Free Writing Tools and AI Training: What Happens to a Pasted Draft

Pasting a draft into a free grammar checker or rewriting tool usually means that text leaves the browser and gets processed on the company's own servers — and a meaningful share of these tools use that text, by default, to help train or improve their AI models. The setting to turn that off, when it exists, is frequently a few screens deep in account settings rather than something shown at the point of pasting.

Two Different Claims That Get Conflated

"We don't sell your data" and "we don't train AI on your data" are two entirely separate promises, and privacy policies often make the first claim prominently while saying much less about the second. A company can honestly state it never sells user data to third parties while still using that same data internally to improve its own AI models — neither statement contradicts the other, but only one of them addresses what actually matters for someone worried about their writing ending up in a training set.

What an Independent Privacy Analysis Found

Incogni's 2026 ranking of AI-powered Chrome extensions by privacy risk specifically flagged several popular writing tools — including Compose AI and QuillBot — as carrying elevated data-collection scores, based on the permissions requested and the scope of data these extensions can access. The analysis noted that ten of the lower-risk extensions it evaluated claim not to collect any user data at all, though a claim like that isn't something a user can independently verify without technical auditing.

A free writing tool has to make money somehow, and if there's no subscription fee changing hands, the text passing through the tool is often part of how that gap gets closed — not necessarily sold outright, but used to make the underlying AI product better, which has its own commercial value to the company running it.

The Default Matters More Than the Setting Existing

Several major writing tools do offer a way to opt out of having submitted text used for model training — but the practical difference between an opt-in and opt-out default is significant, since most users never open settings menus after installing a tool. A setting that exists but defaults to "on" protects far fewer people in practice than one that defaults to "off," even though both technically offer the same choice on paper.

Tools That Explicitly Commit to Not Training on Text

Several established writing and grammar tools state directly in their documentation that submitted text is not used to train their AI models — a genuine, checkable commitment distinct from the vaguer "product improvement" language that shows up elsewhere. That specific phrasing is worth searching for directly in a tool's privacy policy or FAQ rather than assuming it from general reassurances about data handling.

An Emerging Alternative: On-Device Processing

Some browsers have begun shipping AI features that run entirely on-device rather than sending text to a server at all — proofreading, summarizing, and rewriting handled locally on the machine, with nothing transmitted anywhere. Where available, this sidesteps the training-data question structurally rather than through a policy promise, since there's no server-side copy of the text to train on in the first place — the same principle a fully local, browser-based tool applies by not having servers process the content at all.

What's Worth Checking Before Pasting a Draft

Before pasting anything meant to stay private — an early draft, a manuscript, client work — checking specifically for the phrase "not used to train" or "not used for model training" in a tool's privacy policy, rather than relying on general statements about data security, is the more reliable filter. Where that specific language isn't present, treating the tool as one where training use is possible by default is the safer assumption until proven otherwise.

---

Free writing tools aren't inherently unsafe, but "free" as a business model has to be funded somehow, and text used to improve an AI product is one of the more common ways that happens without a subscription fee. Checking for the specific, narrow claim — not used for training — rather than trusting broader language about data security, is what actually answers the question of what happens to a draft once it's pasted in.