in July 2026, or marked as something we could not verify.

Which AI chatbots train on your conversations?

The pattern in that table is not about which companies are ethical. It is about who pays. Every row that trains by default is a product you can use without a subscription, and every row that does not is either something you paid for or something you are running yourself.

Why free chatbots train on you

Inference costs real money on every single message. That expense does not disappear because the interface says free. So a free service is covering it somewhere else: advertising revenue, model improvement that makes the core product more valuable, data licensing, or investor money extended on the belief that one of those arrives eventually. None of that is sinister. It does mean your conversations have to be worth something to somebody other than you, and training is the cheapest way to make them worth something.

This gives you the single most useful filter available. If a chatbot is free and the company is not visibly selling something else at enormous scale, assume conversations are part of the business model until a policy says otherwise in specific words. Paid does not automatically mean private, and you should still read the terms. But a subscription removes the reason to monetize what you typed, and that alignment is the thing worth paying for.

The regional opt-out trap

This one catches Americans repeatedly. Several companies advertise a model training opt-out that only exists where a regulator required it. Character.AI's opt-out is available in the EEA and UK, which means a US user reading a review that mentions the opt-out will go looking for a setting that is not in their account. The review is not lying. It is describing a different jurisdiction.

So when you read that a service lets you opt out, the follow up question is always "where?" Check inside your own settings rather than trusting the write up, and if you cannot find the toggle, that is your answer. European privacy law has produced most of the controls that exist in this market, and Americans get them only when a company decides shipping one control globally is easier than shipping two.

What "de-identified" actually buys you

Most training disclosures include a reassurance that personal identifiers are removed first. That is a genuine safeguard and it is better than nothing. It is also weaker for chat than for almost any other data type, because in conversation people describe themselves incidentally while talking about something else entirely.

Stripping your name from a transcript does not strip the transcript. If you mentioned your job title, your city, your diagnosis, your employer's situation, or the specific thing you were awake worrying about, that is the content rather than the metadata, and no identifier removal touches it. Free text has consistently proven more resistant to de-identification than structured records for exactly this reason. Read the phrase for what it says: identifiers are removed before the data is used. It is not a promise that nothing about you persists.

Training is the part you cannot undo

Of everything a company can do with your conversations, training is the one that deletion does not reach. A retention window closes eventually. An ad profile can be reset, imperfectly. But once a conversation has contributed to a set of model weights, there is no delete button that unwinds it, and no company offers to retrain a production model because one user changed their mind.

That is why training belongs at the top of the list when you evaluate an AI chat product, above encryption and above deletion controls. A service that keeps chats for a year and never trains on them is meaningfully more private than one that deletes in thirty days after learning from everything first. If what you are worried about is your personal details circulating well beyond one app, getting your information pulled from data broker listings is a separate exercise worth running in parallel, because chat privacy and exposure privacy are different problems.

How to verify any chatbot in five minutes

Open the privacy policy and search it for four words: train, improve, retain, and advertising. Language about improving our services using your content is training in softer clothes almost every time. Then look for a number attached to retention, because a company that intends to hold data briefly says how briefly, and one that intends to hold it indefinitely writes only as long as necessary. Then check whether any opt-out is offered in your country. Then ask what pays for the compute.

If the policy will not load, or the company publishes no pricing and no clear content rules, treat that as data rather than as a neutral blank. It is not evidence of wrongdoing. It does mean you are relying on marketing instead of on a commitment. Our page on the best private AI chatbot options applies this test across the major services with each company's own language quoted, and our private AI chat page covers what a genuinely specific data policy looks like.

Where we stand

Depravity is paid only, from $15 a month, or $12 a month billed yearly, for US adults 18 and over. Chats are encrypted in transit and at rest, never sold, never shared with data brokers or marketers, and never used to train advertising or to target you. You can export or delete any conversation from settings yourself. There is automated and human review, and we disclose it, because that is how the hard lines against anything involving minors, non-consent, or illegal activity get enforced.

We are also not the most private option that exists, and it would be dishonest to claim otherwise. A model running locally on your own hardware beats every hosted service including this one, because nothing ever leaves your machine. The cost is a capable GPU, an afternoon of setup, and a noticeably weaker conversation. If that trade appeals to you, take it. The point of this article is not that you should pick us. It is that "does this thing train on what I type" has a findable answer for almost every product, and most people have never gone to look.

The short version

Free consumer chatbots train on your conversations by default. ChatGPT and Gemini let you turn it off, Character.AI only lets Europeans turn it off, Meta AI feeds ad targeting with no US opt-out, and Chai states its fine tuning outright. The ones that do not train are local models, contractual enterprise tiers, and paid consumer products that say so. Check the policy, not the marketing, and start by asking who is paying for the compute.