AI Detector False Positives: What the 2026 Evidence Actually Shows

Quick answer

False positives from AI detectors are real and documented. The Authors Guild examined five detection tools on ten of its own articles dating from 2022 or earlier, a period before generative AI was available. Two of those tools identified several of those pre-AI pieces as predominantly AI-generated. A prize-winning short story published in Granta scored 100% AI-generated on a detector with a very low false-positive rate, and the accusation stuck for weeks before the prize committee opened a review. None of this means detection is worthless. It means a single score is a signal to investigate, not a verdict, and the newest research explains exactly why polished writing keeps tripping the alarm.

What happens when you run pre-AI writing through today's detectors

May 2026 saw the Authors Guild handpick ten articles from its archive. These pieces dated to 2022 or earlier. A round of checks by five common AI detectors, ZeroGPT, Originality.ai, Sidekicker.ai, Grammarly, and Pangram, examined each. All writings were older than ChatGPT's launch. A properly working detector should indicate near-complete absence of AI influence everywhere.

Every article slipped past Pangram and Originality.ai with a score of 0% or 1%. Grammarly came close, flagging two pieces at 7% and 9%. The other two tools did not:

Across a sample of ten articles, ZeroGPT returned scores ranging from 5.3% to 76.3%. No discernible pattern separated the lower scores from the higher ones. It flagged the Authors Guild's obituary for Joan Didion at 66% AI-generated and its note congratulating Louise Erdrich on her Pulitzer Prize at 76%. Sidekicker.ai flagged every entry as predominantly AI-written, with scores from 71% to a full 100% on two of them, on text written and published years before the technology it claims to detect even existed.

The Guild's own framing of the result: "a false sense of security." Publishers relying on a single score to make an authorship call, it argues, are one inconsistent tool away from a wrongful accusation, and the organization is now advising members to keep dated drafts and version history on hand as a defense.

A live example: the prize-winning story that scored 100%

A month earlier, "The Serpent in the Grove" by Jamir Nazir was named a regional finalist for the Commonwealth Foundation Short Story Prize and published in Granta, out of more than 7,800 entries. Wharton professor Ethan Mollick ran the story through Pangram, a detector he'd separately verified has a very low false-positive rate, and it came back 100% AI-flagged (Literary Hub, "A prize-winning story published in Granta was (very likely) written by AI"). The author's public footprint was thin outside a single 2018 self-published poetry collection and a recent LinkedIn history of posting about generative AI. The Commonwealth Foundation confirmed it is reviewing its selection process.

This case cuts the other way from the Authors Guild test. There, a well-regarded detector wrongly flagged human writing. Here, a detector many consider comparatively reliable appears to have caught something real. Both outcomes are true at once, and that's the actual state of AI detection in 2026: no tool is uniformly wrong or uniformly right, which is exactly why one score alone should never be the whole investigation.

Why polished writing sets off the same alarm as AI

Detectors work by measuring perplexity (how statistically predictable the word choices in a passage are) and burstiness (how much that predictability varies from sentence to sentence). Large language models produce low-perplexity, low-burstiness text because they're built to select likely next words. The problem the Authors Guild names directly: experienced human writers, after years of editing toward clarity and economy, also produce low-perplexity, low-burstiness prose, not because they're imitating a model, but because the models were trained on exactly that kind of polished writing in the first place.

That overlap is the same mechanism behind an older, better-documented pattern: Stanford researchers testing seven detectors on TOEFL essays from non-native English speakers found a 61.22% false-positive rate, against near-perfect accuracy on essays from native-English eighth graders (Stanford HAI; underlying paper: Liang et al., arXiv) — a finding we've covered in more depth here. Simpler, more formulaic phrasing reads as "predictable" to a detector whether the reason is a second language or twenty years of professional editing. A March 2026 mathematical paper goes further, arguing this isn't a bug any future model can fix: any text-only detector with real detection power will necessarily misclassify some real human writing, because human writing styles overlap with AI output by population, not by individual intent (arXiv:2603.20254).

It isn't only fiction — the same tension shows up at population scale

NeurIPS's 2026 Position Paper Track required submissions to be substantially human-written. The track partnered with Pangram to screen 969 papers for compliance. Pangram's default settings identified 28.2% of submissions as 100% AI-generated. The organizers desk-rejected 178 papers and requested evidence of human authorship from another 123.

Using reduced text windows, the teams re-evaluated identical papers. The share scoring 90-100% plummeted from 42.7% to 12.7% under more detailed scrutiny, and they built a control comparison against pre-ChatGPT conference papers (0% flagged) to sanity-check the tool. That's the honest-broker version of using a detector: treat a high score as a trigger, not a verdict, check it against a reference baseline, and offer an appeal route to authors before turning the number into confirmation. Most schools and publishers skip every one of those steps.

The law is about to assume detection works, before the research says it does

The EU AI Act's Article 50 transparency obligations take legal effect on August 2, 2026. Providers of generative AI systems must mark synthetic outputs in machine-readable form, and anyone publishing AI-generated or AI-manipulated text on matters of public interest without human editorial review must disclose it (European Commission, official Article 50 FAQ). The compliance model leans on AI providers watermarking their own outputs at the source, not on third-party detectors catching unlabeled text after the fact. That's a meaningfully different bet than what schools, publishers, and prize committees are already doing today: paying a vendor to guess.

What to do if your own writing gets flagged

• Ask for the specifics. Which tool, which version, what threshold triggered the flag, and what the raw score actually was. Institutions that have survived legal challenges over false-positive findings all point to this as step one.
• Keep your process, not just your product. Drafts, version history, and time-stamped revisions are the evidence that actually settles a dispute, because a detector score alone isn't direct proof of authorship either way.
• Run it through more than one tool, and treat disagreement as information. The Authors Guild table above is the clearest illustration on record of why a single score from a single vendor isn't enough to act on.
• Push back on the policy, not just the score. If an institution is treating one AI-detection percentage as the sole basis for a finding, point them to their own stated policy — most major detector vendors, including the ones used in the Authors Guild test, publicly say their own scores shouldn't be used that way.

A detection score might prove false before it becomes someone else's evidence. Feed a sample to TheChecker.AI's free demo. Don't skim past the headline number, read the sentence-level breakdown behind it.

FAQ

Can a completely human-written piece score 100% AI-generated? Yes. The Authors Guild test found that Sidekicker.ai assigned a 100% AI-generated score to two articles published during 2020-2021, years that predate the writing tools the detector claims to catch. A single tool marking an article as fully generated calls for further scrutiny, not proof by itself.

Is one AI detector more accurate than another? Test results vary by tool and by text sample, which is exactly the problem: the Authors Guild's own ten-article test showed a wide spread of results across five detectors on identical human-written text, from near-zero flags to a full 100%. See our breakdown of what "accurate" even means for a detector for the studies behind that variance.

Why do detectors flag skilled or professional writers more often? Detectors flag text by measuring perplexity and burstiness, two proxies for how predictable word choice and sentence rhythm are. An experienced writer who spends years editing toward clarity ends up with the same low-predictability signature a language model produces on purpose, because that model trained on exactly this kind of polished prose. Different population, same underlying mechanism, and it's why non-native English writers hit the same well-documented false-positive gap.

What should I do if my writing gets falsely flagged? Get the specific tool name, version, and threshold that triggered the flag. Keep time-stamped drafts and version history on hand, since that's the process evidence that actually settles a dispute, and get a second opinion from a different detector before you accept the first score as final. We cover the fuller evidence-gathering process here.