What Actually Happens to What You Type Into ChatGPT, Claude, or Gemini? The Answer Isn't the Same for All Three

In March 2023, three Samsung engineers pasted, within twenty days of each other, the source code of a semiconductor database, the transcript of a confidential internal meeting, and a test sequence for detecting defective chips into ChatGPT. None of them was stealing anything: they just wanted the AI to help debug an error, draft some notes, optimize a process. Samsung banned ChatGPT company-wide the moment it found out. You don't need to work in semiconductors, or handle anything nearly that sensitive, to make the same decision without realizing it: every time you type something personal, a medical detail, a draft contract, or a complaint about your boss into an AI chat, you're making the same call those engineers made. And the answer to "what happens to this afterward" isn't a single one — it depends on which tool you use, which plan, and whether you ever touched a setting most people never open.

Three real policies, not one generic one

Anthropic (Claude). As of August 28, 2025, Free, Pro, and Max accounts moved to an opt-out model: if you don't touch anything, your conversations get used to train future models and are kept for up to five years; turn the setting off and it reverts to the old 30-day retention. Conversations you delete are excluded from training either way. Anthropic's stated reason: real interaction data "provide valuable insights on which responses are most useful and accurate for users," and 18-to-24-month development cycles mean longer retention helps keep model behavior consistent across versions. None of this applies to Claude for Work, education, government, or the API.

Google (Gemini). With "Gemini Apps Activity" turned on — the normal state — Google uses your chats to "provide, develop, and improve its services, including training generative AI models," retaining them for 18 months by default (adjustable to 3, 36 months, or indefinitely). Turn it off, and conversations are still held for 72 hours "to respond to your queries and help protect Google's security" before deletion. And there's a wrinkle almost no one reads to the end: a subset of chats get reviewed by people — "human reviewers, including Google's trained service providers" — to improve the product, and those specific chats are kept for up to three years regardless of whether Activity is on or off.

OpenAI. Here we need to be honest about this piece's limit: OpenAI's and ChatGPT's own privacy-policy pages returned an automated-access block to our server on every attempt, so we couldn't read them ourselves the way we did with Anthropic and Google. What we could verify firsthand is the API policy, which OpenAI documents precisely: data sent through the API "is not used to train or improve OpenAI models, unless you explicitly opt in to share data," with 30-day default retention for abuse monitoring. That API policy is explicitly different from ChatGPT, the consumer product: multiple consistent technical outlets describe training as on by default on Free and Plus accounts, switchable off via the "Improve the model for everyone" toggle, with the same 30-day abuse-monitoring retention even with training turned off — but we couldn't confirm that against OpenAI's own source directly, and we're saying so rather than pretending we did.

There's a logic behind why these three policies look so different from each other, and it's worth naming because it won't change even as the products do: on free and cheap consumer tiers, your conversations are part of what you pay — the company covers the cost of serving you with what it learns from your data; on the API and in enterprise contracts, the customer pays in real money, so the company doesn't need to also collect payment in data, and promising not to use it is part of what it's selling. That's why Anthropic explicitly excludes Claude for Work and the API from its August 2025 change, and why OpenAI's API policy is the most protective of the three we could verify — not coincidence or generosity, just the business model talking.

What no policy explains: the court order

There's one more reason, unrelated to any setting, that "I deleted it" doesn't always mean what you think. In May 2025, in The New York Times' copyright suit against OpenAI, a federal judge ordered the company to preserve all ChatGPT output logs that would otherwise have been deleted, including more than 20 million anonymized chats OpenAI had to hand over to the court. The order was partially lifted in October, but logs preserved during that window remain accessible, and accounts specifically flagged by the Times remain under a retention obligation. Your personal retention setting matters not at all against a court order — it's the variable no company explains on its website, because it isn't up to them.

Four questions for any AI chat, whichever one it is

You don't need to memorize these three policies — they'll change, and there'll be a fourth and fifth tool next year. You need to know what to ask any of them:

1. Is training on my data on or off by default? If you don't know, assume it's on: that's the more common setting among the ones we could verify.
2. How much is kept even if I turn it off? Almost no policy is zero: it's usually somewhere between 72 hours and 30 days for security, regardless of the training setting.
3. Can a person read it? Look for the word "reviewer" in the policy: if it's there, someone with human access can end up seeing that specific chat, usually with longer retention than the rest.
4. Is there a temporary or incognito mode, and what does it actually change? It usually keeps the chat out of your history and out of training, but it almost never removes the short security-retention window.

Applied to Gemini, for instance: (1) on by default; (2) 72 hours even when off; (3) yes, a portion of chats go through human review; (4) temporary mode exists and keeps the chat out of history and training, but not out of that 72-hour minimum. Four answers, neither alarming nor reassuring: just what's actually there.

And if a court order, an investigation, or a corporate merger enters the picture, none of those four answers is the last word anymore — as OpenAI found out with the Times, not because it changed its policy, but because a court told it to ignore that policy for a while. That's not a reason to avoid these tools. It's the reason not to type into them anything you wouldn't put in an email that might, someday, end up in front of a judge.

For anyone who wants to go deeper

• Anthropic's official announcement on the consumer terms change (August 2025), with the full timeline and rationale.
• Google's official help page on Gemini Apps Activity, with the detail on human review and retention.
• OpenAI's official guide to data usage in its API, the only one of the three we could verify firsthand for OpenAI.
• Bloomberg Law's coverage of the order to preserve 20 million ChatGPT logs in the Times case.
• Forbes's reporting on the Samsung leak that opens this piece.