OpenAI's AI text classifier lasted seven months. Launched in January 2023, it correctly identified 26% of AI-written text while falsely flagging 9% of human writing as machine-generated. The company discontinued it that July. "We are working to incorporate feedback and are currently researching more effective provenance techniques," OpenAI announced upon pulling the tool. (Source: OpenAI, primary)

Publishers kept using detection tools anyway.

Hachette cancelled a contract for Shy Girl over alleged AI use. The accusation was later disputed. The incident became emblematic. Houses that sue Meta and Google for unauthorized AI training of their own backlists now deploy the same unreliable classifiers to void author contracts. Authors accused of AI assistance cannot meaningfully dispute findings that toolmakers acknowledge are wrong roughly one time in ten. Academic research confirms these tools produce "high false discovery rates" across multiple platforms. (Source: ScienceDirect, primary)

The Authors Guild documented the failure in 2026. Testing five commercial tools against human articles published between 2020 and 2022—predating ChatGPT—the guild found Sidekicker.ai flagged every sample as 71% to 100% AI-generated. ZeroGPT scored a Pulitzer Prize announcement for Louise Erdrich at 76.3% AI-written. Grammarly and Pangram returned 0% for the same text. Originality.ai also returned 0%. The tools contradicted each other absolutely. Their only certainty was mutual unreliability. (Source: Authors Guild, primary)

Hachette, Penguin Random House, and HarperCollins litigate against Meta for "market dilution" through AI training. These same houses have been accused of using undisclosed AI for translations, marketing copy, and manuscript summaries. Judge Vince Chhabria, presiding over Kadrey v. Meta, noted in 2025 that "no other use...has anything near the potential to flood the market with competing works the way that LLM training does." Publishers embraced this formulation in court while ignoring its implications for their own enforcement practices. (Source: Court order, primary)

Their fears have basis. An analysis of 14,419 self-published genre-fiction titles on Amazon from 2023 to 2026 found the market added selling books 19.2 times faster than revenue grew (8.9-fold), indicating severe price compression and per-title revenue decline. The "AI slop" discourse—invoking aesthetic disgust at machine-generated prose—obscures this structural reality. The anxiety is about labor displacement through market dilution, the same phenomenon publishers cite when seeking damages from technology companies. (Source: arXiv, primary)

The guild's response: a "Human Authored" certification program, $10 per title or free for members, requiring affidavit and documentation. Bureaucratic process replaces technical reliability. Authors must prove their innocence; publishers need only point to a contested algorithmic score. No standard governs detection tool deployment in contract disputes. No appeal mechanism exists for false positives. The 9% error rate that doomed OpenAI's classifier in 2023 now governs professional fates without oversight, applied by parties with financial incentive to find violations and no liability for mistaken findings.

New contract clauses require AI disclosure. They do not specify which tools verify compliance, or what recourse authors have when those tools err. The memoirist whose human-written Fortune op-ed triggered recent attention has a publisher that, like others, has not publicly committed to ignoring detection results that OpenAI itself abandoned as insufficiently accurate.

The Authors Guild tested its own 2021 Pulitzer announcement. Sidekicker.ai returned 100% AI-generated. ChatGPT would not exist for another sixteen months.