Authors Can Defend Against AI Allegations With ‘Paper Trail’ Tools. Is It Enough?

Save Story Save this story Save Story Save this story Andrea Bartz is a prolific novelist with an impressive résumé . Since 2019, she has published five thrillers, become a New York Times bestselling author, had her work optioned by the likes of Netflix and Hulu , and picked up an endorsement from Reese Witherspoon’s influential book club. She also has a legal claim to fame: Bartz was the lead plaintiff in a class-action suit against Anthropic that alleged the company had downloaded millions of copyrighted books from pirate sites to use as training data for its models. Anthropic last year agreed to pay $1.5 billion in the largest copyright settlement in US history . But while Bartz and many other writers were on guard against this plunder of their work, another AI threat was brewing. What happens when authors are accused of relying on artificial intelligence themselves? This past March, the publishing industry found out: Following widespread speculation on social media that a horror novel titled Shy Girl was largely AI-generated, and a Times article about those claims, Hachette Book Group canceled an upcoming US publication run for the novel and discontinued it in the UK, where it had already been released. According to Pangram , a robust but by no means infallible AI detection tool, some 78 percent of the text was AI-generated. Mia Ballard, the author of Shy Girl , has denied allegations that she used AI to write the novel, claiming that an editor was responsible for AI-assisted changes to a previously self-published version of the book. She has also said that she is pursuing legal action. Since that scandal broke, several more fiction authors have run afoul of AI suspicions, one after winning a short story prize , another after publishers bid into the millions for his crime novel. Both have denied these allegations as well. But Bartz notes that once the accusations start, the damage is done. As the Shy Girl scandal erupted, she says, “I kind of immediately felt like this was going to lead to a potential witch hunt.” The unavoidable emergence of AI in publishing, she explains, will “have the unintended effect of leading anyone to be able to accuse anyone of using AI,” and at the moment, authors have little recourse in defending themselves against these charges. “There’s no agreed upon standard,” Bartz says, in proving human authorship. The developers of a host of new software tools hope to change that. Platforms including Workings , Blindproof , Wintertext , Human Ink , and ObraVera promise the means to track and document an author’s writing process, in some cases encrypting and fingerprinting the written work or typing rhythms with unique codes, in order to produce a record of authorship (whether entirely human or AI-assisted). The idea is not just to create a paper trail that writers can point to if someone questions whether they used an LLM but to demonstrate provenance when, for example, text is scraped into AI training sets without attribution or other identifying metadata. Mary Rasenberger, CEO of the Authors Guild, a professional organization focused on free expression and copyright protection, says that the guild can’t recommend any one of these tools over another but wants writers to know they exist as something of a security net. (Many of the companies have approached the Authors Guild to pitch their services, according to Umair Kazi, the guild’s director of policy and advocacy.) At the same time, Rasenberger tells WIRED, it’s “insulting” and a “burden” to have to jump through this hoop, and “we don't feel it should be required” for authors to preserve receipts. “If what you're selling are your words and your voice and your imagination, your creativity, using AI doesn't make sense,” Rasenberger says, noting that authors’ pride in their own work means AI slop is far less common in publishing than one might think. “I've talked to editors, particularly editors of literary nonfiction and fiction,” Rasenberger says, “and some of them will say, ‘I never see AI-generated submissions, I'd know it if I did.’” Rasenberger worries that authors are up against a climate of distrust—they should “refrain from accusations of other authors because it just does not support a healthy literary community,” she says—and overconfidence in AI detection tools, which can be fooled and have been known to produce false positives. While Pangram claims an “industry-leading” false positive rate of 0.01 percent when used on creative writing, Rasenberger said the Authors Guild has “heard of lots of instances where people have put their own human-authored work in Pangram and gotten a result back saying it was AI-generated.” There are a variety of factors, she adds, that can “make an AI detector think something's AI-generated when it's not.” By the same token, a process-tracking app isn’t foolproof; the AI arms race means there’s an ongoing cycle of vulnerabilities and fixes on both sides. “We offer a product that [tracks writing] in Google Docs, and many users find it helpful,” says Max Spero, cofounder and CEO of Pangram. “With that said, ‘paper trail’ tools are not a perfect catch-all. For example, someone could hand-type AI text, or use an auto-typer.” Spero points out that the messiness of writing itself creates further problems. “The average writer doesn't just sit down and type out a book,” he says. “They work with multiple files, send drafts back and forth, use voice notes while on the go. Not all of this gets captured in writing process tools.” The tools, meanwhile, may be more geared toward copyright concerns than confirmation of human or AI authorship. Brady Ridgway, technical founder of ObraVera—which can cryptographically “fingerprint” patterns in writing as well as audio voice recordings—says his platform is designed to trace reuse of material you have stored in their database as a registered human user. With a time stamp and a crypto hash, he explains, if the content “crops up later somewhere it's not supposed to be, or somebody else claims it as theirs, we can search the database and say, ‘These words were created at this time, by this person.’” That may demonstrate provenance and ownership, Ridgway says, yet it’s up to the registered individual to declare whether they used AI or not in the creation of the text. He refers to the kerfuffle in May over the writer who was accused of winning a Commonwealth Short Story Prize with an AI-generated submission. “Everybody seems to agree that he's the person who submitted the short story,” Ridgway says. “That's all that the software would do. It would say that somebody on this day made a declaration that it was human only or not AI-assisted.” It can’t confirm whether that’s a true statement or not. Some argue that trying to institutionalize vetting systems for human authorship is simply impractical—and may soon be a moot point anyway. Jane Friedman , an author and publishing business expert, recently argued in her industry newsletter The Bottom Line , that the era of AI scandals is likely to be short-lived, given the improvements of Pangram and similar detectors. Moreover, she wrote, authors have now seen what may happen if they attempt to pass off AI-generated writing as their own. Friedman tells WIRED that agents and publishers are unlikely to implement any method of verifying human authorship that “costs extra money or extra time,” not least because it would be impossible to get them to commit to a shared standard. She predicts that “it will be like herding cats to get agreement on this, because publishing is slow to adopt new technologies, especially new technologies developed by industry outsiders.” Even if diverse publishers with highly variable needs and resources were to coalesce around some theoretical form of authentication, hurdles remain. Friedman alludes to the logistical problems of integrating this benchmark into existing production model