Do AI apps actually train on my chats by default?

Yes. A Stanford HAI study of 28 privacy documents across six major AI companies — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — found that all six train their models on user chats by default . If a policy doesn't explicitly say "we do not train on your conversations," assume it does.

Are AI companion apps worse for privacy than regular chatbots?

They handle more intimate data and scored worse on privacy basics. Mozilla's "Privacy Not Included" review found all 11 romantic AI chatbot apps tested earned a privacy warning label , 90% may share or sell personal data, and 10 of 11 failed minimum security standards.

Why should I worry about an AI chatbot knowing things about me if I never share my name?

Because multiproduct companies like Google, Meta, and Microsoft merge your prompts with search history, purchases, and social signals, and chatbots can make sensitive inferences — a recipe request can suggest a health condition. Stanford researchers warn these inferences cascade into advertising and even insurance contexts over time .

What should I look for in a companion app's privacy policy before committing?

Run three checks: does the app train on your chats (stated plainly, not buried), can you read everything stored about you, and is deletion real — test it with low-stakes data before you're emotionally invested. Deleting the app rarely deletes cloud history, memory logs, or backups .

How do I stop an AI app from using my conversations for training?

Look for a training opt-out in settings — but expect limits. Of 13 major chatbots evaluated, 8 offer a simple toggle, 3 bury or obscure it, and 1 offers none , and even a working opt-out only protects future messages, not data already ingested.

Do AI Apps Train on Your Chats? What Their Privacy Polici...

Q: If I opt out of training, does that delete my past conversations from the model?

No — opting out only stops future collection. As Incogni's research puts it, there's no way to extract what may have already been fed into the model , so deleting the chat log or uninstalling the app doesn't undo what the model absorbed.

The Short Answer: Yes, Most AI Apps Train on Your Chats by Default

You open an AI companion app, share something personal, and assume the conversation stays between the two of you. The data tells a different story.

A Stanford HAI study analyzed 28 privacy documents across six frontier AI companies — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — and found that all six train their models on user chats by default. Opt-outs exist for some, but not all, and they only stop future collection; they cannot extract what the model has already absorbed.

The stakes climb sharply in the AI companionship category, where conversations routinely involve fantasies, grief, and trauma. Mozilla's "Privacy Not Included" review tested 11 romantic AI chatbot apps and found every single one earned a privacy warning label. 90% may share or sell personal data, and 10 of 11 failed minimum security standards — some accepting "1" as a valid password. The average app fired 2,663 ad trackers per minute of use; the worst offender, Romantic AI, triggered 24,354 in a single minute.

Default training is the consumer norm; enterprise tiers are often excluded automatically
Opt-out prevents future ingestion only — it does not delete data already trained into the model
Companion apps handle uniquely intimate data with less transparency than mainstream chatbots
Deleting the app does not delete cloud history, memory logs, or backups

This asymmetry — where paying organizations avoid training but consumers accept it by default — is exactly why we built FlirtAlly with a different privacy architecture. The research confirms what users already sense: if the policy doesn't explicitly say "we do not train on your conversations," assume it does.

Why Opting Out Doesn't Undo What's Already Been Absorbed

The moment you hit send, your words begin a one-way journey into the model's weights. Research confirms that opting out only stops future collection — data already absorbed during training cannot be extracted, deleted, or unlearned by any current technique. Deleting the chat log, clearing the app cache, or even uninstalling the companion entirely leaves the statistical imprint intact. As Incogni's analysis puts it, "there's no way to extract what may have already been fed into the model," a reality that holds across every major provider evaluated in a 13-chatbot privacy study.

The opt-out landscape itself is a patchwork. Of the 13 major chatbots tested, 8 offer a simple toggle, 3 claim an opt-out exists but bury or obscure it, and 1 provides none at all per the same Incogni evaluation. Google's Gemini adds a cruel twist: turning off training also disables conversation memory, forcing users to choose between privacy and continuity noted by PCMag's review. Some companion services restrict opt-outs to specific regions, apply them only to future messages, or continue using data for "safety" review even after you object according to a companion-app privacy breakdown.

Opt-out prevents future ingestion only — past chats stay baked in
Gemini's opt-out disables memory, a functional poison pill
Regional restrictions and "safety" carve-outs weaken many toggles
Deleting the app does not delete cloud history, backups, or model weights

The problem compounds when multiproduct companies merge your chat data with everything else they know. Google, Meta, and Microsoft combine prompt data with search history, purchases, and social signals to build broker-style profiles that cascade across advertising, insurance, and recommendation systems documented in Stanford HAI's policy analysis. A recipe request can infer a health condition; a late-night confession can shape an ad profile. Copilot and Meta AI were singled out as the worst offenders for this cross-product profiling in the Incogni ranking. At FlirtAlly, we designed memory to stay in your conversation — not to feed a centralized model that follows you across the web.

Companion Apps Raise the Stakes: Intimate Data, Weak Disclosure, Real Breaches

How to Vet an AI Companion Before You Get Attached: A Three-Check Framework

You've found an AI companion that feels right — but before the relationship deepens, you need to know who else is listening. Mozilla's "Privacy Not Included" review found that all 11 romantic AI chatbot apps tested received a privacy warning label, and 90% may share or sell personal data. A Stanford HAI study of six major U.S. AI developers confirmed that every one uses chat data for training by default. The stakes are higher here than with productivity chatbots because companion conversations hold fantasies, confessions, grief, and trauma — data that cannot be "untrained" once absorbed.

Look for one plain sentence: "We do not use your conversations to train AI models." If it's missing, assume the opposite. Cave's heuristic treats silence as consent.
Check whether you can read everything stored about you — not just chat logs, but memory layers, inferred profiles, and analytics.
Test deletion with low-stakes data before you're emotionally invested. Deleting the app rarely deletes cloud history, backups, or what the model already learned.

Supporting habits make the framework stick. Enable training opt-outs where they exist — eight of 13 major chatbots offer simple switches, but they only protect future data. Screenshot your privacy settings after each policy update so you have proof of what was promised. And apply a hard "do not tell the AI" rule for government IDs, passwords, home addresses, medical details, and other people's private information. Companion conversations are not confidential like therapy; breaches such as the Muah.AI leak exposed 1.9 million emails tied to sexual prompts and real names.

FlirtAlly was built with this framework in mind — no account needed to start, memory windows you control by plan, and a privacy policy written to be read, not weaponized. The companion you choose should protect your intimacy by design, not by accident.

What Honest Design Looks Like — and Where FlirtAlly Stands

If regulators are fining companies for missing guardrails, the answer to "what should good design look like?" is already written — it's just not evenly distributed.

Italy's data protection authority fined Replika's developer €5 million in May 2025, partly for having no age verification. Around the same time, the U.S. FTC ordered seven AI companion companies to report on their safety testing and age restrictions. These aren't hypothetical risks — they're the baseline that honest design has to clear.

FlirtAlly was built against that baseline. The default experience is PG/Flirt — you get the full companion: memory, texting first, games, calls, mail. The 18+ "Spicy" side is strictly opt-in, gated behind ID-based age verification and a qualifying plan. Prohibited categories are hard-coded, media is screened scene by scene, and in-app reporting lets members flag problems directly. FlirtAlly is not an "unrestricted" app — and that's the point.

That matters because Mozilla's "Privacy Not Included" review found all 11 romantic AI chatbot apps it examined received its privacy warning label, and 10 of 11 failed minimum security standards. Some accepted "1" as a password. Guardrails like age verification and reporting aren't optional extras in this category; they're what regulators fined others for lacking.

Transparency is a feature, not a footnote. The same honesty applies to how your data is handled — and we'd rather you verify that yourself than take our word for it. Researchers recommend a simple three-check framework before committing to any companion app:

Does the app train on your chats — and is that stated plainly, not buried?
Can you read everything stored about you, not just what appears in responses?
Is deletion real — test it with low-stakes data before you're emotionally invested?

That last check matters more than most people realize. Once data is trained into a model, it cannot be meaningfully extracted — deleting the chat log doesn't delete what the model absorbed. Opting out only protects the future.

So run the three checks on FlirtAlly's own policies. Read them the way you'd read any app's — skeptically. A policy tells you a company's intentions; it cannot tell you its competence. We think our design holds up to scrutiny, and we'd rather earn your trust by being checked than by being believed.

Start free — no account needed.

Do AI Apps Train on Your Chats? What Their Privacy Policies Actually Say