A second-year engineering student in Pune submits a lab report. Three days later she is asked to explain why 64% of it was "likely AI-generated". She wrote it at her desk, from her own notes, in the careful English she was taught at a Marathi-medium school. Nobody in the room can tell her what the 64 means.
That scene, with small variations, now plays out in colleges, content agencies and HR departments across India every week. This post is about why it happens, who it happens to, and what both sides should do about it.
Lacewing Technologies is an AI and software development company in Navi Mumbai. We build client systems and we also run our own products, one of which is a text detector, so we see the mechanics from the inside and the complaints from the outside. What follows is the honest version of both.
What is a detector actually looking at when it flags you?#
Two numbers, mostly. The first is perplexity: how surprised a language model is by each word in your text, given the words before it. If the model would have guessed your next word anyway, perplexity is low. Machine-written text tends to have low perplexity because a machine chose those words for exactly that reason.
The second is burstiness: how much perplexity varies across the text. Human writers speed up and slow down. We write a long winding sentence and then a short one. We use an odd word in one paragraph and plain words in the next. Machine output is more even.
Newer detectors layer trained classifiers on top of these features, but the underlying question has not changed: does this text look like something a model would have produced? That question has an uncomfortable property. It is a question about style, and some humans write in exactly the style it is tuned to catch.
Stanford's Liang and colleagues put it plainly in their 2023 paper in Patterns: most detectors use perplexity, "which might inadvertently penalize non-native writers who use a more limited range of linguistic expressions."
Which kinds of human writing trip detectors most often?#
The pattern is consistent across vendors' own guidance and the independent research. Predictable prose gets flagged, and predictability has several ordinary causes.
| Trigger | Why it reads as machine-like | What to do about it |
|---|---|---|
| Writing in a second or third language | Smaller active vocabulary, safer sentence structures, textbook transitions | Keep drafts and notes; ask the assessor to compare with your earlier work, not with a native-speaker baseline |
| Short texts (under roughly 300 words) | Too little signal to separate style from chance; vendors raise minimums for this reason | Scan the whole document, never an excerpt; treat any confident score on a paragraph as noise |
| Essays written to a rubric | Five-paragraph structure, topic sentences, "firstly / secondly / in conclusion" | Assessors should expect higher baseline scores for templated assignments and adjust |
| Legal, compliance and policy text | Fixed phrasing, defined terms, repeated clauses | Exclude boilerplate before scanning; flag only the original analysis |
| Technical documentation and lab reports | Standard section headings, passive voice, formulaic method descriptions | Compare against the author's other reports; procedural sections should be read as procedural |
| Grammar-checked or translated text | Tools smooth rhythm and standardise word choice | Disclose which tools were used; keep the pre-edit version |
| Bibliographies, definitions, quoted standards | They are supposed to match a template exactly | Strip them before scanning or ignore their contribution to the score |
| Introductions and conclusions | Highly conventional openers and closers | Turnitin changed how it aggregates first and last sentences for exactly this reason; read those sections with that in mind |
Originality.ai's own help article lists most of these: human text "can sometimes produce a high AI score, particularly when the text is short, highly structured, heavily edited, translated, or written in a predictable style", and it names bibliographies, standard definitions and legal language as formats that resemble model output.
None of these are tricks people use to hide AI. They are what competent, careful writing looks like for a large share of the people who write in English in India.
What did the Stanford study actually find?#
The paper is Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers", published in Patterns in July 2023. The design was simple enough to repeat.
The authors took 91 TOEFL essays written by non-native English speakers, posted on a Chinese forum, and 88 essays by US eighth-graders from the Hewlett Foundation's ASAP dataset. They ran both sets through seven widely used detectors.
The US essays were classified correctly. The TOEFL essays were labelled AI-generated at an average false-positive rate of 61.3%. All seven detectors agreed on 19.8% of the human-written TOEFL essays, and at least one detector flagged 97.8% of them.
The essays flagged by every detector had significantly lower perplexity than the rest. That is the mechanism in the previous section showing up in the data.
Then came the part that should worry anyone running a detector on students. The authors asked ChatGPT to enrich the vocabulary of the TOEFL essays to mimic native-speaker word choice. The average false-positive rate fell from 61.3% to 11.6%, and only one essay was still flagged unanimously. Doing the reverse, simplifying the US essays to mimic non-native writing, pushed their false-positive rate up sharply.
Read that carefully. The human essays became "more human" to the detector after a machine rewrote them. The authors' conclusion was to "caution against their use in evaluative or educational settings, particularly when they may inadvertently penalize or exclude non-native English speakers."
For an Indian classroom where English is a second language for most of the room, that is not an abstract fairness concern. It is a prediction of who will be called in.
Why did OpenAI withdraw its own detector?#
Because it did not work well enough, and the company said so.
OpenAI launched its "AI classifier for indicating AI-written text" on 31 January 2023. The announcement was unusually candid. On the company's own challenge set of English texts, the classifier correctly identified only 26% of AI-written text as "likely AI-written", while labelling human-written text as AI-written 9% of the time. It added that reliability "typically improves as the length of the input text increases".
On 20 July 2023 the page was updated with a single line: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." The company said it was researching provenance techniques instead, which means marking content at the point of creation rather than guessing afterwards.
A 9% false-positive rate means roughly one in eleven honest writers is flagged. The organisation with the most knowledge of how these models generate text looked at that number and shut the product down. Every other vendor is working on the same problem with less information about the generators.
How do the major detectors present uncertainty?#
Since the 2023 studies, vendors have changed how they show results. The differences matter, because the presentation is what an assessor actually sees. Here is how five of them handle it, in no particular order of merit.
Turnitin shows a document-level percentage inside its Similarity Report. After testing 800,000 pre-ChatGPT academic papers, it found a higher incidence of false positives when it detected less than 20% AI writing, so it marks scores under 20% with an asterisk "to call attention to the fact that the score is less reliable". It also raised the minimum length from 150 to 300 words. For documents scored above 20%, it claims a document-level false-positive rate under 1%, and it adds that behind each one "is a real student who may have put real effort into their original work".
GPTZero classifies a document as human-only, mixed or AI-only, attaches a confidence category of uncertain, moderately confident or highly confident, and highlights sentences. It says it keeps its false-positive rate at no more than 1% and advises using results "as a conversation starter, and not as the final verdict". Its FAQ goes further: "these results should not be used to punish students."
Originality.ai returns a percentage and highlights sentences, and its help centre spends a whole article explaining that the number is "a probabilistic signal, not the percentage of words written by AI". A 90% result does not mean 90% of the document was machine-written. It means the model is fairly confident the text resembles machine output.
Copyleaks reports an AI percentage alongside its plagiarism score and, through a feature it calls AI Phrases, highlights the sentences driving the result. Its FAQ states that the chance of human text being falsely labelled is "less than 0.2%", and that detector output "should be used to inform the situation and offer the option for a learning opportunity".
TextSight, which we build, scores every sentence into one of three bands, AI-generated, mixed or edited, and human-written, and shows the count of flagged sentences next to an overall probability. Its guidance to educators is "Use any score as one signal, not proof. Talk to the student before you act." Free scans cover 1,500 words; paid plans start at $9.99 a month, ₹899 in India. It has the same structural limits as everything above, and we do not claim otherwise.
Two things stand out from that list. First, every vendor now publishes a caveat that amounts to "do not act on this alone". Second, every false-positive rate quoted is the vendor's own, from the vendor's own test set. Stanford measured 61.3% on a specific population with a specific set of tools. Both numbers can be true at the same time, and the gap between them is the whole problem.
What should you do if you have been wrongly flagged?#
Do not panic, and do not start by running your text through five more detectors. A second score is just a second opinion from a second machine with the same blind spots, and a low score from another tool proves nothing to a person who trusted the first one.
What persuades people is evidence of process. Collect it in this order.
1. Version history. Google Docs keeps it under File, then Version history. Word keeps it if the file lived in OneDrive or SharePoint. It shows the document growing over hours or days, with deletions and rewrites. Machine-pasted text appears in one large block at one timestamp. Originality.ai's own guidance tells assessors to look for "revisions, paste events, and contributors", and adds that "absence of history is not proof of AI use". Quote that line if you need to.
2. Earlier drafts and notes. Handwritten notes, an outline, a photo of a whiteboard, a WhatsApp message to a classmate about the structure, the PDF of the source you cited with highlights. Dated material from before the submission is strong.
3. Writing samples. Three or four earlier pieces in the same voice, ideally ones assessed before the detector was introduced. If your plain style is consistent across two years of work, the style is you. Originality.ai recommends reviewing "several substantial samples written by the same person" for this reason.
4. Tool disclosure. If you used Grammarly, a translator or a spell-checker, say so, and show the before and after if you have it. Hiding it looks worse than using it, and most policies permit it.
5. A conversation. Ask to walk through the argument in person. Explain why you ordered the sections the way you did, what you considered and dropped, where you were unsure. People who wrote something can do this. People who pasted it usually cannot.
6. The vendor's own words. Bring the vendor's published guidance to the meeting. Turnitin says it "does not make a determination of misconduct". GPTZero says results should not be used to punish students. Originality.ai says no detector "should be used as the sole basis for a disciplinary, employment, publishing, or academic decision". An institution that overrules its own vendor's instructions should have to explain why.
7. Put it in writing. Ask for the exact score, the tool and version used, the date of the scan, and the policy under which action is proposed. A written record protects you if the matter escalates.
One more thing. Do not rewrite the piece to lower the score and resubmit. It implies the original was the problem. Your original is your evidence.
How should an institution or employer build a fair process?#
The cost of a false positive does not fall on the institution. It falls on an individual, often one with the least power in the room. A fair process accepts that asymmetry and designs for it.
Publish the policy before the first scan. State which tool is used, what threshold prompts review, what a reviewer must do before anyone is contacted, and how the person can respond. Turnitin's first tip to educators is exactly this: "Know before you go", decide the process in advance and tell students what it is.
Set the floor at the vendor's own floor. If the vendor says scores under 20% are unreliable, do not review them. If it says 300 words is the minimum, do not scan shorter work. If it says an 800-word essay with a 30% score may contain 240 machine words or none at all, do not describe it as "30% AI".
Require a second signal before any contact. A score alone should trigger nothing but a closer read. Version history, a comparison with the person's earlier work, or an inconsistency in content should be present before anyone is asked to explain themselves.
Separate the detector from the decision. The person who reads the score should not be the person who decides the penalty. Reviewers should be told the base rate: on a plain-English, template-driven assignment, a share of honest submissions will score high.
Account for language background. Indian institutions assess students whose first language is Marathi, Hindi, Tamil, Bengali or any of a dozen others. The Stanford result is a direct warning that these students will be over-represented among flags. A fair process asks whether the writing matches the student's past writing, not whether it matches a native-speaker corpus.
Give the appeal real teeth. The person should be able to see the full result, including which sentences drove it, submit process evidence, and speak to someone who was not involved in the original flag. Set a time limit so an accusation does not hang over an exam term or a probation period.
Prefer process checks to post-hoc detection. Draft submissions, in-class writing, oral defences, and assignments built around personal data or local examples are harder to outsource to a model and do not require accusing anyone. GPTZero's own advice to educators is to "craft assignments that cannot be solved with AI" and to let students "demonstrate their understanding in a controlled environment".
Review the process, not just the cases. Count flags, count how many were upheld after review, and look at who was flagged. If the flagged group is disproportionately second-language writers, the tool is measuring the wrong thing for your population.
What does this mean for Indian students, agencies and employers?#
Three practical consequences.
For students, the honest advice is to keep drafts as a habit, write in a tool with version history, and know the institution's policy before submitting. Not because you have anything to hide, but because the burden of proof has quietly shifted onto you.
For content agencies and in-house teams, a client who runs copy through a detector before paying is now common. Agree in the contract which tool, which threshold, and what happens on a flag. A score is not a defect, and an invoice should not depend on one.
For employers screening written applications or assessments, remember that a detector tuned on native-speaker English will systematically disadvantage candidates whose writing is careful and plain. If you use one, use it to decide where to look, record the outcome of every review, and be ready to show that record to a candidate who asks.
The technology will improve, and so will the tools people use to write. The one thing a detector will never be able to see is who was sitting at the keyboard. Build every process around that fact, and the false positives become an inconvenience rather than an injustice.