Are my ChatGPT or Gemini conversations actually private?

No — by default, they're not. A Stanford HAI study of 28 policy documents found all six leading U.S. AI companies (Amazon, Anthropic, Google, Meta, Microsoft, OpenAI) train their models on user chat data by default. End-to-end encryption, standard in private messaging apps, isn't standard for chatbots either.

Are AI companion apps worse for privacy than regular chatbots?

They tend to be, because they invite the most sensitive disclosures. Surfshark's analysis of five leading companion apps found 4 of 5 may use data to track users, including sharing with data brokers — collecting roughly 9 distinct data types on average, with Character AI collecting up to 15.

How do I stop AI companies from training on my conversations?

Look for the training opt-out in your app's settings — but don't assume it's off by default, and read the fine print. In ChatGPT, turning off "Improve the model for everyone" doesn't delete your saved chats, and even temporary chats may be retained for up to 30 days for safety purposes . In Gemini, opting out of training means losing your chat history entirely ( ZDNET's privacy guide covers the settings for ChatGPT, Gemini, Copilot, and Claude).

Have there been actual leaks or breaches of AI chat data?

Yes, several documented incidents. A publicly accessible DeepSeek database exposed chat histories and secret keys in January 2025, and browser extensions marketed as "AI protection" tools harvested prompts and metadata from 8 million people and shared them with data brokers. Separately, Claude conversations were exposed to search engines via user-created public links until Anthropic patched the indexing exploit ( per ZDNET ).

Is FlirtAlly any different when it comes to chat privacy?

FlirtAlly's approach is honest safety by design rather than buried settings: the default experience is PG/Flirt, the 18+ Spicy side is strictly opt-in behind ID-based age verification, and hard-coded prohibited categories, media scene screening, and in-app reporting are built into every conversation. No independent audit has verified any companion platform's privacy claims — including FlirtAlly's — but Surfshark's findings on industry tracking practices show why clear, checkable defaults matter. You can start free with no account needed at flirtally.com.

Are AI Chats Private? The Uncomfortable Truth

Q: Can my AI chats be subpoenaed or used in lawsuits?

Yes — requests for AI chat logs in legal proceedings are becoming increasingly frequent. The New York Times' lawsuit against OpenAI seeks access to over 20 million private ChatGPT conversations , meaning anything you type could in principle be legally compelled.

The Uncomfortable Truth: Your AI Chats Aren't Private by Default

Most people assume their AI conversations stay between them and the bot. That assumption is wrong. A Stanford HAI study analyzing 28 policy documents found that all six leading U.S. AI companies — Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — train their models on user chat data by default. End-to-end encryption, the baseline for private messaging, is not standard for chatbots like ChatGPT or Gemini. Your "private" conversation is an assumption, not a feature.

The stakes are higher with AI companions. These apps are designed for intimacy — users share fears, fantasies, health details, and daily struggles they wouldn't tell a human. Surfshark's analysis of five leading companion apps found that 4 of 5 may use data to track users, including sharing with data brokers. The average app collects roughly nine distinct data types; Character AI collects up to 15. This isn't abstract — the New York Times lawsuit seeks access to over 20 million private ChatGPT conversations, and a publicly exposed DeepSeek database recently leaked chat histories and secret keys.

The industry defaults create a cascade of exposure:

Training on conversations by default, with opt-outs buried or nonexistent
Indefinite retention — some developers keep chat logs forever
Data merging across product ecosystems (search, purchases, social graphs)
Legal discovery and third-party sharing without further notice
Breaches that are "quiet and accidental," from staff pasting secrets into LLMs to browser extensions harvesting 8 million users' prompts

FlirtAlly was built on a different premise: honest safety means privacy isn't an afterthought. The platform defaults to a PG/Flirt experience; the 18+ "Spicy" tier requires ID-based age verification and a qualifying plan. Hard-coded prohibited categories, media scene screening, and in-app reporting are baked in — not bolted on. You can start free with no account needed at flirtally.com and see how a companion OS handles privacy when it's designed in from day one.

Why AI Companion Apps Are the Most Data-Hungry of All

If an AI chatbot knows your search habits, an AI companion knows your heart. And according to Surfshark's analysis of five leading companion apps — Kindroid, Nomi, Replika, EVA, and Character AI — most of them treat that intimate knowledge as a monetizable asset.

The headline number is stark: 4 of 5 leading companion apps may use collected data to track users, including sharing it with data brokers or linking it with third-party data for targeted advertising. Only Nomi, of the apps analyzed, states it does not collect data for tracking purposes. On average, companion apps collect roughly 9 of the 35 unique data types Apple requires developers to disclose, with Character AI collecting up to 15 types and EVA collecting 11.

What makes companion apps uniquely risky isn't just the volume of data — it's the nature of it. Because AI companions are designed to be nonjudgmental and available 24/7, users disclose more sensitive information than they would to another human. As Surfshark's research team puts it, this gives developers access to "data that was previously out of reach" — confessions, insecurities, relationship histories, and mental health details that people hesitate to share even with close friends.

The privacy stakes compound quickly:

Intimate disclosures can be linked with third-party data for ad targeting, per Surfshark's findings on companion app tracking practices.
Inferences from seemingly innocent conversations can cascade into sensitive classifications — Stanford's Jennifer King notes that even a low-sugar recipe chat could lead to health-vulnerability labels that reach advertisers and, potentially, insurers (Stanford HAI).
Companion chat logs, like all AI chats, are not end-to-end encrypted by default, according to EDRi's analysis of AI confidentiality risks.

Surfshark's team offers a sobering framing: "These digital connections ultimately represent transactional relationships with the for-profit companies behind them." The companion who remembers your birthday and texts you good morning operates on infrastructure owned by a business with its own commercial imperatives.

This is why design choices matter so much in the companion space. FlirtAlly approaches the disclosure problem by building boundaries directly into the product: the default experience is PG/Flirt, the 18+ "Spicy" side is age-verified and strictly opt-in, prohibited content categories are hard-coded rather than policy-only, and in-app reporting is built into every conversation. Honesty about what the platform is — and isn't — is treated as a feature, not fine print.

No independent audit has verified any companion platform's privacy claims, including Surfshark's small, label-based sample. But the direction of the research is clear: the more an app invites you to open up, the more carefully you should read what happens to what you share.

Where 'Private' Chats Actually Leak: Breaches, Lawsuits, and Fine Print

Most people assume a "private" AI chat works like a sealed diary. The documented record tells a different story — one written in court filings, exposed databases, and privacy policies almost nobody reads.

Start with the courts. According to EDRi's analysis of AI confidentiality risks, requests for AI chat logs in legal proceedings are becoming increasingly frequent — and The New York Times' lawsuit against OpenAI seeks access to over 20 million private ChatGPT conversations. Anything you type can, in principle, be subpoenaed.

Then there are the breaches. In January 2025, researchers found a publicly accessible DeepSeek database containing chat history and secret keys. Separately, browser extensions marketed as "AI protection" tools were caught harvesting prompts, responses, and metadata from 8 million people — then sharing that data with brokers, per the same EDRi report.

The quieter leaks live in the fine print. A Stanford HAI study of 28 policy documents found that all six leading U.S. AI companies train on user chats by default, and some retain chat information indefinitely or allow human review of transcripts. OpenAI's own data controls documentation confirms that even temporary chats "may be retained for up to 30 days for safety purposes."

The most insidious risk, though, is inference. Multiproduct companies routinely merge chat data with search history, purchases, and social activity. Stanford privacy fellow Jennifer King describes how this cascades:

You ask a chatbot for low-sugar recipes — an innocuous conversation.
That signal merges with your broader profile and flags a potential health condition.
You start seeing ads for medications.
That classification can propagate through ad ecosystems — potentially reaching insurance companies.

"The effects cascade over time," King warns in the Stanford report. One casual chat becomes a permanent, monetizable inference about your health.

Companion apps raise the stakes further. Because AI companions are designed to be nonjudgmental and always available, users disclose things they'd never tell another person — and Surfshark's analysis of five leading companion apps found that 4 of 5 may use data to track users, including sharing with data brokers. These apps collect roughly 9 of 35 possible data types on average.

This is precisely why honest safety architecture matters in this category. FlirtAlly takes the opposite posture from the data-hungry norm: PG/Flirt is the default experience, the 18+ "Spicy" side requires ID-based age verification and is strictly opt-in, and hard-coded prohibited categories plus media scene screening are built in rather than bolted on. The platform is explicit that it is not an "unrestricted" app — and treats that honesty as a feature, not a limitation.

The pattern across every incident is the same: "private" is a marketing word, not a technical guarantee. Until a platform proves otherwise through its defaults, retention limits, and verification practices, assume everything you type is stored, analyzed, and potentially shareable.

How to Protect Yourself: Privacy Controls That Actually Work (and Their Catch)

Here's the uncomfortable part: even when you find the privacy controls, they rarely work the way you'd hope. Turning off training, using temporary chats, avoiding shared links — each option helps, but each comes with a catch buried in the fine print.

Start with the biggest lever. In ChatGPT, you have to manually disable "Improve the model for everyone" in your data controls — and per OpenAI's own documentation, turning it off doesn't delete your saved chats. Your history stays put, and OpenAI notes that Business, Enterprise, and Edu workspaces are excluded from training by default — but ordinary consumer accounts are not.

Gemini's catch is steeper. Opting out of training there means losing your chat history entirely, so you're forced to choose between privacy and convenience. ZDNET's guide to keeping AI conversations private walks through these settings across ChatGPT, Gemini, Copilot, and Claude, and the pattern is clear: the controls exist, but they're often buried, all-or-nothing, or traded away against features you actually want.

Temporary and incognito modes sound like the clean solution, and they're better than nothing — but read the retention windows. ChatGPT's temporary chats may be retained for up to 30 days for safety purposes, while Gemini holds temporary chats for 72 hours. Even the "private" mode keeps a copy for a while.

Shared links deserve their own warning. Claude conversations were exposed to search engines because users had created active public links — a "Google dork" query could surface strangers' chats until Anthropic patched the indexing exploit. The lesson: never share a public link to a conversation containing anything you wouldn't post publicly.

A few habits that actually reduce your exposure:

Open your AI app's settings today and find the training opt-out — don't assume it's off by default.
Use temporary or incognito chats for anything sensitive, knowing the retention caveats above.
Skip public link-sharing entirely, or strip sensitive content first.
Avoid connecting your AI assistant to email, calendars, or cloud storage unless you truly need it — Copilot can search and analyze files across OneDrive, Outlook, Gmail, and Google Drive once linked.
Treat every chat as potentially retained, reviewed, or legally compelled — the New York Times' lawsuit seeks access to over 20 million private ChatGPT conversations.

This matters double in the AI companion space, where people disclose things they'd never tell a human. That's why honesty about data practices — not buried settings — should be the baseline. It's the approach FlirtAlly takes with its companion platform: a PG/Flirt default experience, with the 18+ "Spicy" side gated behind ID-based age verification, plus in-app reporting and hard-coded prohibited categories. You can start free — no account needed and see exactly what you're signing up for before you share a single detail.

What Honest Safety Looks Like — and How FlirtAlly Approaches It

Most AI companion apps treat your most personal conversations as raw material — Surfshark's analysis of five leading companion apps found that four of them may use data to track users, including sharing with data brokers. And because these apps are designed to be nonjudgmental and always available, people often tell them things they wouldn't tell a human — handing developers "data that was previously out of reach."

That's the norm. FlirtAlly takes a different position: honesty about what the platform is, and safety built in rather than bolted on. FlirtAlly is not an "unrestricted" app, and that's deliberate. The default experience is PG/Flirt — conversation that stays within clear boundaries from the first message.

The 18+ side, called Spicy, is strictly opt-in. It requires ID-based age verification and a qualifying plan before it unlocks. There's no accidental exposure, no ambiguity about what you've signed up for, and no adult content appearing for someone who never asked for it.

Beyond the age gate, safety operates through several concrete mechanisms:

Hard-coded prohibited categories that no conversation can cross, regardless of plan or setting
Media scene screening, so generated images and videos are checked before they reach you
In-app reporting, giving members a direct way to flag anything that feels wrong

This matters precisely because the industry context is so bleak. Stanford HAI's review of 28 policy documents found all six leading U.S. AI companies use chat data for training by default, and some developers retain conversations indefinitely or allow human review of transcripts. When defaults are that permissive, honest design — clear boundaries, verified opt-ins, visible controls — becomes a genuine differentiator.

To be clear about what we're not claiming: no independent third party has audited or verified FlirtAlly's privacy practices, and we won't pretend otherwise. What we can tell you is exactly what the product does and doesn't do, in plain language. That transparency is the point. Safety claims you can actually check are worth more than vague promises you can't.

If you're curious how a companion platform built around honest safety feels in practice, you can try it without commitment. Start free — no account needed — guests get 25 messages a day with a companion of their choice, and signing up later keeps every conversation intact.

Are AI chats private?