Authors seek ways to prove their books were written by humans

Writers have begun using services that record how a text was created to defend themselves against accusations of using AI and establish authorship. The trigger was a series of controversies around the novel Shy Girl and other books: Hachette Book Group cancelled the novel’s US release and stopped publishing it in the UK after suspicions that it had been written with AI. But these tools do not provide reliable proof, and publishers may not want to spend time and money establishing common verification rules.

Authors seek a way to prove their books were written by humans

Controversies and accusations

Andrea Bartz is the author of five thrillers published since 2019. She became a New York Times bestselling author; her work attracted interest from Netflix and Hulu, and Reese Witherspoon’s book club recommended one of her books. Bartz was also the lead plaintiff in a class-action lawsuit against Anthropic, which was accused of downloading millions of copyrighted books from pirate sites to train its models. Last year, Anthropic agreed to pay $1.5 billion — the largest copyright settlement in US history.

As writers tried to protect their work from being used to train AI, another problem emerged: authors themselves began facing suspicion that they were asking artificial intelligence to write their books.

In March, after social media discussions and a New York Times report, Hachette Book Group cancelled the planned US release of the horror novel Shy Girl and stopped publishing it in the UK, where the book had already come out. AI detector Pangram estimated that about 78% of the text was AI-generated. The service considers its detector one of the best in the industry, but acknowledges that detectors are not infallible.

The novel’s author, Mia Ballard, denies using AI to write it. She says AI helped an editor make changes to an earlier version of the book, which she had published at her own expense. Ballard also said she intends to take legal action. After the story, other fiction writers came under suspicion: one had won an award for a short story, while publishers had offered millions for another author’s crime novel. Both rejected the accusations.

Bartz believes accusations cause harm simply by being made. Amid the Shy Girl controversy, she feared that the possibility of AI use would lead to a witch hunt in which almost anyone could be accused. She says authors do not yet have a generally accepted way to prove that a work was written by a human.

What the services record

New tools are targeting this gap: Workings, Blindproof, Wintertext, Human Ink and ObraVera. They record the writing process; some encrypt materials or create digital fingerprints of the text and typing patterns. These records may help establish where a text came from, whether it was written entirely by a person or with AI assistance. They may also be useful if a text is included in AI training data without its author being credited.

Mary Rasenberger, CEO of the Authors Guild, a professional organisation focused on free expression and copyright protection, says the organisation cannot recommend one service over another. But she wants writers to know these tools exist as a kind of insurance. According to Umair Kazi, the Authors Guild’s director of policy and advocacy, many companies have offered their services to the organisation.

Rasenberger considers it humiliating and burdensome to require authors to gather proof of their own authorship. In her view, keeping such records should not be mandatory. She also notes that writers value their own voice, imagination and creativity, so AI-generated texts are less common in publishing than some might think. Rasenberger says editors of fiction and literary nonfiction have told her they see almost no AI-generated writing and would notice it.

At the same time, Rasenberger is concerned about the climate of distrust. She urges readers not to accuse other authors without sufficient grounds, since doing so harms the literary community. She also warns against putting too much faith in detectors. Pangram says its false-positive rate for creative writing is 0.01%, but the Authors Guild, Rasenberger says, has heard of cases where human-written texts were flagged by the service as AI-generated. Different features of a text can affect the result.

Even software that records the writing process does not offer complete protection: ways to bypass it and fix vulnerabilities keep changing. Max Spero, Pangram’s co-founder and CEO, says the company’s service records work in Google Docs and many users find it useful. But, he says, AI-generated text can be retyped by hand or entered into a document automatically. Authors also work across multiple files, exchange drafts and record voice notes on the go — actions the software may not capture.

Proof of authorship does not prove human involvement

Some services are better suited to establishing a text’s provenance and protecting copyright than to determining who actually wrote it. Brady Ridgeway, ObraVera’s technical founder, says the service creates cryptographic fingerprints of text characteristics and voice recordings. A registered user can save material in a database and later search for matches if their text appears elsewhere or someone claims it as their own. A timestamp and cryptographic hash can link a text to a person and the time it was created.

However, users must say for themselves whether they used AI. Ridgeway points to the May controversy involving a writer accused of winning the Commonwealth Short Story Prize with an AI-generated story. He says no one disputed that the writer had submitted the story. The service could have confirmed only that, on a particular date, a registered user stated that the text was written by a human or created without AI assistance — not whether the statement was true.

Jane Friedman, a writer and publishing-industry expert, believes that introducing systems to verify human authorship may be impractical, and that AI controversies could lose relevance over time. In her industry newsletter, The Bottom Line, she wrote that detectors, including Pangram, are improving, and authors have already seen the consequences of trying to pass off AI-generated text as their own.

Speaking to WIRED, Friedman suggested that agents and publishers are unlikely to adopt verification if it requires extra money or time. Another obstacle is the lack of a common standard: in her view, it is difficult to bring the publishing industry together around new technologies, especially those developed by outside companies.

Even if publishers agree on an authorship verification system, there will still be challenges integrating it into existing processes, as well as the risk of forged or altered certificates. Friedman believes extra costs or delays would make such a system unattractive. Writers may pay for peace of mind, but they cannot be sure that agents and publishers will accept verification from every app.

Suspicion reaches future books too

With no common solution in place, authors have to defend themselves. Bartz is concerned by how sharply readers react to the possibility of AI-enabled deception. She has heard of cases where Goodreads users flooded a book with low ratings after one reader raised suspicions. Some, she says, have declared they will not read books published after 2022 to avoid generative AI text. Bartz fears this approach could harm the entire publishing industry.

Generative AI will be a central theme in Bartz’s book Delete Your Account, scheduled for release in 2027. Its protagonist is a writer facing a creative crisis. She takes a dangerous shortcut and asks a chatbot to imitate her favourite thriller, then finds herself inside the original story and facing a deadly threat.

Bartz worries that the satire itself could attract suspicion. The book includes transcripts she made up and promotional material for a fictional company deliberately written to resemble AI-generated text. She does not know how she will defend herself if detectors flag those passages as AI-generated.

Dataist’s take: Tools that record the writing process can confirm when a text was created and link it to a particular user, but they cannot prove that the user did not rely on AI. Until publishers agree on a common standard and are willing to devote resources to verification, these services remain an individual precaution — and do not guarantee that accusations can be avoided.