Last updated August 26, 2026

Academic Integrity

What is usually fine, what is not, and why an AI detector score is never enough to accuse someone.

Tool-by-tool guidanceHonest about detector limitsWritten for students and educators

In short

These tools help you write. They do not do your thinking, and they never move responsibility for the finished work away from you.
Your institution’s policy overrides everything on this page. Policies differ enormously between universities, and often between departments. Read yours.
No AI detector — ours included — is accurate enough to accuse someone. Never open a misconduct case on a detector score alone.
Non-native English writers are falsely flagged far more often than native speakers. That is a documented finding, not a rumour.
Using the Humanizer to hide AI authorship where you are required to disclose it is a breach of these rules, and we say so even though we make the tool.

1. Where we stand

We sell writing tools to a lot of students, and it would be easy to write this page as a disclaimer — a paragraph telling you to follow your university's rules, so that whatever happens next is your problem. That is what most of our competitors do. We would rather be useful.

Our position is simple. A writing tool that fixes your grammar, tightens your argument or helps you say in English what you already understand in your first language is doing what a good supervisor or a writing centre does, faster and at three in the morning. A tool that produces the analysis you were asked to produce, which you then submit as your own, has done your assignment for you. Both of those are available on this site. The difference between them is not which button you pressed — it is whether the thinking was yours.

Almost every academic integrity policy in the world is built around that same distinction, even when it is written in different words. If you keep it in mind, you will rarely go wrong.

2. Your institution’s rules come first

Nothing on this page is permission. We are not your university, we do not know what your course allows, and a page on a vendor's website has never once succeeded as a defence in a misconduct hearing.

Policies vary far more than students expect. Some institutions permit AI assistance freely with disclosure. Some permit it for drafting but not for final text. Some ban it entirely for assessed work. Some leave it to the individual module leader, which means the rules can differ between two assignments in the same term. A few still have no written policy at all, which is its own risk — an absent rule is not a permission.

Before you use any AI tool on assessed work

Read the assessment brief. Restrictions are often stated there rather than in the general policy.
Read your institution's academic integrity or academic misconduct policy, and search it for “artificial intelligence”, “generative” and “AI”.
If it is unclear, ask your tutor in writing before you submit, and keep the reply. A written answer from the person marking your work is worth more than any other assurance you can get.
If you cannot get an answer in time, take the cautious path and disclose what you used.

3. Tool by tool: what is usually fine

This table describes what is typically acceptable under mainstream academic policies. It is guidance, not a ruling on your case, and your institution can be stricter than any row below.

Tool	Usually fine	Usually a problem
Grammar Checker	Correcting spelling, punctuation and grammar in writing you produced. This is the closest thing to a universally accepted use — it is what a spellchecker has done for forty years.	Almost nothing, unless the assessment is specifically testing your unaided command of the language.
Paraphraser	Rewriting your own draft to read more clearly, or to cut length.	Rewording a source so that a similarity checker will not match it. This is plagiarism. Changing the words does not change whose idea it was, and the fact that the output looks original is precisely what makes it dishonest rather than clever.
Summarizer	Condensing a long paper so you can decide whether to read it properly, or checking you understood its argument.	Submitting a generated summary as your own reading, analysis or literature review.
Translator	Reading sources in another language. Drafting in your first language and translating — frequently allowed, and for many multilingual students the difference between participating and not.	Where the assessment is testing your ability to write in the target language. Language degrees and language-proficiency assessments almost always prohibit it.
Tone Changer	Adjusting register in your own writing — making an email more formal, a section less stiff.	Rarely an issue in academic work, though the same disclosure rules apply.
Email Writer	Administrative correspondence: emailing a supervisor, requesting an extension, contacting a library.	Not usually assessed work at all. Just be aware that a reader may notice.
AI Writer	Generating an outline to react against, producing counter-arguments to test your position, drafting sections you then rewrite substantially in your own words and with your own sources.	Generating text you submit as your own. Under most policies this is the clearest breach on the list, and it is the one most likely to be treated as serious misconduct rather than poor practice.
Humanizer	Improving the flow of writing that is already yours.	Disguising machine-generated text as your own. See the next section — it deserves more than a table cell.
AI Detector	Checking your own work for your own reassurance before submitting.	Using a score to accuse another person. See section 5.

4. The Humanizer, plainly

We make a Humanizer. Most tools like it are marketed on a single promise: that they will get AI-written text past a detector. We do not make that promise, and this section explains why in terms that are not marketing.

If your institution, employer, publisher or the law requires you to disclose that text was generated by AI, running it through the Humanizer to conceal that fact is a breach of our Terms of Service and of these guidelines. We say this even though it costs us customers who came looking for exactly that, because the alternative is selling a tool on a promise we would be ashamed to defend.

There are legitimate uses. Machine-generated prose has recognisable habits — it hedges, it repeats sentence shapes, it reaches for the same connectives, it is oddly reluctant to be brief. If you have written something yourself and it reads flatly, or you drafted with AI assistance that your course permits and you are editing the result into your own voice, smoothing that out is ordinary editing. That is the work the tool is built for.

What it is not built for is laundering. And there is a practical point worth knowing even if the ethics do not persuade you: it does not reliably work anyway. Detection and evasion move against each other continuously, any given detector may be updated between the day you submit and the day your work is reviewed, and institutions increasingly look at process evidence — drafts, version history, viva questions about your own argument — rather than a single score. A student who cannot discuss the reasoning in their own essay has a problem that no text transformation solves.

5. What AI detectors can and cannot tell you

We build an AI Detector, and it is on this site because people want one. It is also the tool we are most careful about, because detector output is routinely treated as evidence when it is nothing of the kind.

What a detector actually produces

A statistical estimate of how closely a piece of text resembles patterns typical of machine generation. That is all. It is not a record of how the text was made. There is no watermark to read and no metadata to check — the model is looking at word choice, sentence rhythm and predictability, and inferring backwards. Prose can be uniform and predictable for many reasons that have nothing to do with AI.

Two things worth knowing

OpenAI withdrew its own detector. In July 2023, roughly six months after launching it, OpenAI shut down its AI Text Classifier citing a low rate of accuracy — a company detecting output from models it had itself built, and concluding it could not do it well enough to keep the tool online.
Peer-reviewed research found systematic bias. Liang, Yuksekgonul, Mao, Wu and Zou, publishing in Patterns (Cell Press) in July 2023, found that GPT detectors consistently misclassified non-native English writing as AI-generated while classifying native-speaker writing correctly. The authors cautioned specifically against deploying such detectors in evaluative or educational settings.

How ours is built in response

Our detector reports a range rather than a single number, and shows sentence-level evidence rather than one verdict for the whole document. That is deliberate: a confident-looking percentage invites exactly the misuse this section is warning about. A range is harder to quote in a misconduct email, which is the point.

Our Terms of Service state this as a binding commitment, not a suggestion: AI Detector results are estimates, not proof, and must never be relied on alone for academic, employment, legal or disciplinary decisions about a person.

6. Who gets falsely accused

False positives are not distributed evenly. They fall hardest on people who are already at a disadvantage, which is what makes casual use of detectors an equity problem rather than a technical one.

Students writing in a second or third language. The Patterns study above is the clearest evidence, and the mechanism is intuitive: writers working in an additional language often use a more limited, more careful, more conventional range of vocabulary and structure — which is exactly what a detector reads as machine-like.
Autistic and other neurodivergent writers, whose prose may be more systematic, more literal or more consistently structured than a marker expects.
Anyone taught to write formulaically. Students drilled in a rigid essay structure, and disciplines with highly conventional academic registers, produce the regularity detectors are tuned to notice.
Careful editors. Heavy revision removes idiosyncrasy. Polishing a piece until every sentence is tight can make it look less human, which is a genuinely perverse incentive.
Technical and scientific writing, where precision and standard phrasing are required by the field.

If you belong to one of these groups, keeping evidence of your process — drafts, version history, notes — is worth the small effort. Section 9 covers what to do if it comes to that.

7. Disclosing and citing AI use

Where AI assistance is permitted, disclosure is usually what makes it permitted. It costs nothing and removes the entire question.

How to disclose

Follow the form your institution asks for. If it has not specified one, a short, specific statement in your methodology or acknowledgements is the accepted practice — naming the tool, what you used it for, and which parts of the work it touched. “I used an AI grammar checker to correct spelling and punctuation throughout” is a complete and honest disclosure. Vagueness is what causes problems: “AI was used in the preparation of this work” tells a marker nothing and invites the question you were trying to avoid.

Citing generated content

The major style guides — APA, MLA, Chicago and others — have all published guidance on referencing generative AI, and they update it. Check the current version of the one your department uses rather than relying on a summary, including this one.

Keep your working record

The most useful protection is also the least dramatic: keep your drafts. Version history in your word processor, dated files, reading notes, annotated sources and outlines together demonstrate a process that no generated text can retrospectively produce. If your work is ever questioned, this is the evidence that resolves it.

Our own Drafts library keeps versions in your browser rather than on our servers, so the history stays under your control.

8. For educators and institutions

We would rather be useful to you than adversarial, so this section is written straight.

What we ask

Do not open a case on a detector score alone — ours or anyone's. Treat a score as a prompt to look more closely, never as a finding. Section 5 sets out why, with sources.
Ask about process. Drafts, version history and a short conversation about the argument distinguish a student who did the work from one who did not, far more reliably than any classifier. They are also fairer, because they do not systematically disadvantage second-language writers.
Say what you allow. Most misuse we hear about comes from students who genuinely did not know where the line was. A specific sentence in the assessment brief prevents more misconduct than any detection tool.

What we will not do

We will not confirm whether a named individual holds an account, or whether a particular piece of text was produced using our tools. Two reasons, both firm. First, it would breach the Privacy Policy. Second, we could not answer honestly even if we wanted to: we do not store documents, so no record exists that could link a passage of text to a person. Any vendor who offers you that confirmation is either keeping far more of their users' writing than they admit, or guessing.

We do respond to valid legal process from authorities with jurisdiction, as described in the Privacy Policy.

9. If you have been wrongly accused

This happens, and it is frightening. Some practical steps.

Gather your process evidence first. Version history, dated drafts, notes, reading, search history, outlines. Do this before you reply, and do not edit the original files.
Ask what the evidence actually is. If it is a detector score, ask which tool, what score, and what the tool's own documented false-positive rate is. Ask whether the institution has validated it on writing by students in your language group.
Point to the research. The Patterns paper cited in section 5 is peer-reviewed, widely known, and directly on point if English is not your first language.
Offer to discuss the work. A student who can explain their argument, their sources and the choices they made is providing the strongest evidence available.
Use your institution's process. You are almost always entitled to representation from a students' union or equivalent, and to appeal. Ask what the procedure is and follow it.
Be honest about what you did use. If you used a spellchecker or a translator, say so plainly. Being caught understating it is far more damaging than the use itself.

We cannot intervene in your institution's process, give you legal advice, or certify that a piece of work was not machine-generated — nobody can honestly do that last one. What we can do is point you to the sources above, which are public and citable.

10. Help and contact

Questions about this page, or about how a tool works: info@aiwritingassistant.net. We cannot advise on whether a specific use is permitted at your institution — only your institution can answer that.

Sources cited on this page

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). doi.org/10.1016/j.patter.2023.100779
OpenAI (2023). New AI classifier for indicating AI-written text — updated to record the tool's withdrawal for low accuracy. openai.com