AI tools allowed in university work: which ones, and when?

Direct answer: No tool is allowed everywhere, and silence in a brief is not a neutral signal. Reading treats it as no AI, the University of Maryland as not allowed, Stanford's business school as permitted for take-home work, and Southampton tells you to ask. The rule belongs to the task and the university, not to the product.

Picture one hypothetical brief: a 2,000-word essay on the causes of inflation, with no sentence about AI. Four students open it on the same morning, and each lives by a different default.

At the University of Reading, the page on using generative AI in assessments sorts work into three categories and says that if a brief says nothing about GenAI, you should assume Category 1, which means independent work with no AI. The University of Maryland's guidelines tell students to assume AI is not allowed unless the syllabus says otherwise. Stanford's business school course policies say instructors may not ban AI for take-home MBA coursework. The University of Southampton's policy lists what is allowed and says "if in doubt, ask".

Source: MAAS summary of the four university pages above, read 2026-10-27.

Direct answer: Usually the course coordinator, inside a framework the university sets. The University of Helsinki allows AI as a rule but lets each course decide, and bans it in maturity tests. Surveys of top universities show that many publish guidance, but guidance mostly advises staff rather than listing approved products.

The University of Helsinki permits AI as a rule, lets course coordinators decide per course, forbids it in maturity tests, and says students cannot be forced to use AI services the university does not provide. That layering is typical: a general statement, then a course rule that wins.

Published guidance is also patchy. Moorhouse, Yeo and Wan (2023) searched the 50 top-ranked universities on 15 June 2023 and found 30 of them (60%) with a GenAI page. Only 23 pages met their criteria of addressing instructors and being university-level guidelines: 15 from the USA, 4 from the UK, 2 from Canada, 1 from Japan and 1 from Australia. Their abstract concludes: "The findings show that just under half of the institutions have developed publicly available guidelines."

A later review by Jin et al. (2025) examined the top 10 universities in each of 6 regions, 60 in total. Of these, 41 (68%) had published teaching and learning documents, and 40 were analysed. Only 22 of the 40 described how they communicate their policy to students, and the authors write that "gaps remain in comprehensive policy frameworks, particularly in addressing data privacy concerns and ensuring equitable access to GAI tools."

Direct answer: Not reliable enough to settle a case alone. Weber-Wulff and colleagues (2023) tested 14 detection tools on 54 documents, and the best reached 76% accuracy on the strictest (binary) scoring method. Perkins and colleagues (2024) found a mean accuracy of 39.5% on unedited AI text, so a score is a prompt for a conversation, not proof.

Weber-Wulff et al. (2023) built 54 test documents, 9 in each of 6 classes, and ran every one through 14 tools, 12 public detectors plus two commercial systems. On the strictest (binary) scoring method, the best tool classified 41 of 54 documents correctly, 76%, and the authors note: "Note that even the highest accuracy values are below 80%" (Weber-Wulff et al., 2023, p. 13). Class averages explain why. Human text was recognised 94% of the time, machine-translated human text 69%, manually edited AI text 30% and paraphrased AI text only 15%. The authors concluded that the tools "are neither accurate nor reliable and have a main bias towards classifying the output as human-written."

Perkins et al. (2024) ran 805 tests on 114 samples through detectors and recorded a mean accuracy of 39.5% on unmanipulated AI text and 67% on human-written controls. After simple evasion tricks, accuracy fell by 17.4% on average, to a mean of 22.14%. One detector flagged 50% of the human samples as AI.

Two lessons follow for you. A detector can miss real AI use, so universities rarely rely on it alone. It can also flag honest work, which is why keeping notes and drafts matters, as our guide on how accurate AI writing detectors are explains.

Direct answer: Detectors score text that looks predictable as machine-made, and many careful non-native writers use simpler vocabulary. Liang and colleagues (2023) found seven detectors flagged 61.22% of human TOEFL essays on average as AI-generated, so an international student can be wrongly flagged for work that is entirely their own.

Liang et al. (2023) tested 7 GPT detectors on 91 human-written essays from the TOEFL exam and 88 essays by US eighth-graders. They report: "However, they misclassified over half of the TOEFL essays as AI-generated (average false positive rate: 61.22%)." All seven detectors flagged 18 of the 91 essays (19.78%), and at least one flagged 89 of 91 (97.80%).

The mechanism is vocabulary. When the researchers asked ChatGPT to enhance the word choice of the TOEFL essays, the false positive rate fell from 61.22% to 11.77%, and only 1 essay in 91 (1.10%) was flagged by every detector. The reverse also held: simplifying the US eighth-grade essays raised their flag rate from 5.19% to 56.65%. A false positive here is a human essay judged to be AI.

The false-flag risk matters for Vietnamese students writing in English, because none of the university policy pages above mention detector accuracy or false accusation. If a flag arrives, our guide on what to do if you are accused of using AI lists the steps. Practical protection is dull but effective: keep version history, handwritten notes and source highlights until results are released, so you can show how the text developed.

Direct answer: Most do. The HEPI survey of 1,054 UK full-time undergraduates found 95% use AI in some way and 94% use it for assessed work. Only 12% say they include AI-generated text directly in assessed work, which is the line most policies draw.

Stephenson and Armstrong (2026) surveyed 1,054 full-time UK undergraduates through Savanta in December 2025, with a margin of error of about 3%. The headline was 95% using AI in at least one way, up from 92% in 2025 and 66% in 2024. For assessed work the figure was 94%, against 89% a year earlier. Generating text with tools such as ChatGPT was reported by 56%, down from 64%.

The share who include AI-generated text directly in an assessment rose from 3% in 2024 to 8% in 2025 and 12% in 2026. The report states: "There is a change in the proportion of students who include AI-generated text directly in a piece of assessed work, with 12% stating they have done so."

Students also feel the pressure of rules. About 65% say assessment has changed significantly because of AI, and 42% say fear of being accused of cheating makes them less likely to use it, down from 53% the year before. Another 21% are put off because their institution discourages or bans AI, down from 31%.

Using a tool for explanation, search terms or feedback is a different act from pasting its text into your answer. Policies such as Southampton's separate the two, and so should you.

Direct answer: Often not. HEPI found 37% of UK undergraduates agree and 36% disagree that their institution encourages AI use. At the Australian National University only 37% of surveyed students agreed they know what is expected, and 28% disagreed, rising to 40% among research students.

Habel (2025) reports the ANU students' association survey of 378 valid responses collected between 9 August and 10 September 2024. It was a convenience sample: 76% of respondents were domestic students, against 61% of ANU's population, so treat the figures as indicative and not as a national rate.

Within it, 37% agreed they know what ANU expects of them on GenAI, 30% were neutral and 28% disagreed. Disagreement was 40% among research students and 34% in humanities and social sciences, against 19% in STEM. Use itself was widespread: ChatGPT 71%, Copilot 15%, Gemini 13% and Claude 11%, and 86% of 362 respondents had concerns about GenAI accuracy. The executive summary says: "Students are not expecting a free pass to use GenAI for their assessments, but they are asking for clarity, equity, and guidance."

The HEPI numbers show the same gap in the UK, where 37% of undergraduates agree their institution encourages AI use and 36% disagree, and 48% feel staff help them build AI skills even though 68% call those skills essential. The practical conclusion is to treat a missing rule as a question to put to your lecturer, in writing, before you start. Our guide to referencing in university assignments helps with the citation side once you know what is allowed.

Direct answer: Follow your university's format, which usually asks for the tool, the purpose and a record of prompts. Helsinki goes further and asks students to disclose non-use as well as use. Write the statement while you work, because memory at the deadline is the weak point.

Policies differ on what you declare. Southampton says some modules require a declaration statement, and it counts citing GenAI use among allowed behaviour. The University of Helsinki requires students to disclose use or non-use of AI in written work, a rule the other pages in this post do not state. The distinction is useful: a line saying "no AI was used" is a claim you can later be asked to defend.

A good declaration answers three questions in two or three sentences: which tool and version, for which stage of the work, and what you did with the output. For example, you might state that a chatbot suggested search terms and that you then searched the library database yourself. Keep prompts and outputs in a folder until results are released, because some universities ask to see them.

Gonsalves (2025) studied why students failed to comply fully with AI-use declarations and argued that declaration rules alone do not ensure honest reporting. The fix is habit: update the declaration whenever you use a tool, and check it against your notes before submitting. Sorting your uses into "I wrote it" and "a tool wrote it" is also easier than judging a grey case at 11 p.m.

Direct answer: Start with your lecturer and the learning-skills team, who know what the brief allows. MAAS also offers paid one-to-one tutoring in which a subject expert coaches your study methods while the thinking and writing stay yours. Tutoring sets no grade target and makes no result promise.

Free support comes first. Most universities publish an AI guide for students, and learning advisers can show you how to turn an AI explanation into a properly sourced note. Bring the assessment brief to any appointment, because advice that ignores the specific rule for your task is of little use.

If you want regular coaching, MAAS one-to-one university subject tutoring pairs you with an expert in your subject for live sessions on the concepts and methods your course assesses. You do the work yourself: the tutor explains, questions and reviews your understanding. If the expert is not a good fit, you can ask to change to another one. For the wider topic, see Wikipedia on academic integrity and generative artificial intelligence.

• Accused of using AI: what to do: steps to take if a marker questions your work
• How accurate are AI writing detectors?: what the research says about detector errors
• When do you need referencing in university assignments?: citing sources, including material from AI tools

• Habel, E. (2025). ANUSA's generative AI survey report 2024. Australian National University Students' Association. https://anusa.com.au/wp-content/uploads/2025/06/2024_GenAI_Survey_Report_final.pdf
• Jin, Y., Yan, L., Echeverria, V., Gašević, D., & Martinez-Maldonado, R. (2025). Generative AI in higher education: A global perspective of institutional adoption policies and guidelines. Computers and Education: Artificial Intelligence, 8, Article 100348. https://doi.org/10.1016/j.caeai.2024.100348
• Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779
• Moorhouse, B. L., Yeo, M. A., & Wan, Y. (2023). Generative AI tools and assessment: Guidelines of the world's top-ranking universities. Computers and Education Open, 5, Article 100151. https://doi.org/10.1016/j.caeo.2023.100151
• Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). Simple techniques to bypass GenAI text detectors: Implications for inclusive education. International Journal of Educational Technology in Higher Education, 21, Article 53. https://doi.org/10.1186/s41239-024-00487-w
• Stephenson, R., & Armstrong, C. (2026). Student generative AI survey 2026 (HEPI Report 199). Higher Education Policy Institute. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/
• Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, Article 26. https://doi.org/10.1007/s40979-023-00146-z
• Gonsalves, C. (2025). Addressing student non-compliance in AI use declarations: Implications for academic integrity and assessment in higher education. Assessment & Evaluation in Higher Education, 50(4), 592–606. https://doi.org/10.1080/02602938.2024.2415654

This article is academic integrity guidance for students who use AI within their institution's rules. It does not replace your course policy, and MAAS does not produce assessed work on a student's behalf.