Good Writing Is Now Suspect — Here's How to Defend Yourself Against False AI Accusations

Humans can’t reliably detect AI-generated writing from human-produced writing, but most think they can. No app can reliably do so either — my newspaper and magazine articles from 10, 15 years ago have been flagged by Grammarly, ZeroGPT, Pangram and others I’ve tested as various percentages of AI generated. Grammarly, which recently had to backtrack after stealing the editorial voices of various writers and offering them to its customers — without the writers’ permission — flagged entire paragraphs as AI generated, paragraphs written decades ago. I suspect it’s my use of the em dash, but that’s a hill I will die on. It’s mine, and the AI-overlords can drag it out of my cold, dead hands. My use of cliches is probably also a trigger.

I don’t have great answers for this. We live in an era where good writing is immediately suspect to the masses, and there’s little we can do to guard ourselves against false accusations — little that isn’t entirely disruptive to our workflows. Protecting yourself will likely require some changes to how you work, but they don't have to be disruptive.

• Version control. Keeping a version history lets you later produce a trail of your evolving work. The simplest way to do this is to number your files as you go, and use a different file for every set of revisions. For example, each day before you start work, copy the previous file, give it a new version number and do that day’s work in that file (I use a form of semantic versioning favored by software developers). Keep all of the historical drafts as their own files. You can automate this process through GIT, though it requires varying levels of technical skills, depending on how you do it. For work where evidence is critical, like chapters produced under a six-figure book contract or legal filings, storing these documents on servers whose timestamps can’t be modified is stronger evidence, but so is file hashing (a process that also requires technical skills). File versioning is more difficult if you’re using an app that abstracts the file system away, like a simple notes app.
• Process artifacts. A lot of fiction writers collect this epherma already, because they believe it’s useful for marketing. It can also be evidence: screenshots or exports of research notes, browser history exports during research phases, library loan records, interview recordings or correspondence with sources, margin notes in physical books. The goal is to show a research-to-draft-to-publication pipeline that AI use wouldn't produce.Apps that take time-shots of your screen and can be confined to the relevant windows are a good middle ground. Something like TimeSnapper or Hunchly (great for journalists as a research tool, too) can create reconstructable workflows and the process artifacts mentioned above.
• Metadata and stylometric consistency. A long publishing history itself is evidence (whether accepted as such or not is another question), especially when stylometric analysis shows consistent voice across decades. Worth noting that stylometric analysis, unlike AI detectors, actually has a scientific track record in authorship attribution but is very far from perfect. Writers with backlists could preemptively commission or self-run stylometric comparisons showing continuity of voice, but many writers at risk of false accusations are writing to a publication’s voice, not their own. Not everything I write, for example, sounds like a Wired front-of-book piece or an inverted-pyramid-style breaking news article, but a large volume of my work does.
• Handwriting. Really, Natalie Goldberg would be proud. Writing drafts by hand produces process artifacts that are difficult for generative-AI to convincingly fake, and there are other benefits, too. “Handwriting activates a broader network of brain regions,” and it slows you down, giving your brain more time to process and more time to process deeply. And if you can’t write in cursive, that’s even better — studies show the fluidity of cursive over block letters recruits an even broader network of brain regions. At any rate, it’ll be difficult for most people to argue with a hand-written first draft.

No single piece of evidence here will protect you from being falsely accused or losing a contract or job — and none of these solutions are currently easy or workflow-insensitive. They require thought, planning and an organized storage method for later retrieval. The more different pieces of evidence of your process you collect, the better off you’ll be.

And it’s worth noting — there are significant digital privacy costs with some of these methods. Even facing a public accusation, it’s not necessarily wise to publish private emails or even drafts in progress to satisfy an online mob. Screenshots and screen recordings come with even more dire privacy risks. But where it counts, with people who can be trusted (like editors and bosses and judges), proactively preparing to defend against the accusation can make all the difference.