Posted on The Velvet Quill by KL Adams

Here is something that should make every writer stop cold.

Andrea Bartz is a published thriller author. She writes literary-quality fiction. She has been doing this professionally for years. Earlier this year, she ran some of her own work through an AI detection tool just to see what would happen.

It came back 82% AI-generated.

Her response, posted on Substack, attracted dozens of writers who had experienced the same thing. One of them commented, “I guess that’s what happens when your books were stolen to program AI.” Another had been flagged at 76%. A third at 91%.

None of them had used AI to write their books.

This is the problem nobody wants to talk about clearly — because one side of the conversation is busy building mobs and the other side is too scared to speak up.

What AI Detection Actually Is

Before we go further it helps to understand what these tools are actually doing, because most people assume they work like plagiarism checkers. They don’t.

AI detectors don’t compare your text to a database of known AI writing. They analyze statistical patterns — specifically, how predictable your word choices are and how consistent your sentence rhythm is. The technical terms are “perplexity” (how surprising your word choices are) and “burstiness” (how much your sentence lengths vary).

The theory is that AI writing tends to be smooth and predictable in ways that human writing isn’t. Humans write in bursts — complex sentences followed by short ones. AI produces more even, measured prose.

The problem is that this description also fits plenty of human writing.

Formal prose — the kind trained professionals write after years of practice — tends to be measured and consistent. Literary fiction that has been carefully edited tends to be smooth. Writing by non-native English speakers, who often rely on learned patterns and familiar constructions, tends to read as more predictable to a detector, even when it is entirely human.

A 2023 Stanford study found that AI detectors misclassified over 61% of essays written by non-native English speakers as AI-generated, while achieving near-perfect accuracy on essays written by native English speakers. Over 25 major universities — including Yale, MIT, Vanderbilt, Northwestern, UC Berkeley, and NYU — have banned or restricted AI detection tools precisely because the false positive problem is too serious for high-stakes accusations.

These are the tools being used to end authors’ careers.

Meet Pangram — The Company at the Center of Everything

At the center of almost every major AI authorship controversy in publishing in 2026 is a single company: Pangram.

Pangram’s CEO, Max Spero, calls himself a “slop janitor” on Twitter/X. The pattern around Pangram’s involvement in publishing controversies goes like this: readers suspect a problem, Pangram validates the suspicion, the outrage spreads, and Pangram — which sells its detection services to publishers and institutions — benefits from the resulting demand.

Pangram flagged the Shy Girl novel that Hachette ultimately canceled. Pangram flagged the Modern Love column that went viral on social media. Pangram’s own study scanned 250,000 newspaper articles and claimed 9% of U.S. news contains AI-generated text — including op-eds from writers at the New York Times, Wall Street Journal, and Washington Post — and then Spero used those findings to publicly name and accuse individual journalists on LinkedIn and Twitter/X.

When the Wall Street Journal’s opinion editor ran the flagged articles back through Pangram and got different scores, he published a combative response. Pangram’s reaction was to accuse him of dodging the real issue.

This is not a neutral actor operating in good faith.

But Here’s Where It Gets Complicated

Pangram is also, according to independent research, genuinely the most accurate AI detector currently available.

A University of Chicago Booth School of Business study tested four major detection tools and found that Pangram achieved near-zero false-positive and false-negative rates on medium- to long texts — significantly outperforming every competitor. An independent economist who conducted his own evaluation said flatly: “This narrative that we shouldn’t use AI detection doesn’t seem to hold anymore.”

So Pangram works. On aggregate data, at scale, analyzing substantial passages of text, it performs extraordinarily well.

The problem is that “extraordinarily well at scale” and “reliable enough to destroy an individual person’s career” are not the same standard. And Pangram is being used for both.

One researcher who studies AI detection put it precisely: “If you copy-paste chunks of ChatGPT with minimal edits, then Pangram is fairly accurate. If you significantly edit AI-generated text, then it becomes human, and this is a harder problem in general.” That matters because in the real world, AI use exists on a spectrum — and so does the detector’s accuracy.

The Real-World Consequences

The authors who are being caught in this are not the ones who fed a prompt into ChatGPT and hit publish. Those cases are usually obvious and unambiguous. The people being destroyed by this are in the middle — writers whose prose style happens to be smooth and consistent, writers who edited heavily and efficiently, writers who come from non-English-speaking backgrounds, writers whose earlier books were used to train AI models without their consent, and whose own voice now registers as “machine-like” to the very tools trained on their stolen work.

That last one is almost too grotesque to sit with. An author’s writing is scraped without permission to train AI. The AI learns to write in patterns that resemble the author. Detection tools are then trained on AI output. Now the author’s own voice — the one that was stolen — flags as artificial.

A study published in August 2025 found that while AI detectors have useful applications, “the possibility with which these detection tools misclassify original, unaided creative work as AI-generated raises significant concerns,” and that “none achieved 100% reliability.”

None. Not one.

What This Means for You

If you are a writer in 2026, here is what you need to know:

Keep your drafts. Version history, timestamps, handwritten notes, outlines, revision files — all of it. The only way to prove human authorship after the fact is evidence of process. Keep yours.

Know your style. If you write in a formal, measured register — if you tend toward smooth prose, consistent rhythm, careful construction — you are statistically more likely to trigger a false positive. This doesn’t mean change how you write. It means know your risk.

Don’t trust a screenshot. A single detector result from a single tool is not evidence of anything. Reputable institutions have concluded that no AI detector is reliable enough to be the sole basis for an accusation. If someone waves one at you, that is the response.

Read your contracts. Some publishers are now including language about AI use in their agreements. Know what yours says before you sign. Know what counts as a violation. Know what the process is if you are accused.

The mob is not the truth. Several of the highest-profile AI accusations in publishing in the past year have involved a Reddit post going viral, Pangram validating the suspicion, and the resulting pile-on moving faster than anyone could respond. That sequence is not evidence. It is a pattern.

The Harder Question

The conversation about AI detection is currently focused almost entirely on process — did this person use AI to write this thing? That is a reasonable question. It is not the most important question.

The more important question is what we lose when suspicion becomes the default. When writers are afraid to write efficiently. When having a distinctive, polished voice becomes a liability. When readers’ opportunities to build community and develop empathy through engagement with literature are replaced by algorithmic judgment about whether the human on the other side of the book is real.

Andrea Bartz ran her own work through a detector and got 82%. She is unambiguously a human writer. The tool was wrong about her.

It might be wrong about you, too.

Have you run your own work through an AI detector? What happened? Tell me in the comments.

KL Adams is a literary blogger and fiction writer specializing in dark fantasy, vampire fiction, and paranormal romance. Follow on WordPress, Inkitt, and Wattpad for reviews, reading lists, and stories that haunt you long after the last page.

📚 Building your own dark fantasy world? Grab the Dark Fantasy Worldbuilding Kit on Gumroad — worksheets, magic system builders, character prompts, and storytelling tools designed specifically for dark fantasy, gothic fantasy, and romantasy writers.