Five services track the writing process but cannot prove AI was not used

Imagine spending several months writing a book, submitting the manuscript to a publisher and being accused of using AI. A detector has flagged a suspicious result. How do you respond? New services offer to preserve a record of your work: drafts, revisions, creation times and other traces. For a developer, this is a familiar task: building an event log. For a product manager, the challenge is harder: explaining to customers what conclusions that log actually supports.

WIRED examines this market through Workings, Blindproof, Wintertext, Human Ink and ObraVera. Their approaches differ, but the promise is the same: writers will have records that help explain how a text came into being. The problem is that a record of actions in an editor does not, by itself, establish who came up with the sentences.

What a writing history records

Some tools document the writing process. Others use timestamps, cryptographic hashes or typing patterns. These are different mechanisms with different capabilities, and treating them as a universal test of human authorship would be a mistake.

Take a simple example. A writer creates a scene, rewrites the dialogue several times and saves the result. The revision history shows the sequence of work. If a digital fingerprint of a particular version has been saved, it can later be checked against the submitted text. Such records can provide useful evidence when resolving a dispute.

But they do not establish that the writer never consulted a chatbot. The writer could have obtained a finished paragraph in another application, retyped it and then edited it. The log will faithfully record the typing and revisions. The source of the original paragraph will remain outside its view.

ObraVera's technical founder, Brady Ridgeway, describes this limitation explicitly: the platform can preserve material and a registered user's declaration about AI use. It cannot verify whether that declaration is truthful. It is important here to distinguish between recording a version, linking it to an account and verifying how the text was created.

Why a detector does not settle the question

Demand for these tools has emerged against a backdrop of tangible consequences from accusations. According to WIRED, Hachette canceled the US release of the novel Shy Girl and stopped distributing it in the United Kingdom following suspicions of AI use. Its author, Mia Ballard, denies writing the novel with AI assistance and attributes the changes to an editor's work.

The publication cites an assessment by Pangram: the tool classified about 78% of the text as generated. That is a detector's output, not an established proportion of machine-written text. Confusing the two is particularly dangerous when a book's publication and an author's reputation depend on the decision.

Authors Guild head Mary Rasenberger reports complaints from writers whose human-written work was classified as machine-generated by a detector. Pangram itself also acknowledges the limitations of tools that record the writing process: someone can manually type a model's response or use automated typing.

There is a problem in the other direction, too. A writer works across several files, dictates notes on the go and receives corrections from an editor. The application sees only part of that chain. A gap in the history does not mean AI created the missing passage. Requiring a continuous record could put an honest writer in a difficult position simply because of their usual working methods.

What developers and product managers should consider

If you are building an editor, a platform for writers or a content verification service, start with a precise description of the output. Which events do you record? Which files do you cover? What happens when text is pasted, dictated or edited collaboratively? Users should understand the limits of the record before they need to use it in their defense.

A separate question is who will agree to accept such a report. Publishing expert Jane Friedman doubts that publishers will agree on a common standard, especially if verification requires additional money and time. Buying an application therefore does not yet guarantee that its data will convince an editor or agent.

The practical takeaway for product teams: sell the specific ability to preserve and present a writing history. A promise to prove that AI was not used goes far beyond the capabilities of the tools described here. For organizations assessing writers, it makes more sense to treat revision histories and detector assessments as evidence to examine: each source leaves questions that require human judgment.