A single detector scan takes about four seconds and costs almost nothing. Acting on the result can cost a student a semester or an editor a contributor. That gap is the whole story. AI content detectors are accurate enough to sort a pile of documents by suspicion, and nowhere near accurate enough to prove authorship. Vendors advertise figures near 99%, while peer-reviewed testing shows those figures collapse on paraphrased text and on writing by non-native English speakers. Treat a score as a triage signal, never as evidence.
The Honest Accuracy Range
Under laboratory conditions, on unedited output from a model the detector was trained against, published accuracy is genuinely high. Grammarly states 99% detection accuracy and a first-place finish on the RAID independent benchmark. GPTZero advertises 99% accuracy across models such as ChatGPT, Claude and Gemini.
Now change one variable. Researchers showed that a recursive paraphrasing attack pushed detection rates down sharply across watermarking schemes, neural classifiers, zero-shot classifiers and retrieval-based detectors, while barely degrading the text (Sadasivan et al.). Nothing in the vendor number is false. It simply describes a narrow condition your documents rarely match.
The three numbers that matter more than accuracy
- False positive rate. How often human writing gets flagged. This is the number that hurts people.
- Sample length. Short passages carry little signal. The paraphrasing study worked with passages of roughly 300 tokens.
- Base rate. If only 5% of a batch is machine written, even a small error rate produces mostly wrong flags.
What a Detector Is Really Scoring
A detector is a statistical classifier, not a witness. It never sees who typed the document. It only sees the finished string of words and compares its shape against what generated text tended to look like in training.
Predictability and variation
Two properties do most of the work. The first is how predictable each word is given the words before it. The second is how much sentence length and structure vary across a passage. Generated prose tends to sit in a narrow band on both. Careful human prose sometimes sits there too, which is exactly where trouble starts.
Probability output, not verdict
Most tools return a percentage. That percentage describes the classifier’s confidence in a pattern match. It does not describe what share of the document a machine produced, and it does not describe how likely the writer is to be dishonest. Reading it as either is the most common misuse in the field.
Sample length changes everything
Signal accumulates across sentences. One short paragraph gives a classifier very little to work with, so its confidence swings wildly between runs. Longer passages average out individual quirks. This is why a scan of an abstract, a headline or a forum reply should be discarded outright, no matter how alarming the percentage looks on screen.
Why every model release resets the clock
Detectors learn from output that already exists. Each new generation of language models shifts the distribution of that output. A benchmark run last year against last year’s models does not automatically carry forward, and vendors rarely republish results per model version.
Vendor Claims Next to Independent Findings
Both sets of numbers are real. They answer different questions.
| Claim source | Headline figure | What it was measured on | How to use it |
|---|---|---|---|
| GPTZero | 99% accuracy, 17 million users, 1 million educators | Vendor benchmarking of its own detection model | Screening at volume |
| Grammarly | 99% accuracy, ranked first on RAID | Independent benchmark plus internal testing | Self-checking a draft |
| Turnitin, as reported by Vanderbilt University | 1% false positive rate at launch | Vendor claim, later withdrawn from use by the university | Understanding scale of harm |
| Liang et al. | Consistent misclassification of non-native English writing | Writing samples from native and non-native writers | Fairness review |
| Sadasivan et al. | Detection rates fall under recursive paraphrasing | Passages of roughly 300 tokens, multiple detector families | Stress testing |
The table compresses one lesson. Vendor accuracy answers “can this tool separate clean samples in a controlled set”. Research answers “does it hold up against edited, translated or adversarial text”. Buy on the first number, set policy on the second, and check pricing and current model coverage directly with each vendor before committing, because both change quickly.
Where Detection Breaks Down
Four failure modes account for nearly every disputed case.
Non-native writers pay the highest price
The clearest documented bias is linguistic. In a study of widely used GPT detectors, samples from non-native English writers were consistently misclassified as AI-generated while native samples were identified correctly (Liang et al.). The same paper found that simple prompting strategies both reduced the bias and let generated text slip past the detectors. The authors explicitly cautioned against use in evaluative or educational settings.
A 1% error rate is not small at scale
Vanderbilt University worked the arithmetic in public. Turnitin claimed roughly a 1% false positive rate. The university had submitted 75,000 papers in 2022, so an equivalent year would mean around 750 papers wrongly labelled as containing AI writing (Vanderbilt Brightspace). The university disabled the detector, citing that math, reported false accusations elsewhere, and the absence of any published explanation of how the tool reaches its determination.
Light rewriting is a reliable bypass
Paraphrasing does not need to be clever. The recursive paraphrasing experiments cut detection meaningfully with only slight quality loss (Sadasivan et al.). Anyone motivated to evade a scan will succeed. That inverts the incentive: the honest writer with plain prose is more exposed than the deliberate one.
Hybrid drafts sit in permanent grey zone
Most real documents are neither pure. A human outline expanded by a model, or a human draft tightened by one, produces mixed signal. Detectors were built for the clean case. When a document is genuinely collaborative, the number that comes back is closer to noise than measurement.
Field Cases, Lessons, and Reliability Practices
A university that stopped scanning
Vanderbilt’s decision is the most useful published case because it names its reasons: unexplained methodology, privacy exposure from routing student work to a third party, and the projected volume of false flags (Vanderbilt Brightspace). Their replacement was process based. Talk to students early, define permitted use, redesign assignments, and ask directly when something looks off.
A worked example from a content team
Picture a marketplace publisher reviewing 2,400 freelance submissions in a quarter. Assume the tool really does hold a 1% false positive rate, and assume 8% of submissions, 192 pieces, were actually machine written. On the 2,208 human pieces, roughly 22 get flagged wrongly. Suppose the detector catches 70% of the real cases after light editing, so 134 true flags. The reviewer now holds 156 flags of which about 14% are innocent writers. If the response is an automatic ban, one in seven punished writers did nothing. If the response is a request for drafts and a short call, the same list becomes a workable queue. Identical accuracy, opposite outcomes, decided entirely by policy.
The privacy question nobody scores
There is a second cost that never appears in an accuracy table. Scanning means sending someone else’s writing to a third-party service. Vanderbilt raised exactly this concern about routing student work into a system with separate data policies. Before adopting any tool at scale, read what it stores, how long it keeps text, and whether submissions can feed future training.
What experienced reviewers report
The pattern across published guidance is consistent. Scores are most trusted by people who have never audited one, and least trusted by people who have run known-human text through several tools and watched them disagree. Calibrate before you rely on any single number.
Practices That Improve Reliability
You cannot fix a classifier, but you can improve the decision built on top of it.
- Scan passages of several hundred words. Never judge a paragraph.
- Run two independent tools and compare. Disagreement is information.
- Record the tool, version and date with every score you keep.
- Set a written threshold before you look at results, not after.
- Treat mid-range scores as unreadable rather than mildly suspicious.
- Ask for the working file, revision history, notes or outline.
- Check whether cited sources exist. Fabricated references are stronger evidence than any percentage.
- Compare the document against the writer’s earlier work for tone and level.
- Publish your policy so people know scanning happens and how flags are handled.
- Give the writer a chance to explain before any consequence lands.
- Track your own false positives and review the threshold each term or quarter.
- Exclude quoted material, templates and boilerplate before scanning.
Common mistakes worth avoiding
- Reading a 78% score as “78% of this was written by AI”.
- Scanning a 90-word abstract and treating the result as meaningful.
- Applying one threshold to first-language and second-language writers alike.
- Keeping a flag on file without recording which tool produced it.
- Letting a score, rather than a person, trigger a penalty.
Product, Course, App and Platform Experience
Detection has quietly become a suite market rather than a single-button market. GPTZero now bundles its scanner with a plagiarism checker, a hallucination checker for fabricated academic citations, authorship verification, and integrations for Google Docs, Canvas and Chrome. Grammarly takes the opposite route and folds detection into the editor writers already have open, returning a share-of-text figure alongside rewrite suggestions. Copyleaks and Turnitin sell into institutions, where the buying criteria are bulk scanning, learning-management integration and audit trails rather than raw score quality.
The practical differences are workflow ones. An instructor grading 120 essays needs the score to appear where grading already happens. A freelancer needs a browser tab and nothing else. Free tiers, character limits and model coverage move often, so confirm the current terms on each vendor’s own page before you standardise a team on one tool. If you would rather understand the models producing this text than only the tools policing it, you can explore Coursiv AI lessons and see how generation and detection actually work.
Decision Framework: What to Know Before Acting on a Score
Run these six questions in order. Stop at the first one that fails.
- Is the sample long enough? Under a few hundred words, discard the result.
- What is the consequence? Higher stakes demand more evidence outside the tool.
- Do two tools agree? One flag is a hypothesis. Two is a reason to look closer.
- Is process evidence available? Revision history and drafts outrank any percentage.
- Who wrote it? Second-language and highly formulaic writing carries known false positive risk.
- What happens if you are wrong? Design the process backwards from that answer.
Next steps
Start with a calibration exercise this week. Take three documents whose origin you already know: one you wrote, one a colleague wrote, one you generated. Scan all three in two different tools and write down the spread. That one hour will tell you more about real accuracy than any published benchmark, and it gives you a defensible threshold instead of a borrowed one. Then write your policy down before the first disputed case arrives, because writing it afterwards always looks like a reaction.
Two neighbouring guides carry this further: how to tell if something was written by ai and does relying on ai hurt your skills.