How Much AI Assistance Counts as Your Own Work

# How Much AI Assistance Counts as Your Own Work

Most students who ask this question want a number. A percentage of words, a list of permitted tools, a line they can see before they cross it. There is no such number. Hunting for one makes the problem harder.

A better question: not whether you used AI, but what you did with it. The distinction that matters is between using a tool and letting the tool do the intellectual work. That distinction is a way of thinking about your own draft, not a finding from research, and this article presents it as such. [S6] and [S2] are relevant here, but neither frames the issue as a question of authorship. Treat the framing below as a working device, not a rule you can cite.

This piece covers three things: how to read your own institution's rule, how to separate assistance from authorship, and how to disclose either way. It cannot tell you what your university's policy says. No Ugandan institutional policy, faculty guideline, or national regulator guidance was supplied when this article was written, so the rules at Makerere University, Kyambogo University, Gulu University, Mbarara University of Science and Technology, Uganda Christian University, and other institutions are not described here. Where that gap matters, it is stated plainly.

Start with your own institution's rule, not with this article

Your programme handbook, faculty policy, or examination regulations are the authority on what AI use is permitted in your work. Nothing here overrides them.

Where the rule is likely to live. Check, in this order: your programme or student handbook; any faculty-level academic integrity or AI policy; your dissertation or research module guide; and the general examination regulations. This is a suggestion about where to look, not a claim about where Ugandan universities publish such rules. Some institutions post policies on a faculty noticeboard or circulate them by email rather than on a website. A rule that is not online may still exist. If you cannot find one in any of these places, that is itself information: you need to ask.

Read it for four specific things. Look for permitted uses (what the policy explicitly allows), prohibited uses (what it explicitly forbids), disclosure requirements (whether you must declare AI use, and in what form), and the stated consequence of a breach. Write these four points down in your own words. If any is missing, you have found a gap to resolve before you submit, not after.

When the rule is silent or vague. Treat silence as a reason to ask your supervisor in writing before you proceed, and keep the written answer. An email saying "for your literature review, using a tool to check grammar is fine, but not to draft the chapter" is worth more than any general guidance, because it is specific to your work and it is on record. Do not assume that because a policy does not mention a use, the use is permitted.

Practical constraints you may face. Not every student has the same access. Data bundles, electricity, and reliable internet are uneven, and a tool that works well on campus may be unusable at home. If your access is limited, say so in your disclosure rather than letting the reader assume you had none or plenty. The cost of paid AI subscriptions in Ugandan shillings, and how many students can afford them, is not known from any supplied source; this article does not estimate it. For integrity, what matters is the same either way: what you did with the tool, not what you paid for it.

Assistance versus authorship: a working distinction

The difference that matters in practice is between support for a judgement you made and a judgement that originated somewhere else. No supplied source defines this distinction for academic work, so the definitions below are this article's own working device. Use them to interrogate your draft, not as a standard you can attribute to anyone.

Assistance: you decided, AI helped you execute. You chose the argument, the structure, and the evidence. The tool checked your grammar, reformatted your references to APA 7, or suggested that two sentences might read better in the other order. You looked at each suggestion and decided whether to accept it. The intellectual choice stayed with you.

Authorship: the claim, structure, or analysis originated with the tool. You asked a tool to draft your literature review and presented the result as your reading of the field. You asked it to propose research questions and adopted them without your own reasoning. You fed it your results and presented its interpretation as your analysis. In each case the judgement did not originate with you, even if you edited the wording afterwards.

The defend-it test. A short check you can apply to any sentence, choice, or number in your draft: can you explain, without the tool in front of you, why it is there and what alternative you rejected? If you can, the work is defensible as yours. If you cannot, it is not, whatever the written policy says. The underlying idea — that effective use of AI output depends on the user's own judgement and oversight — is consistent with findings on how people respond to AI assistance in a data-analysis setting [S2]. The test itself is this article's device, not a proposal from that study.

Why more assistance is not automatically safer

It is tempting to assume that a little AI is safe and a lot is risky, with a smooth line in between. The available evidence suggests the relationship is not that simple, and that heavier assistance carries a cost that shows up in what you can do afterwards.

What the research actually found. In a study of AI-supported note-taking, more AI assistance reduced cognitive engagement [S6]. A 2026 study models a pathway from AI assistance toward AI dependence [S1]. Both are findings about patterns across participants or modelled trajectories. Neither predicts what will happen to any individual student, and neither studied Ugandan students or dissertation writing specifically. Read them as reasons to watch your own process, not as a verdict on your draft.

The risk is a chapter you cannot defend. The practical hazard is not only the written policy. It is that a student who cannot reconstruct their own reasoning is exposed in supervision meetings and oral examinations, where the questions come without warning and the tool is not available. This is reasoning, not a research finding. No supplied source examined viva or supervision practice. But it follows from the point above: if heavier assistance reduces engagement with the material, the gap will show wherever you are asked to explain your work aloud.

Assistance quality depends on the human in the loop. How well people use AI output varies with their own oversight and judgement [S2]. The same tool, used on the same task, can be assistance for one student and substitution for another, depending on how much the student checks, rejects, and understands. That extension is this article's inference. [S2] studied data analysts, not students, and it supports the general point about variation in user response, not the specific claim about substitution.

Where the line usually falls: four common tasks

These are the tasks students most often ask about. For each, the assistance/authorship distinction gives you a way to think it through. The final answer for your own work depends on your institution's rule, which this article cannot state. No supplied source examined these tasks in an academic-writing setting, so the guidance is conditional advice.

Reading and summarising sources. Using an AI summary as a check on your own reading is assistance: you read the source, formed your own view, and used the summary to test whether you missed something. Letting the summary replace the reading is an authorship risk, because the claims then enter your literature review without you having verified them against the source. If you cannot say what the original argued in your own words, you have not read it. This matters especially when the source sits behind a paywall or in a library database you can only reach on campus. See How to Find Credible Academic Sources and How to Write a Literature Review Chapter.

Drafting and editing prose. Editing your own sentences, with a tool suggesting tighter phrasing you then accept or reject, is assistance. Generating sentences and presenting them as your own writing is not, even if you change some words afterwards. What counts is where the sentence came from, not how much you altered it. See How to Paraphrase Without Plagiarising and Clearer Academic Writing Without Losing Rigour.

Analysis and interpretation. Here the authorship question is sharpest, because the interpretation is the contribution of an independent study. This article's assessment, not a research finding, is that interpretation carries the highest risk. AI tools can produce plausible output whose reasoning the user cannot inspect [S5]. That source is about alignment and AI behaviour, not about student analysis, so cite it narrowly: fluent output is not the same as output you can check. If you cannot explain why a result means what you say it means, the interpretation is not yet yours. See Interpreting a p Value Without Overclaiming.

References, citations, and formatting. Formatting references is low-risk assistance: the intellectual work of choosing and reading sources is already done, and the tool is applying a style. But every citation must be verified against the actual source before it goes in your reference list. No supplied source examined AI-generated citation accuracy, so this article makes no claim about error rates. Verification is simply your responsibility. See How to Verify a Source Before Citing It and APA 7 Referencing Explained.

The attribution problem: why "the AI wrote it" is not a complete answer

If a piece of work was produced with AI, saying so sounds like full disclosure. It is not, because attributing AI output to a specific tool, prompt, or person is genuinely difficult. That difficulty is exactly why institutions ask for disclosure rather than leaving the question to inference after the fact.

Outputs are not cleanly traceable. Work on AI agents argues that counting and attributing them raises real individuation and liability problems [S3]. The same paper appears in two venues [S4]; treat these as one source, not two independent findings. This is a conceptual and legal argument about AI agents in general. It is not a study of student work, and it says nothing about plagiarism policy. What it supports is the narrow point that working out which output came from which tool, and who is responsible for it, is harder than it looks.

What this means for you in practice. If attribution is difficult in general, it will be difficult for you after the fact. Reconstructing your process at submission time is unreliable: you will not remember which paragraph you drafted and which the tool did, or which version of a prompt produced which sentence. The record you keep as you work, and the disclosure you write from it, are what protect you. That is an inference from the point above, not a sourced conclusion.

Disclose either way: what a short disclosure statement looks like

Disclosure resolves most of the uncertainty, and it is worth doing whether or not you think your use was substantial. The template below is a suggestion, not a required format. If your institution specifies a format or a location, follow that instead.

A four-line template. Four lines cover most cases:

1. Tool and version. Name the tool and the version or model, as specifically as you can.
2. What you used it for. For example: "checked grammar and sentence clarity in Chapters 2 and 3"; "reformatted the reference list to APA 7"; "suggested alternative phrasings that I accepted or rejected."
3. What you checked or changed yourself. For example: "I read every suggestion and accepted only those that preserved my meaning; all citations were verified against the original sources."
4. What you did not use it for. For example: "I did not use AI to generate text, choose my research questions, or interpret my results."

The fourth line matters as much as the others. It tells the reader where the boundary was, which is the question they are actually asking.

Where it goes. Students commonly place a disclosure in a methods note, an acknowledgements section, an appendix, or a declaration page. No supplied source gives examples of institutional placement rules, so this is a list of common locations, not a recommendation from any authority. Your institution's stated placement takes precedence. If it states none, ask your supervisor where they would like it.

Keep a simple log as you work. A dated record of prompts, outputs, and your own edits, kept from the first day rather than reconstructed at submission, is the single most useful habit in this article. It can be a notebook, a document, or a folder of screenshots. No institution is known to require a log on the basis of any supplied source, so do not treat it as a formal requirement. Treat it as the evidence base for your disclosure statement and for your own memory.

If you are unsure, ask before you submit

The sequence is short. Read your institution's rule and note its four elements: permitted uses, prohibited uses, disclosure requirements, and consequences. Apply the defend-it test to your draft. Write a disclosure statement whether or not you think you need one. And when the rule is silent or the answer is unclear, ask your supervisor in writing and keep the reply.

None of this is a procedure recommended by any authority. It is this article's synthesis of its own advice, offered because the alternative — guessing, then discovering the problem at examination — is worse. For the wider picture on accuracy, citations, originality, and integrity, see Using AI in Academic Research Responsibly. For the writing skills that make assistance less necessary in the first place, see How to Paraphrase Without Plagiarising.

The line between assistance and authorship is not drawn by a percentage. It is drawn by whether you can defend the work as yours, and by whether you have told your institution honestly what you did.

Sources used

• [S1] Tariq, U., Malik, A., & Farooq, H. (2026). From AI Assistance to AI Dependence: Predicting Students' Reliance on Generative AI for Academic Work. ACADEMIA Nextgen Management Journal. doi:10.63056/anmj.2.3.2026.2481
• [S2] Gu, K., Grunde-McLaughlin, M., McNutt, A. M., Heer, J., et al. (2024). How Do Data Analysts Respond to AI Assistance? A Wizard-of-Oz Study. CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3641891
• [S3] Arbel, Y., Salib, P., & Goldstein, S. (2026). How to Count AIs: Individuation and Liability for AI Agents. arXiv (Cornell University). http://arxiv.org/abs/2603.10028
• [S4] Arbel, Y. A., Goldstein, S., & Salib, P. (2026). How to Count AIs: Individuation and Liability for AI Agents. SSRN Electronic Journal. doi:10.2139/ssrn.6273198
• [S5] Buyl, M., Khalaf, H., Verdun, C. M., Paes, L. M., et al. (2025). AI Alignment at Your Discretion. ACM Conference on Fairness, Accountability, and Transparency (FAccT). doi:10.1145/3715275.3732194
• [S6] Chen, X., Ruan, K., Ju, K. P., Yap, N., et al. (2025). More AI Assistance Reduces Cognitive Engagement: Examining the AI Assistance Dilemma in AI-Supported Note-Taking. Proceedings of the ACM on Human-Computer Interaction. doi:10.1145/3757632

Related reading

• How to Paraphrase Without Plagiarising: A Checklist
• How to Verify a Source Before Citing It: A Checklist