The Privacy Side of AI Detectors: What Happens to Your Submitted Text

You paste a paragraph into an AI detector, hit check, and get a percentage back. Quick, harmless, done. But that text just left your computer and went somewhere — and “somewhere” has a privacy policy, a retention schedule, and possibly a copy of your words sitting on a server you’ll never see.

Most coverage of AI detectors argues about accuracy. Almost nobody asks the simpler question: what happens to the writing you hand over? Here’s the honest version.

Key takeaways

• Most AI detectors send your text to a server, and many keep at least a temporary copy. “We don’t store it” and “we don’t keep it forever” are not the same promise.
• Turnitin is the case students worry about most: submitted papers can be added to a comparison repository, which is a retention issue, not just a one-time scan. Schools choose whether papers get stored.
• You can usually learn what happens by reading three lines in a privacy policy: the retention period, whether your text trains models, and whether data is shared with third parties.
• Free tools often pay for themselves with your data. If you can’t tell how a tool makes money, assume your text is part of the deal.
• You can cut your exposure: strip names and identifying details before pasting, and favor tools with a clear no-training, short-retention policy.

Where your text actually goes

When you click “check,” your browser packages up the text and sends it over the internet to the detector’s servers. This isn’t optional sneakiness — it’s how the tools work. The statistical models that estimate whether writing looks machine-generated are too big to run inside a web page, so the analysis happens remotely. (If you’re curious about what those models are measuring, we walk through how detectors actually score text elsewhere.)

So transmission is a given. The privacy question is about what happens *after* the score comes back. Three things can occur, and a tool might do one, two, or all three:

• It deletes your text immediately. The analysis runs, the score returns, the input is discarded.
• It retains your text for a while. Companies keep submissions for debugging, fraud prevention, or so you can see your history. “A while” might be hours or might be years.
• It reuses your text. Your words could be folded into a training set to improve the detector, or shared with a partner, or used to build features.

The trouble is that a single phrase like “your privacy is important to us” tells you nothing about which of these is true.

A quick scenario

Picture Maya, a grad student finishing a thesis chapter. She wrote it herself but heard her department runs everything through a checker, so she wants to see her own score first. She pastes 4,000 words — including a paragraph naming an interview subject and that person’s employer — into the first free detector that ranks on Google.

The score comes back at 12% AI. Reassuring. But Maya never read the policy. That tool retains submissions for 30 days and reserves the right to use “submitted content” to improve its services. Her interviewee’s name and employer now sit on a third-party server, attached to nothing Maya agreed to share on that person’s behalf.

Nothing dramatic will probably happen. But “probably nothing” is a weak privacy guarantee when a 90-second read could have told her the answer.

The Turnitin question students keep asking

Turnitin deserves its own section because it works differently from a consumer checker, and because students ask about it constantly.

Two things are bundled under “Turnitin.” One is the long-standing similarity (plagiarism) check. The other is its AI writing indicator, added more recently. They’re separate analyses.

The retention concern is mostly about the similarity side. Turnitin can add a submitted student paper to a comparison repository, so that future submissions — yours included, and other students’ — can be matched against it. That means your paper may persist as a reference document, not just get scanned and forgotten. But this is a setting, not a fixed rule. For each assignment an instructor can store papers in the standard global repository, store them in an institution-only repository, or choose not to store them at all. So the behavior isn’t identical everywhere.

The practical move: with a school’s tool, *you* aren’t the customer — your institution is, and it sets the configuration. If retention matters to you, ask your instructor or the academic-integrity office two direct questions. Is my paper added to a repository? Who can see the AI indicator score? You’re entitled to those answers, and asking is not an admission of anything.

How to read a privacy policy in two minutes

You don’t need to read the whole document. Open it, use your browser’s find function (Ctrl/Cmd-F), and search three terms.

“Retain” (or “retention”)

Look for a number. “We retain submitted content for X days” is a real answer. “We retain content as long as necessary” is a non-answer dressed as one. Shorter is better; a stated period is far better than vague language.

“Train”

You want to know if your text feeds the company’s models. A clear policy says so and ideally offers an opt-out. If the word “train” doesn’t appear at all, that silence is itself a yellow flag — don’t assume the absence of a claim means the absence of the practice.

“Third part” (catches both “third party” and “third parties”)

This surfaces whether your data gets shared with analytics vendors, advertising partners, or anyone else. Some sharing is mundane (a hosting provider). Selling your content to data brokers is not. The policy should distinguish between the two.

If you can’t find satisfying answers to all three in a couple of minutes, that difficulty is the answer.

Free tools and the cost you don’t see on the invoice

Running detection models isn’t free for the people who run them. Servers and GPU time cost real money. So when a checker is free, shows no ads, and sells no subscription, it’s reasonable to ask where the money comes from.

Sometimes the answer is fine — a company offers a free tier to upsell a paid plan. Our pricing page lays out a model that doesn’t depend on hoarding your drafts. Sometimes the answer is your data: longer retention, broader reuse rights, sharing with partners. Paid doesn’t automatically mean private, and free doesn’t automatically mean exploitative. But “free with no visible business model” is the combination that should make you read the policy before you paste.

Practical steps to protect yourself

You don’t have to stop using detectors. You just have to use them deliberately.

• Strip identifying details first. Before checking, remove names, email addresses, employers, student IDs, and anything about third parties (like Maya’s interview subject). The detector scores style and statistics, not identities — it doesn’t need real names to do its job.
• Check excerpts, not whole documents. Run a representative paragraph or two rather than your entire dissertation. You’ll usually learn what you need from a sample, and you’ve exposed far less.
• Favor tools with a stated no-training policy and a short retention window. A company willing to commit to specifics in writing is signaling that it has thought about this.
• Be extra careful with anything confidential. Client work under NDA, unpublished research, legal or medical content — for these, default to not pasting into a third-party tool at all unless its policy is explicit and acceptable.
• Use one tool you trust rather than spraying your text across ten. Every additional checker is another server holding a copy. Pick one with a policy you’ve read and stick with it.

If you want to keep reading on detection and the ethics around it, you can browse the rest of our writing on detection and ethics.

Frequently asked questions

Do AI detectors store the text I submit? It depends on the tool. Almost all of them send your text to a server, since the models are too large to run in your browser. Whether they keep a copy afterward varies — some delete within seconds, some retain for days or weeks, a few keep it indefinitely or add it to a repository. The only reliable way to know is to read the tool’s privacy policy and look for a stated retention period.

Does Turnitin keep my paper after it checks for AI? Turnitin’s similarity check can add submitted student papers to a comparison repository so future work gets matched against them, but it’s configurable. Instructors can pick the standard global repository, an institution-only repository, or “do not store” per assignment. Its AI writing indicator is a separate analysis. Because schools control the settings, ask your instructor or academic-integrity office how submissions are retained rather than guessing.

Can a detector use my text to train its AI? Some can, if their terms allow it. Detection companies improve models partly by collecting examples of human and AI writing, and your submission could become one. A good policy states plainly whether your content trains models and offers an opt-out. Silence on training is a reason to avoid pasting anything sensitive.

Are free AI detectors less private than paid ones? Not always, but free tools more often monetize through data. If a tool is free with no ads and nothing to sell, ask how it stays funded. Some free checkers retain submissions longer or share data with partners. Paid tools aren’t automatically private, but they’re likelier to have a business model that doesn’t lean on your text.

How can I check a tool’s privacy before I paste anything? Open the privacy policy and search for “retain,” “train,” and “third part.” Those land you on retention, training use, and data sharing. If you can’t find clear answers in a couple of minutes, that’s informative on its own.

The short version

The privacy of an AI detector isn’t a mystery you have to accept blindly — it’s a policy you can read in about the time it takes to brew coffee. Know that your text travels to a server, find out how long it stays and what it’s used for, and strip out anything you wouldn’t want sitting on a stranger’s hard drive. Treat your words like data, because that’s exactly what they become the moment you click check.

If you’d rather start with a tool that’s upfront about how it treats your text, see how PaperBleach handles your text.