Is Using AI Cheating? The Rule That Actually Decides

The decisive line is not whether a student touched an AI tool; it is whether the student misrepresented who did the assessed work. I have watched the phrase “is using AI cheating” collapse several different questions into one anxious headline. This article separates them. It explains when AI assistance is permitted, when disclosure is required, when citation is insufficient, how leading UK universities frame misconduct, how to use AI for brainstorming without plagiarism, what detectors can and cannot prove, and how to preserve evidence of an honest process. The immediate answer is no: using AI is not automatically cheating. It becomes cheating when the rules forbid the use, when generated material is submitted as original human work, when sources or data are fabricated, or when the tool performs the skill that the assessment is designed to measure.

The distinction matters because AI use is now ordinary rather than exceptional. The Higher Education Policy Institute’s 2026 survey of 1,054 full-time UK undergraduates found that 95% used AI in at least one way and 94% used generative AI to help with assessed work. Direct inclusion of AI-generated text reached 12%, up from 8% in 2025 and 3% in 2024. Those figures describe a rapidly changing study environment, not a verdict that almost every student cheats. The same tool can be a tutor in one task, an editor in another and an unauthorised ghostwriter in a third.

During this 2026 evaluation, I compared current university policies, citation guidance, peer-reviewed detector research, vendor documentation and student-use data. The resulting standard is practical: check the brief, identify the assessed skill, keep control of the reasoning, verify every claim, disclose material assistance and retain process evidence. That standard is stricter than “I cited the chatbot”, but more useful than a blanket ban.

Is Using AI Cheating? Apply the Four-Part Test

A reliable answer begins with four questions: Was the tool authorised? What did it actually do? Did the student remain the intellectual author? Was the assistance disclosed in the required form? A “no” on authorisation or authorship can make the use misconduct even when the output is factually correct. A failure to disclose can make an otherwise permitted use a transparency breach. Conversely, a clearly permitted use for practice questions, concept explanations or grammar checks is not transformed into cheating merely because the software is powerful.

When is using AI cheating?

Using AI is cheating when it crosses the assessment’s boundary of acceptable assistance. The most obvious case is submitting generated prose, code, analysis or images as personal work where independent production is required. It can also be cheating when a student feeds confidential exam material into a model, asks the system to fabricate references, paraphrases copied material to evade similarity checks, or uses an unauthorised agent during a closed-book task. The relevant wrong may be labelled plagiarism, fabrication, collusion, contract cheating or breach of assessment conditions, depending on the institution.

The reverse is equally important. An instructor may explicitly permit AI to generate counterarguments, test code, translate a draft, improve accessibility or provide feedback, provided the student checks and declares the assistance. In that situation the tool functions more like approved tutoring or software support. A useful companion is this ethical AI essay-writing framework, which separates idea support from authorship substitution. The governing principle is not “AI equals cheating”. It is “unapproved substitution plus misrepresentation equals cheating”.

“Assessment needs to change for a time of AI.” Professor Phillip Dawson, Deakin University, speaking at QAA’s Quality Insights Conference in 2025.

Dawson’s point redirects attention from policing a tool towards designing valid assessment. If a model can complete a task and the institution cannot tell whether the student learnt the target skill, the weakness is partly structural. That does not excuse dishonesty, but it explains why clear briefs, staged submissions and oral verification are more durable than vague warnings.

Situation	Cheating?	Why	Safer response
AI is forbidden, but used to draft the answer	Yes	It breaches explicit assessment conditions.	Do the task independently or request written permission.
Generated text is submitted as personal writing without disclosure	Usually yes	It misrepresents authorship and may be plagiarism or fabrication.	Rewrite from your own reasoning and declare permitted assistance.
AI is allowed for brainstorming and the use is acknowledged	No	The tool supports rather than replaces assessed authorship.	Keep prompts, notes and source checks.
AI checks grammar without changing meaning, where permitted	No	This is bounded editorial support.	Review every change and preserve the original draft.
AI invents a reference that the student includes	Yes	The submission contains fabricated evidence.	Open and verify every source before citing it.

Permission, Disclosure and Authorship Are Different Tests

Students often assume that citing an AI tool settles the ethical question. It does not. Citation addresses attribution. Permission addresses whether the tool could be used at all. Authorship addresses whether the submitted reasoning and expression are genuinely the student’s. These tests overlap, but none replaces the others. A student can accurately disclose a prohibited use and still have breached the brief. A student can use an approved tool yet conceal it, creating a disclosure problem. A student can also cite a chatbot while allowing it to perform the central intellectual task, leaving the assessment unable to measure the intended learning.

This produces one of the most useful insights for 2026 policy: permission is granular, not binary. A single assignment can contain green, amber and red activities. Green activities might include asking for definitions, generating practice questions or checking spelling. Amber activities might include suggesting an outline, translating a paragraph or critiquing a draft, all subject to disclosure and close review. Red activities might include producing the final argument, writing assessed code from the specification or generating empirical results. The same action can move between colours when the learning objective changes.

Consider two statistics students. One asks AI to explain why heteroscedasticity affects standard errors, then independently runs and interprets the model. The other uploads the dataset and asks for the complete analysis and discussion. The first use may strengthen learning; the second may replace it. In a coding module, autocomplete may be permitted for routine syntax while an agent that builds the entire application is forbidden. In language study, translation can be an accessibility aid in one module and the very skill being assessed in another.

The safest practical rule is to treat the assessment brief as a local licence. Read the exact wording, not a university-wide summary alone. Record what is permitted, what must be declared and which stages must remain unaided. Where wording is silent, ask a narrow question in writing: “May I use AI to generate outline options if I write and source the final argument myself?” A specific answer creates a usable boundary and avoids the false comfort of general approval.

What UK University AI Policies Commonly Require

Current UK policy converges on a simple hierarchy: the assessment-specific instruction controls; permitted use must be acknowledged; and unacknowledged generated content can constitute misconduct. Oxford’s policy, last updated in October 2025, requires assessment setters to declare whether and how AI may be used, to specify the expected declaration and to align permission with the purpose of each task. It states that breaching the stated specifications constitutes cheating and may constitute plagiarism. Notably, Oxford also said no AI detector had received university endorsement at the time of that update.

Cambridge permits appropriate AI use for personal study, research and formative work, but says that unacknowledged AI-generated content presented as a student’s own in summative assessment constitutes academic misconduct unless the brief explicitly states otherwise. Its humanities declaration template asks for the tool and version, the purpose, the prompt or interaction and how the output was changed. UCL similarly allows some planning, grammar checking and revision uses while making departmental and module instructions decisive. It identifies unacknowledged use as plagiarism, fake sources or data as fabrication, and instruction-breaking use as misconduct.

Institution	Permitted baseline	What triggers risk	Evidence expected
Oxford	Use depends on each discrete summative assessment.	Any use outside the stated specification.	Formal declaration in the setter’s prescribed format.
Cambridge	Appropriate personal study, research and formative use.	Unacknowledged generated content in summative work, unless expressly allowed.	Tool/version, purpose, prompts and account of changes.
UCL	Some grammar, planning and revision support, subject to module rules.	Unacknowledged use, fabricated material or breach of instructions.	Acknowledgement in appendix, methods or references as directed.

Students should therefore use a current, assessment-level checklist rather than rely on what a friend was allowed in another module. A broader review of student-focused AI tools can help identify capabilities, but capability never establishes permission. The institution’s policy, the module handbook and the task brief form the operative rule set.

“AI literacy and capability must be embedded across the curriculum.” Charlotte Armstrong, Policy Manager at HEPI, responding to the 2026 student survey.

The Learning Objective Decides Whether Assistance Is Legitimate

The most defensible boundary is the skill the assessment intends to measure. If the objective is to test recall under time pressure, an AI answer service defeats the task. If the objective is to evaluate and improve machine-generated material, refusing AI would defeat the task. If the objective is persuasive writing, a grammar checker may be peripheral while a model-generated thesis and argument are central. The closer the tool comes to performing the target skill, the stronger the case for restriction, disclosure or an alternative form of verification.

This approach is more precise than judging by output length. Ten generated words can be decisive if they contain the original hypothesis. Five hundred generated words may be legitimate if the task explicitly asks students to audit the model’s errors. Intent matters, but outcome and design matter more. A student may sincerely intend only to “get unstuck”, yet accept a generated structure that determines the whole argument. Another may use extensive AI output as raw material for a documented critique while retaining full analytical control.

A useful test is counterfactual: if the AI assistance disappeared, could the student still explain, defend and reproduce the core work? If not, the assistance probably reached the assessed competence. This is why oral follow-ups, code walkthroughs, research diaries and staged drafts are powerful. They test possession of the knowledge without pretending that a detector can recover authorship from prose alone.

“Validity matters more than cheating.” Professor Phillip Dawson, arguing that assessment must first measure what it claims to measure.

There is also an equity dimension. Oxford’s policy asks setters to ensure an equality of baseline provision where AI is authorised. That matters because paid plans can offer more messages, larger contexts, stronger reasoning modes and faster access. An assessment that silently rewards the best subscription risks measuring purchasing power. Institutions should either provide the required tool, design around free access or mark the student’s judgement rather than the model’s raw performance.

A Seven-Step Workflow for AI Brainstorming Without Plagiarism

Brainstorming is one of the most widely accepted uses of generative AI, but it can still drift into ghostwriting. The safest workflow keeps the model upstream of authorship and downstream of a clearly defined question. It also creates a record that a tutor can inspect. The following sequence is reproducible in ChatGPT, Claude, Gemini, Perplexity or a university-provided assistant, subject to the assignment rules.

1. Write your own problem statement first. State the topic, audience, assessment criteria and what you already think before opening the tool.
2. Ask for categories or questions, not finished prose. A prompt such as “List five tensions I should investigate” preserves more intellectual space than “Write my introduction”.
3. Challenge the output. Ask what assumptions it made, which counterarguments are strongest and what evidence would falsify each claim.
4. Move to primary sources. Treat model citations as leads only. Open the source, confirm authorship, date, context and page or section, then cite the original.
5. Close the AI window and build your own outline from notes. This break reduces unconscious copying of the model’s phrasing and sequence.
6. Draft in a versioned document. Keep timestamps, tracked changes and source notes so the development of the argument remains visible.
7. Write a declaration that names the tool, model if available, date, purpose, representative prompts, material retained and verification performed.

A good prompt log need not contain every conversational turn. It should preserve the interactions that materially influenced the work. For sensitive research, do not upload unpublished data, personal information, exam content or confidential client material unless the institution has approved the environment and its data terms. Consumer systems may offer account controls, but those controls do not replace research ethics, data-protection obligations or contractual confidentiality.

For literature work, compare the model’s suggestions with AI tools designed for researchers and conventional databases. The important step is provenance: can the student trace each claim to a source that was actually read? Brainstorming remains ethical when AI expands the question space but the student selects, verifies, structures and expresses the final argument.

How to Cite and Acknowledge AI Use Properly

Citation practice depends on the institution, discipline and type of output. A reference entry may be appropriate when a student quotes or closely paraphrases recoverable AI content. A methods statement or appendix is often better for describing a multi-turn process. Some tutors require both. The crucial point is that an AI citation should not be used as a substitute for citing the underlying journal article, statute, dataset or historical source. Models can misattribute, compress context or invent references, so the human author remains responsible for verification.

APA Style’s guidance treats generated output as software-generated material and recommends enough information for readers to identify the tool and interaction. MLA’s revised August 2025 guidance says not to treat the tool as an author. It recommends describing what was generated, naming the tool as the container, specifying the model or version, recording the publisher and date, and using a stable share link where available. For a practical publication workflow, this APA citation guide for AI offers a useful starting structure, but the local assessment instructions still take priority.

Use case	Minimum acknowledgement	What not to do
Brainstorming only	Tool, date, purpose and representative prompt; state that final wording and sources are yours.	Do not claim the tool had no influence if its categories shaped the argument.
Quoted or paraphrased output	In-text acknowledgement plus reference or shareable conversation, if policy permits.	Do not cite the AI instead of the primary source it summarised.
Editing or translation	Name the tool and scope of changes; preserve the pre-edit draft.	Do not let meaning, evidence or voice change unnoticed.
Code or data assistance	Record prompts, generated components, tests and human modifications.	Do not include untested code, fabricated data or confidential inputs.

A concise declaration might read: “I used [tool and model] on [date] to generate alternative research questions and critique my outline. I did not use generated prose in the submission. I verified all sources independently and retained the prompt log and version history.” Where generated language is retained, specify the sections and explain how it was edited. Honesty should be concrete rather than ceremonial.

AI Tools, Features, Pricing and Hidden Academic Constraints

Price does not determine whether AI use is ethical, but product limits shape what students can do and what evidence they can preserve. As of 17 June 2026, consumer plans change frequently, usage caps are often dynamic and education or enterprise contracts may differ by institution. The matrix below records publicly accessible plan information from official pricing pages and announcements. It should be treated as a dated snapshot, not a promise of permanent capacity.

Tool and plan	Current public price	Relevant features and integrations	Limits that matter academically
ChatGPT Free / Go / Plus / Pro	Free; Go US$8 monthly; Plus US$20 monthly; Pro US$200 monthly.	Web, mobile and desktop access; file uploads; image tools; web research; projects, tasks and custom GPTs on higher tiers; API billed separately.	Message, file, research and model access vary by tier and demand. Consumer subscription does not include API usage.
Claude Free / Pro / Max	Free; Pro US$20 monthly or US$17 monthly equivalent annually; Max from US$100 monthly.	Web, mobile and desktop; files and code execution; web search; Google Workspace and Slack; remote MCP connectors; Microsoft 365 and Outlook on eligible tiers.	Usage is variable and resets by policy. Higher context or tool use can consume allowance faster. API is separate.
Google AI Free / Plus / Pro	Free account; AI Plus US$9.99 monthly; AI Pro US$19.99 monthly on the UK-facing page.	Gemini access, storage bundles, Gmail and Docs integration on eligible plans, NotebookLM benefits and Google ecosystem connections.	Model limits, age rules and regional availability vary. Public pages do not guarantee fixed prompt caps.
Perplexity Free / Pro	Free; Pro advertised at US$17 monthly equivalent when billed annually. Other checkout prices may vary.	Cited web answers, file analysis, research modes, access to multiple models and separate API platform. Education deployments can connect approved content and drives.	Exact query allowances and institution pricing are not consistently public. Verify checkout and campus contract terms.

Pricing sources were the official OpenAI plan announcement, Claude pricing page, Google One plans page and Perplexity Pro page. These pages should be rechecked immediately before publication because regional prices, taxes, features and limits can change.

The technical difference from a traditional search engine is synthesis. An AI assistant can transform, infer, draft and execute multi-step instructions, not merely retrieve pages. That power creates new bottlenecks: hallucinated citations, context-window omissions, stale web results, prompt leakage, variable model behaviour and opaque changes between versions. It also creates reproducibility problems because the same prompt may produce a different answer later.

For source-led assignments, a useful comparison is Perplexity AI versus Google Scholar. Perplexity can accelerate discovery and summarise linked material, while Scholar and library databases remain better evidence layers for bibliographic tracing and peer-reviewed coverage. The defensible workflow uses AI for navigation and questioning, then verifies claims in the original publication. Exact pricing or hidden plan caps that cannot be confirmed should be labelled unavailable rather than guessed.

Can AI Detectors Identify Generated Content Reliably?

AI detectors estimate whether text resembles patterns associated with machine generation. They do not observe the writing process, identify the actual author or prove a policy breach. Their score can be affected by genre, length, editing, translation, formulaic academic style and later paraphrasing. A student who writes concise, predictable English may be flagged, while heavily edited generated text may pass. That asymmetry makes detector-only accusations unsafe.

Turnitin’s own guidance acknowledges false positives and suppresses exact scores in the 1% to 19% range, showing an asterisk instead. Its documentation says the AI writing score should not be the sole basis for adverse action. The system also requires qualifying prose, generally at least 300 words and no more than 30,000 words, with supported file and language conditions. These are not minor implementation details. A detector result outside its supported domain can look precise while resting on weak evidence.

Peer-reviewed research raises a fairness concern. Liang and colleagues reported that several detectors frequently misclassified writing by non-native English writers as AI-generated. Other studies have shown that paraphrasing can reduce detection and that performance varies sharply by model and text type. A current comparison of AI detector tools is useful for understanding product differences, while this guide to detecting AI-written content explains the linguistic signals people often inspect. Neither should be treated as forensic proof.

“AI literacy is akin to driver’s ed.” Victor Lee, Associate Professor at Stanford Graduate School of Education, describing responsible use education.

The stronger approach is triangulation. Review the assignment brief, compare drafts, inspect version history, ask the student to explain choices, test cited sources and consider whether the work is consistent with prior performance. A detector can trigger a conversation, but the evidence should come from process and context. Institutions also need an appeal route because false accusations carry academic and psychological costs.

Has ChatGPT Actually Increased Cheating?

The evidence does not support a simple claim that ChatGPT caused a universal surge in cheating. Stanford researchers Denise Pope and Victor Lee reported that, in their longstanding high-school survey work, the proportion of students admitting at least one cheating behaviour had remained around 60% to 70% for years and did not rise after ChatGPT’s release. That result is important, but it has limits: it concerns particular school populations and self-reported behaviour, not every university, country or form of misconduct.

What has clearly changed is the method and visibility of assistance. AI makes some forms of outsourcing faster, cheaper and available at any hour. It can produce plausible prose, code, images and references at scale. It also supports legitimate learning at scale. The 2026 HEPI survey found near-universal use among UK undergraduates, while only 12% reported directly including AI-generated text in assessed work. That figure has risen, but it is not equivalent to a measured cheating rate because some assessments explicitly permit or require generated material.

“The data suggest that AI is not increasing the frequency of cheating.” Victor Lee, summarising Stanford’s survey findings after ChatGPT’s launch.

The useful interpretation is that prevalence and misconduct are different variables. A calculator can be present in almost every student’s bag while being allowed in one exam and prohibited in another. AI is similar, although more capable and less transparent. Institutions should measure specific behaviours, not infer misconduct from tool adoption. They should also distinguish direct text insertion, idea generation, tutoring, editing, coding assistance and source discovery because each has a different relationship to learning.

Cheating is often a symptom of pressure, weak task design, inadequate support or low perceived relevance. Pope argues that students are less likely to cheat when they feel respected, connected and purposeful. That does not remove individual responsibility. It does suggest that clear expectations, reasonable workloads, accessible support and authentic assessment can reduce misuse more effectively than an arms race built around increasingly fragile detection.

Process Evidence Is Stronger Than a Detector Score

A student who uses AI responsibly should be able to show how the work developed. Process evidence does not mean surveillance of every keystroke. It means preserving enough artefacts to demonstrate authorship, verification and compliance. The most useful bundle contains the assessment instructions, an initial problem statement, source notes, representative prompts, exported conversations where permitted, dated outlines, draft history, tracked changes, code commits, test results and the final AI declaration.

A technical implementation workflow

1.  Create a project folder with subfolders for brief, sources, prompts, drafts, data or code, and submission.

2.  Save the assessment policy as a PDF or screenshot with the access date because web guidance can change.

3.  Export material AI conversations to a stable format. Record tool, model label, date, account type and whether web browsing or connectors were enabled.

4.  Use a reference manager for verified publications. Attach the downloaded paper or persistent identifier rather than copying an AI citation blindly.

5.  Write in a platform with version history. For code, commit meaningful stages and include tests that expose generated errors.

6.  Before submission, compare the declaration against the retained artefacts and remove confidential data from any shareable log.

Known bottlenecks include truncated exports, links that expire, changing model labels, institution-managed accounts that restrict sharing, and tools that do not expose exact model versions. Where exact version information is unavailable, record what the interface displayed and the date. For group work, identify which member used which tool and how the team reviewed the output. For API-based workflows, save the model identifier, system instructions, temperature or sampling settings where exposed, code version and request logs, while redacting secrets and personal data.

This approach also improves learning. A prompt log reveals where the student changed their mind; version history shows revision; test outputs show whether code was understood. It gives an educator evidence that can be discussed rather than a probability score that cannot explain itself. OpenAI’s educator guidance similarly suggests asking students to share or cite AI conversations as process evidence, but local privacy and assessment rules should determine implementation.

Edge Cases: Paraphrasing, Translation, Code and Accessibility

The hardest cases are not full essay generation. They are ordinary-looking transformations that can either support access or conceal authorship. Paraphrasing is legitimate when a student has understood a source, writes a faithful synthesis and cites it. It becomes misconduct when AI is used to disguise copied wording, evade a similarity report or preserve an argument the student did not develop. Ethical use of an AI paraphrasing tool therefore requires the same source citation and meaning check as manual paraphrase, plus disclosure where the policy requires it.

Translation raises a similar tension. A student may need language support to access research or express an idea, and an institution may permit it as reasonable support. In a language assessment, however, translation may perform the skill under examination. The safe boundary should be stated in advance, including whether the tool may translate source material, the student’s draft or only isolated vocabulary. The student should compare outputs against the original because models can flatten ambiguity, alter tone and introduce claims.

Code assistants can generate correct-looking software that fails at edge cases, imports insecure packages or reproduces licensed material. A permitted coding workflow should require unit tests, dependency review, explanation of generated components and disclosure of substantive assistance. The student must be able to trace data flow and defend design decisions. “It ran once” is not evidence of understanding.

Accessibility use deserves particular care. AI can simplify text, produce captions, convert formats, support dyslexia or help a disabled student organise work. Blanket bans can remove valuable access. Yet accommodations should not silently alter the competence being assessed. The fair solution is an agreed adjustment that preserves the learning outcome, gives comparable access and specifies what must be declared. The same principle applies to neurodivergent students using planning support and to international students using language feedback.

Finally, group assignments need collective rules. One member’s undisclosed AI use can expose the whole team. Agree permitted tools, logging responsibility, review standards and final sign-off before work begins. Shared accountability should be documented rather than assumed.

A Practical Decision Framework Before You Submit

A defensible decision can be made in less than five minutes if the student asks the right questions before using the tool, not after the draft is complete. Start with authorisation. Does the assessment brief explicitly allow, conditionally allow or prohibit the proposed use? If it is silent, consult the module’s current guidance and ask the tutor. Do not infer permission from a campus licence or from another assignment.

Next identify substitution. Is the model performing the exact knowledge, judgement, writing, coding, translation or creative work being assessed? If yes, stop unless the task explicitly asks for AI involvement. Then test provenance. Can every factual claim, quotation, dataset and reference be traced to an original source? If not, remove it or verify it. Test ownership by closing the tool and explaining the argument aloud. If you cannot defend a choice without rereading the generated answer, the work is not yet yours.

Finally, test transparency and security. Have you recorded material prompts and outputs? Does the declaration match what happened? Did you avoid uploading personal data, confidential research, copyrighted course packs or unreleased assessment content? Have you checked the plan and integration settings, particularly connected drives or workspaces, so the tool did not access more material than intended? These questions catch risks that a plagiarism detector cannot see.

The final answer to “is using AI cheating?”

No, not by itself. Using AI is legitimate when the rules permit the specific activity, the student remains responsible for the intellectual work, sources and outputs are verified, material assistance is declared, and the tool does not replace the competence being assessed. It is cheating when the use is forbidden, concealed, falsely attributed, fabricated or substitutive. Where uncertainty remains, the burden is not to guess what a detector might catch. It is to obtain a clear instruction and preserve an honest record.

This framework also helps educators. State permission at task level, provide examples of allowed and prohibited prompts, require proportionate process evidence, design meaningful checkpoints and avoid treating a detector as an adjudicator. The goal is not to make AI invisible. It is to make learning, authorship and accountability visible.

Takeaways

• Check the assessment-specific rule before opening an AI tool; university-wide guidance is only the starting point.
• Treat permission, authorship and disclosure as separate tests. Passing one does not guarantee compliance with the others.
• Use AI to generate questions, counterarguments and feedback, then close it and draft from verified notes.
• Cite original sources, not a chatbot summary, and remove any reference you cannot open and confirm.
• Preserve prompts, drafts, version history, source notes, tests and a concise declaration as process evidence.
• Never rely on an AI detector score as proof; combine context, drafts, explanation and source verification.
• Record plan limits, model labels and integrations when they affect reproducibility, privacy or equal access.
• Remember that citation cannot authorise a forbidden use, and disclosure cannot make substituted learning authentic.

Conclusion

The question “is using AI cheating?” has a clear answer only after the context is specified. AI is a tool, but it is not a neutral one: it can tutor, retrieve, transform, infer and generate finished work. That range is precisely why blanket approval and blanket prohibition both fail. The durable boundary is the relationship between the tool, the rules and the learning objective.

In 2026, universities are moving towards assessment-specific permission, explicit declarations and stronger evidence of process. That direction is sensible. It recognises that near-universal student use does not make every use dishonest, while refusing to treat disclosure as a magic cure for prohibited outsourcing. It also acknowledges that detector scores are too fragile and unequal to carry disciplinary decisions alone.

Open questions remain. Institutions still differ in terminology and enforcement. Paid tiers may create unequal access. Model updates weaken reproducibility, and accessibility uses require careful, individualised boundaries. Assessment design will continue to change as agents become more capable. Even so, the immediate standard is workable: obtain permission, keep human control of the assessed reasoning, verify sources and outputs, disclose material assistance and retain evidence of how the work developed. Used within those constraints, AI can support learning without impersonating it.

Frequently Asked Questions

Is using AI cheating in school or university?

Not automatically. It is cheating when the assignment forbids the use, when AI performs the assessed skill, when generated material is presented as personal work, or when assistance is not disclosed as required. The task brief and instructor’s written guidance control.

Can I use ChatGPT to brainstorm an essay?

Often yes, when brainstorming is permitted. Ask for questions, themes or counterarguments rather than a finished draft. Verify all suggestions, build your own outline, write in your own words, retain a prompt log and declare the use if your institution requires it.

Is it plagiarism if I rewrite AI-generated text?

It can still be misconduct. Rewriting does not solve unauthorised assistance or false authorship. It may also preserve fabricated claims or copied ideas. Permission, independent reasoning, source verification and disclosure matter more than whether the final wording has changed.

Do I have to cite ChatGPT or another AI tool?

Follow the assignment and citation style. Material generated content may require an in-text acknowledgement, reference entry, methods statement or appendix. Cite the original sources for factual claims. A citation does not make prohibited use acceptable.

Can plagiarism detectors detect AI writing?

They can estimate whether text resembles machine-generated patterns, but they cannot prove authorship or a policy breach. False positives, paraphrasing, language background and text length affect results. Detector scores should prompt review, not determine guilt.

Is using AI for grammar and spelling cheating?

Usually not when the policy permits bounded proofreading. It becomes risky if the tool rewrites substantial passages, changes meaning or supplies the argument. Preserve the original draft, review each change and disclose the tool where required.

Can I use AI-generated code in an assignment?

Only within the stated rules. Record the generated components and prompts, test the code, review security and licensing, and be ready to explain every design choice. If coding is the assessed competence, extensive generation may be prohibited even with citation.

What should I do if the AI policy is unclear?

Ask the instructor a specific written question describing the exact task you want AI to perform. Until permission is clear, use the tool only for general learning outside the assessed production process and retain the response for your records.

References

Higher Education Policy Institute. (2026, March 12). Student Generative Artificial Intelligence Survey 2026. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779

Modern Language Association. (2025, August 13). How do I cite generative AI in MLA style? (Updated and revised). https://style.mla.org/citing-generative-ai-updated-revised/

Quality Assurance Agency for Higher Education. (2025, February 26). Conference keynote addresses challenges of AI. https://www.qaa.ac.uk/news-events/news/conference-keynote-addresses-challenges-of-ai

Stanford Accelerator for Learning. (2023, October 31). What do AI chatbots really mean for students and cheating? https://acceleratelearning.stanford.edu/story/what-do-ai-chatbots-really-mean-for-students-and-cheating/

Turnitin. (2024). AI writing detection model. https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model

University College London. (2025, February 26). Can you use AI in exams and assessments? https://www.ucl.ac.uk/news/2025/feb/can-you-use-ai-exams-and-assessments

University of Cambridge. (2026). Artificial intelligence (AI). Education Quality and Policy Office. https://www.educationalpolicy.admin.cam.ac.uk/plagiarism-and-academic-misconduct/artificial-intelligence-ai

University of Oxford. (2025, October 27). AI use in summative assessment. https://governance.admin.ox.ac.uk/education-committee/policies/ai-use-in-summative-assessment