No detector is 100 percent accurate on every text it sees, and any vendor who says otherwise is selling a billboard. TextSight scores text at the sentence level so you can see which lines triggered the flag, and we publish how often it wrongly flags human writing: 5.85% of 1,180 human-written academic papers. That is a result panel you can defend in a conversation, and a number you can check.
Accuracy in AI detection depends on what the model was trained on, where the signal is scored, and whether anyone measured how often it is wrong on real human writing.
TextSight runs a neural classifier trained on human and AI-written text, with lighter pattern checks blended in. It reads the overall fingerprint of machine-written prose rather than counting a few telltale words, which is why simple rewording rarely moves the score. The model is English-only.
Every sentence gets its own probability score. The result panel colours each sentence green, yellow, or red and lists the specific GPT-4 signals that triggered each flag (sentence-length floor, nested-clause density, the polite-assistant register, the "thoughtful synthesis" closer). A document-level percentage alone is brittle. A document percentage backed by sentence evidence holds up when a teacher, editor, or recruiter has to explain why the call was made.
We ran the detector over 1,180 human-written academic papers published in 2018, all of which predate ChatGPT, so every AI flag is a false positive by construction. It wrongly flagged 5.85% of them (95% confidence interval 4.65% to 7.34%), and a further 20.0% came back Mixed or Uncertain rather than a clean human verdict. The full method and per-document data are on our benchmark page.
Accuracy in this category depends on three variables: the source model, the length of the text, and whether the text has been edited or humanized. Here is where TextSight lands on each.
We do not publish an accuracy percentage, because accuracy moves with text length, genre and how much the text was edited, and formal, encyclopedic human writing is the kind most likely to be wrongly flagged. Unedited ChatGPT output of a few hundred words is where detection has the most signal to work with; short or heavily edited text is where it has the least.
If a passage went through several rounds of human edits, or through a dedicated humanizer, every detector loses signal, TextSight included. Weight the call accordingly and lean on the sentence-level evidence.
We ran the detector over 1,180 human-written academic papers published in 2018, all of which predate ChatGPT, so every AI flag is a false positive by construction. It wrongly flagged 5.85% of them (95% confidence interval 4.65% to 7.34%), and a further 20.0% came back Mixed or Uncertain rather than a clean human verdict. Second-language writers were flagged 4.6% of the time against 8.1% for native writers, a gap whose confidence interval crosses zero, so the honest statement is no significant difference in either direction.
That sentence will eat most competitor marketing, and it is still true. Treat any ChatGPT detector score as a conversation starter, not a verdict. The 0 to 100 Humanization Score and the sentence colours together give you enough evidence to decide whether to escalate a paper, push back on a freelancer, or move on. They do not give you the right to declare a piece "AI" with no further review.
False positives on second-language writers are the most consequential failure in this category, because they send real students into academic-integrity hearings they should not be in. Here is what we measured, and what we have not built.
Second-language writers often use shorter sentences, a more uniform rhythm and a narrower vocabulary, and a detector can mistake those features for machine writing. Liang et al. (2023) found seven detectors flagged 61.3% of TOEFL essays by non-native writers on average, against near zero for essays by native US eighth-graders.
Our training data includes second-language academic prose, and we checked whether our detector treats those writers differently. Second-language writers were flagged 4.6% of the time against 8.1% for native writers, a gap whose confidence interval crosses zero, so the honest statement is no significant difference in either direction. There is no separate calibration layer or per-language threshold in the product.
5.85% is not zero, for any writer. Never treat a TextSight score on its own as proof that a student used ChatGPT. Read the sentence-level evidence and talk to the writer first.
A free scan and a Business scan on the same text produce the same Humanization Score, the same sentence colours, and the same per-model confidence. Paid tiers raise the cap and unlock the humanizer, file upload, and API.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
Accuracy is only useful if the workflow around it is short. The standard TextSight scan takes about thirty seconds end to end on an 800-word essay.
Open app.textsight.ai. Paste up to 5,000 characters on the free tier, or up to 10,000 on Pro. The character counter ticks as you type or paste. No signup is required for your first scan.
Click Scan. The classifier runs in around six seconds for short text, thirty seconds for an 800-word essay. The result panel returns a 0 to 100 Humanization Score, a per-sentence colour map, and a list of the top three GPT-4 fingerprint signals found in the text.
Open the sentence-level panel. Every red sentence is annotated with the specific signals that triggered it: the sentence-length floor, the nested-clause syntax, the polite-assistant register, the synthesis closer, or the "intricate tapestry" vocabulary cluster. Yellow sentences are borderline and listed separately so a reviewer knows where the genuine uncertainty lies.
If three or more red sentences carry classic GPT-4 fingerprint signals on long-form text, the call is solid. If red sentences are isolated quotes or technical definitions, downgrade to "inconclusive". If the text is under 300 words or has been heavily edited, the result panel surfaces a low-confidence warning and you should treat the score as directional rather than precise.
No AI detector is accurate enough to be the sole basis of an academic-integrity ruling or a freelancer firing. These are the cases where the published numbers still leave too much room for a wrong call.
Every detector loses statistical power. Short paragraphs do not give the classifier enough sentence-level data points to score reliably. TextSight surfaces a "low confidence" warning when input length drops below the threshold the model was tuned for.
A passage that went through three rounds of human edits or a dedicated humanizer no longer carries the surface signals the classifier was trained on. Accuracy drops. The Humanization Score is still useful as a directional reading, but a single low score should not be treated as proof.
Text that is 80 percent ChatGPT and 20 percent human edits sits in a no-man's-land that no detector handles cleanly. TextSight reports a Humanization Score rather than forcing a binary classification, and the sentence colour map handles this case better than a document-level percentage. The borderline sentences themselves are still the hardest to call.
The most defensible workflow we have seen: scan with TextSight, read the sentence-level evidence, then have a conversation with the writer. A direct conversation produces more clarity than any percentage. If the conversation goes badly, the sentence evidence is the documentation you keep for the next step.
General-purpose detection across the full GPT family with the same sentence-level evidence panel.
Open the detector →Three scans a day, no signup, same classifier as paid.
Use it free →How TextSight reads GPT-4 output, and where detection gets harder.
Read GPT-4 page →Full tier breakdown for Free, Starter, Pro, and Business. Annual billing saves 25%.
See pricing →Free to try, no card, your first scan in about six seconds. Sentence-level evidence and a published false-positive benchmark.
Detect ChatGPT output wherever you work.