AI detection software does not tell a publisher whether you used ChatGPT. It tells them there is a statistical probability that you did — and that distinction should alarm every indie author who writes clean, polished prose. As the publishing industry scrambles to draw lines around generative AI, a growing number of agents, editors, and even academic publishers have started running manuscripts through detection tools. The current industry leader, according to analyst Jane Friedman, is a platform called Pangram. But here is the uncomfortable truth that rarely makes it into the discourse: these tools operate in probabilities, not certainties, and not all of them are built to the same standard. Some are significantly worse than others. For an indie author two years into a career, with a growing readership and maybe a small-press deal on the horizon, a false positive isn't a hypothetical — it's a reputational grenade.
Let's be precise about what's happening. Some literary agents now ask directly on submission forms whether AI was used to write or prepare any part of a manuscript.
But! Their reasons vary. Some are worried about copyright protection, which is a legitimate legal concern because the U.S. Copyright Office has made clear that purely AI-generated content is not copyrightable. Others simply do not want to work with writers who touch AI at all, full stop. But Friedman makes an important clarification that often gets lost in the noise: using AI for brainstorming, outlining, or editing does not automatically strip a work of copyright protection. The legal landscape is more nuanced than most social media takes suggest, and even agents and editors routinely get the basics wrong. If the gatekeepers are confused, imagine what a probability-based scanner is going to do with your carefully revised fourth draft.
Here is where my position diverges from the cautious, wait-and-see advice you'll find elsewhere: indie authors need to stop treating AI detection tools as neutral arbiters and start treating them as what they are — blunt instruments being deployed in a high-stakes environment before the science is ready. A tool that flags your manuscript at 68% probability of AI involvement is not evidence. It is not a ruling. It is a guess dressed in a confidence interval. And yet, for a growing number of publishers — particularly in academic and professional publishing, where detection tools are already being used on contracted work — that guess can delay a release, trigger a contract review, or permanently damage an author-publisher relationship. Indie authors operating without the buffer of a major agency are especially exposed.
This does not mean you should ignore the conversation. Contracts are evolving fast, and the clauses showing up in publishing agreements reflect genuine anxieties. The Authors Guild has introduced model contract language that addresses AI on multiple fronts: prohibiting publishers from using a book to train AI models without express written consent, requiring author approval before AI-generated audiobook narration or AI translation, and blocking fully AI-generated cover art without permission. These are defensive measures worth fighting for in any deal. But notice the direction of protection — those clauses protect the author's work from being consumed by AI, not from being falsely accused of producing it. That gap matters, and no one is filling it yet.
The misinformation problem compounds everything. Friedman warns bluntly that if you primarily learn about generative AI from social media or average online commentary, you are likely misinformed. The majority of what circulates daily in writer forums is outright false. That means the indie author who diligently avoids AI but writes in a style that happens to trigger a detection algorithm — clean syntax, consistent tone, limited idiosyncrasy — has almost no reliable public resource to understand their risk or defend their work. The ecosystem is punishing the wrong people while the actual legal and technical frameworks remain unsettled.
So here is one concrete thing you can do this week: run your own finished manuscript through Pangram or a comparable top-tier detection tool before you submit it anywhere. Document the results. Screenshot them. Save them with a timestamp. If your work comes back clean, you have a dated record. If it flags sections, you know exactly which passages to annotate with your revision history, drafts, or notes.
This is not about proving your innocence — you shouldn't have to. It's about building a paper trail in an industry that is increasingly making decisions based on tools it doesn't fully understand. Protect your manuscript the same way you'd protect any other piece of intellectual property: with evidence, not trust.