Andrea Bartz has published five thrillers since 2019, secured a deal with Netflix, and earned a spot in Reese Witherspoon’s book club. She also led a class-action lawsuit against Anthropic which resulted in a $1.5 billion settlement last year. Now, the writer faces a different problem: the risk of being accused of using artificial intelligence to write her own books.
Recent incidents
In March, Hachette Book Group cancelled a US publication run for the horror novel Shy Girl after reports suggested it was largely AI-generated. The company had already stopped selling the book in the UK. Pangram, an AI detection service, stated that 78 percent of the text appeared to be generated by an algorithm.
Mia Ballard, the author of Shy Girl, denies using AI. She claims an editor made AI-assisted changes to a previously self-published version. Ballard is pursuing legal action. Since then, other fiction writers have faced similar suspicion after winning prizes or selling novels for high prices. All have denied the allegations.
Bartz says the damage occurs once accusations begin. She fears a witch hunt will follow the current trend. The inevitable presence of AI in publishing allows anyone to accuse anyone else of using it. Currently, there is no agreed standard for proving human authorship.
New software options
Several new tools aim to create a record of an author’s writing process. Platforms such as Workings, Blindproof, Wintertext, Human Ink, and ObraVera offer features like encryption, fingerprinting, and unique typing codes. These systems generate a paper trail to defend against claims of AI use or to prove that text was not scraped for training data.
Mary Rasenberger, CEO of the Authors Guild, says the organisation cannot recommend one tool over another. She notes that many companies have approached the Guild to pitch their services. Rasenberger views the requirement to preserve receipts as insulting and unnecessary.
She argues that selling words and imagination does not make sense if AI is involved. Editors of literary fiction and nonfiction often state they never see AI-generated submissions because they can spot them. Rasenberger believes AI slop is less common in publishing than people think.
Rasenberger also worries about a climate of distrust among authors. She advises against accusing peers because it harms the literary community. She is concerned about overconfidence in detection tools, which can produce false positives.
Pangram claims a false positive rate of 0.01 percent for creative writing. However, the Authors Guild has heard of many instances where human-authored work was flagged as AI-generated. Various factors can cause a detector to make this error.
Process-tracking apps are not foolproof. The competition between AI and detection software creates a cycle of vulnerabilities and fixes. Max Spero, CEO of Pangram, notes that users find the Google Docs tracker helpful but warns it is not a perfect solution.
Spero explains that someone could hand-type AI text or use an auto-typer. The messy nature of writing creates further issues. Writers often use multiple files, exchange drafts, and make voice notes. Not all of this activity gets captured by writing process tools.
These tools may focus more on copyright than confirming human authorship. Brady Ridgway, technical founder of ObraVera, says his platform traces reuse of material stored in their database. A time stamp and crypto hash allow the system to show when words were created by a registered user.
Ridgway notes that the software records a declaration of whether text is human-only or AI-assisted. It cannot confirm if that statement is true. This was the case in May when a writer accused of using AI to win a Commonwealth Short Story Prize was still considered the person who submitted the work.
Industry hurdles
Jane Friedman, an author and publishing expert, argues that the era of AI scandals may be short-lived. She suggests that improvements in detectors and the recent consequences for passing off AI work will change behaviour. Friedman tells WIRED that agents and publishers are unlikely to implement verification methods that cost extra money or time.
She predicts that reaching agreement on a shared standard will be difficult. Publishing is slow to adopt new technologies, especially those created by industry outsiders. Even if diverse publishers coalesce around a theoretical form of authentication, logistical problems remain.
Friedman points out the difficulty of integrating benchmarks into existing production models. She also mentions the risk of faked or manipulated certification. Authors may pay for peace of mind, but they cannot trust that agents and publishers will accept proof from any given app.
Reader reaction
Without a clear solution, writers must fend for themselves. Bartz is alarmed by zealous readers who suspect deception. She has heard of people getting review-bombed on Goodreads because a stranger felt a book sounded like AI. Others have stated they will not read anything published after 2022.
Bartz’s forthcoming book, Delete Your Account, tackles generative artificial intelligence directly. The protagonist is a blocked writer who uses an AI chatbot to imitate her favourite thriller. She becomes trapped in the story she created.
Bartz wonders if satirising AI puts her in a precarious position. She has written portions meant to sound like LLMs, including transcripts and marketing materials from a made-up company. She does not know what to do if any of that is flagged as AI-written.
