The Detection Plateau: Is AI Detection Accuracy Hitting a Ceiling?

There is an AI detection accuracy ceiling, and the field is close enough to feel it: against unedited text from familiar models detectors are excellent, but against fresh models, edited drafts, and paraphrase attacks, accuracy has stalled well short of the marketing — and theory says the gap can’t fully close. The interesting question isn’t whether a plateau exists. It’s where exactly it sits, why it moves, and what to do inside its limits.

Key takeaways

• Detection accuracy is really several numbers: near-ceiling on lazy copy-paste, mediocre on edited or paraphrased text, poor on brand-new models until retraining catches up.
• The theoretical argument is simple and hard to escape: where machine text statistically overlaps human text, no classifier can tell them apart — it can only pick its preferred error.
• Independent benchmarks consistently land far below vendor claims, because vendor numbers are measured on the easiest slice of the problem.
• The plateau’s real-world bite is asymmetric: false positives fall on people, and even a “tiny” false-positive rate flags real students at scale.
• The way past the ceiling isn’t a better classifier — it’s changing the question from “does this text look AI?” to “can this person show their work?”

Three accuracies, not one

“How accurate are AI detectors?” has no single answer because the problem splits into tiers of difficulty.

Tier one: unedited output from known models. Paste a ChatGPT essay straight into a modern detector and it will very likely get caught. This is the case vendors benchmark, and honestly, the case they’ve largely solved. Claims of 98-99% accuracy live here.

Tier two: edited, blended, and human-AI hybrid text. Accuracy sags. Every replaced sentence dilutes the machine signal; every human paragraph pulls the average score toward ambiguity. Detectors still contribute signal here, but confident verdicts evaporate.

Tier three: adversarial and out-of-distribution text. Paraphrased output, new models the detector hasn’t trained on, unusual genres. This is where benchmarks get ugly. The RAID evaluation (ACL 2024) — one of the broadest independent tests, spanning many generators, decoding settings, and adversarial edits — found detectors that look stellar in-domain degrading sharply under conditions no marketing page mentions. Krishna and colleagues showed back in 2023 that a good paraphraser alone gutted detector accuracy across the board.

The plateau claim is about tiers two and three. Tier-one accuracy climbed fast and is now scraping its natural maximum; the hard tiers have improved much more slowly, because they’re hard for structural reasons.

Why the AI detection accuracy ceiling exists

The clearest statement of the limit comes from Sadasivan and colleagues’ 2023 analysis, whose core idea doesn’t need math to grasp: a detector’s job is to separate two piles of text, and its best possible performance is capped by how much the piles overlap. If a passage could plausibly have come from either pile, no algorithm can sort it correctly — the information isn’t in the text.

And the overlap grows from both directions. Models are explicitly trained to write like humans; every generation narrows the statistical gap. Meanwhile a lot of human writing — formulaic, careful, second-language, heavily edited — has always sat near the machine region. Stanford’s Liang study documented the consequence in 2023: detectors flagged non-native English writers’ essays at rates that would be comic if they weren’t attached to misconduct hearings.

You can improve features, ensembles, and training data forever and still only approach the overlap boundary. That’s what a ceiling means: not “no progress,” but progress that buys less each year. OpenAI hit this wall publicly — its own classifier retired in July 2023 for low accuracy, and the company has been notably unwilling to ship a replacement.

The number the marketing hides

Accuracy is the wrong headline number anyway. The operationally decisive figure is the false-positive rate at the threshold actually deployed — and the base rates it meets in the wild.

Turnitin illustrates the whole dynamic. It launched claiming a document-level false-positive rate under 1%, then spent the following months adding caveats, an “asterisk” for low-score ranges, and warnings against acting on small percentages. Under 1% sounds like a rounding error until you multiply by millions of submissions: we’ve walked through how a tiny false-positive rate still flags real students, and the arithmetic is brutal. Push the threshold up to protect innocents and you miss more real AI text; pull it down and you manufacture false accusations. The ROC curve is a budget, and the plateau means the budget has stopped growing much.

That tradeoff — not raw accuracy — is where the ceiling actually bites institutions.

Doesn’t the arms race reset things?

Partly, and it cuts both ways. Each new model generation opens a temporary hole that retraining partially closes — the cycle we mapped in the arms race between generators and detectors. Detector vendors genuinely do claw back tier-one accuracy every cycle.

But notice what the cycle can’t do: it never recovers ground on the overlap itself. Retraining teaches a detector the *new* machine style; it doesn’t make edited hybrid text separable, and it doesn’t un-flag the humans whose writing naturally resembles machine output. The treadmill keeps tier one polished while tiers two and three — the tiers that decide real disputes — stay stubbornly hard. A plateau with periodic dips and recoveries is still a plateau.

Living inside the ceiling

If near-perfect detection isn’t coming, the sane responses follow directly.

For institutions: demote the score from evidence to prompt. A flag should start a conversation — about drafts, version history, and how the work was made — not end one. Process-based verification has no false-positive rate and improves teaching besides.

For platforms and publishers: use detectors for what tier-one accuracy is genuinely good at — triaging floods of low-effort content — and stop pretending a percentage can adjudicate individual authorship.

For writers: the plateau doesn’t mean scores stop mattering; gatekeepers still use them. It means scores are noisy in both directions, so check your own work before someone else does. Our detector is free to try on your own text, shows sentence-level detail rather than one opaque number, and if you’re checking at volume you can see what each plan handles. For the wider context, browse the rest of our writing on detection.

Frequently asked questions

Is AI detection accuracy actually plateauing? On the hardest cases, the evidence says yes. Detectors keep improving against last year’s models and lazy copy-paste use, but against fresh frontier output, edited text, and paraphrased text, independent benchmarks show performance far below vendor marketing — and theory work argues that as model text converges on human distributions, the best possible detector loses its edge. The easy accuracy has been won; what’s left is the expensive tail.

Why can’t detection just reach 99.9% with better technology? Because the errors that remain aren’t engineering slop — they’re overlap. Some human writing genuinely looks statistically machine-like, and some machine writing genuinely looks human. Where the two distributions overlap, no classifier, however clever, can separate them; it can only choose which kind of mistake to prefer. Better tech shrinks the overlap it can see, but new models keep widening it.

What do vendor accuracy claims like 98-99% actually mean? Usually: performance on the vendor’s own test set, often dominated by unedited output from well-known models — the easiest case. Independent evaluations across many generators and light adversarial edits routinely land far lower. And accuracy alone hides the number that matters most in schools: the false-positive rate on the specific kind of human writing being judged.

Does the plateau mean AI detectors are useless? No — it means they’re bounded. Detectors remain genuinely useful for triage: catching low-effort, unedited AI text at scale, flagging drafts for a human look, and giving writers a pre-submission check. What the plateau rules out is the courtroom use case: a single score as proof of authorship, which the error rates never supported anyway.

What should schools and platforms do if accuracy has topped out? Shift weight from detection to process. Version history, drafts, and conversations about the work verify authorship in ways a probability score can’t, and they don’t carry a false-accusation rate. Use detector output as a starting question, never a final answer — and calibrate any threshold against the real cost of a false positive.

The bottom line

AI detection climbed the easy slope fast and is now grinding against the hard one: statistical overlap that no classifier can cut through, refreshed by every new model release. That’s a ceiling — porous, shifting, but real. The mature response isn’t despair or denial; it’s using detectors for the triage they’re good at, refusing them the verdicts they were never accurate enough to deliver, and verifying authorship the way it has always actually been verified: by looking at how the work got made.