What academic integrity boards look for in AI cases

If AI use in your thesis or coursework has been questioned, or you simply want to make sure it never is, it helps to understand how academic integrity processes tend to approach AI cases. Not the mythology, the mechanics: what typically triggers a case, what kinds of evidence get weighed, and what makes a student's account credible. Understanding the mechanics early is worth it even if you never need them.

One thing before anything else: this is general information, not legal or procedural advice. Integrity processes differ significantly between institutions and countries, and your university's own policy and procedures are the only authoritative description of your situation. If you are facing an actual case, read those documents first and use the support your institution provides.

How AI concerns typically arise

Most AI cases start with a human, not a detector. A marker notices a shift in register between assignments, or fluent well-structured prose that goes strangely generic wherever specifics are required, or terminology nobody in the field actually uses, or claims with no traceable origin in the module's material.

The most concrete common trigger is references: citations that do not resolve, quoted text that appears nowhere in the cited source, or bibliographies where real authors are attached to papers they never wrote. Detector scores and similarity reports often play a role too, but typically alongside these human observations rather than instead of them.

Timing shapes the process as well. Concerns about coursework are often handled through faster, lower-level procedures, while concerns about a thesis or dissertation may surface at examination, where both the stakes and the formality are higher. The same habits protect you in both settings, but the thesis case is where a defensible record earns its keep most visibly.

What integrity boards generally weigh

Procedures vary, but published policies and case guidance across many institutions point to a recognisable set of considerations:

• What the policy actually said: what was permitted, restricted, or banned for that specific assessment, and what students were told about it.
• Disclosure: whether AI use was declared, and whether the declared use matches what the work itself suggests.
• The nature of the assistance: language polishing tends to sit differently from generated analysis submitted as the student's own reasoning.
• Fabricated or unverifiable references, which are concrete and checkable in a way that style suspicions are not.
• Process evidence: drafts, version history, notes, search records, data files.
• The student's account, and whether they can explain the submitted work in their own words.

Why fabricated references carry so much weight

A style suspicion is arguable. A citation to a paper that does not exist is not. Fabricated references are the closest thing AI cases have to physical evidence: anyone can try to resolve a DOI or search a title, and the outcome is binary. This is why generated bibliographies are so dangerous. One unresolvable reference invites scrutiny of every other reference, and of everything else in the document.

The failure mode is inherent to ungrounded text generation, not a sign of a careless student in particular; the mechanics are explained in why AI makes up citations. But in an integrity process, a fabricated reference reads one of two ways: AI use combined with a failure to check, or fabrication by hand. Neither is a comfortable position to argue from.

Why AI detector scores rarely settle anything

Detectors occupy a strange position in these processes: widely used as screening signals, widely distrusted as proof. Vendors themselves generally describe outputs as probabilistic rather than definitive. Concerns have been raised publicly about false positives, including for writers whose first language is not English, and a number of institutions have publicly limited or reconsidered detector use in their processes.

None of this means a score gets ignored. It means that at many institutions a score alone is treated as a reason to look closer, not as a conclusion. The practical consequence for students cuts both ways: a low score is not protection, and a humanizer tool is not safety. What protects you is a documented process showing the work is yours.

Common misconceptions about AI misconduct cases

A few beliefs circulate among students that do not survive contact with how these processes actually run. The first: a detector score is proof, so a low score means safety and a high score means a finding. As above, many institutions treat scores as signals to investigate rather than findings, and cases commonly turn on other evidence entirely. None of these beliefs helps you, and two of them actively hurt.

The second: if the policy did not name my exact tool, I am covered. Policies are usually written at the level of conduct, presenting work that is not your own as though it were, rather than at the level of named products, and boards generally read them that way. The third: deleting drafts and history protects you. In practice the absence of any process record is itself conspicuous, and the student with nothing to show has surrendered their best evidence. The strong position is not an empty history. It is a full one.

What outcomes tend to look like

Outcomes vary as much as procedures do, and nothing here predicts any specific case, but published institutional policies typically describe a range rather than a single penalty: no finding, informal warnings and educational outcomes, mark penalties or resubmission for the affected assessment, and escalating sanctions for serious or repeated findings. Where a case lands generally tracks the factors already discussed: what the policy said, whether use was disclosed, the nature of the assistance, and the credibility of the student's account.

The pattern worth noticing is how much of that sits within a student's control in advance. You cannot control a marker's suspicion or a detector's output. You can control whether your use matched the written rules, whether it was disclosed, and whether a real process trail exists. That is where preparation belongs.

Process evidence: what a defensible record looks like

When a board asks whether work is genuinely the student's, the strongest answers are records that accumulate naturally during real work and are hard to fake after the fact. Notice that nothing below is produced for a board's benefit: ordinary working artefacts persuade precisely because they were not created to persuade. The record that helps is the one real work leaves behind:

• Dated drafts and version history showing the document growing and changing over weeks, not appearing at once.
• Reading records: annotated PDFs, notes tied to specific papers and specific passages.
• Search and screening records, especially for literature reviews.
• Data files, analysis scripts, or outputs that match the reported results.
• A disclosure statement written during the work, consistent with everything above.

The explain-your-work conversation

Many processes include some version of a viva: a conversation where you talk through the submitted work. The questions tend to be simple. Why this method and not that one. What does this term mean. Walk me through this argument. Where does this claim come from, and what does the source actually say.

A student who did the work, with whatever properly disclosed assistance, usually finds this conversation straightforward. A student defending text they never engaged with usually does not. This is worth internalising long before any case exists: every AI-assisted sentence you cannot explain is a liability, whatever your institution's detection posture happens to be.

If your programme allows it, rehearse this long before anyone requires it. Supervision meetings are a free version of the same exercise: talking through your argument, methods, and sources with someone who probes. Students who use supervision that way tend to find formal questioning unremarkable, because they have been explaining their work in their own words all along.

If you are asked to respond to a concern

Speaking generally, and not about any specific case: read the allegation and the policy it cites carefully, and note exactly what is being claimed, because responding to the wrong claim helps nobody. Note the deadlines and the format your response may take. Find the institutional support available to you: most universities have student advice services, unions, or ombuds staff who know these processes well. And gather your process evidence early, while it is easy to assemble.

Above all, be accurate. Overclaiming, such as insisting you never used AI at all when records suggest otherwise, tends to damage credibility more than the original concern did. Many institutions treat honesty during the process itself as significant, in both directions.

Keep perspective, too. An inquiry is a question, not a verdict, and integrity procedures exist partly to protect students from unfounded suspicion. The process is designed to be answerable, and students with genuine work behind their submissions answer it every year.

Building a defensible record from day one

The best time to think about an integrity board is when you will never meet one. Working in a system that documents your process as a side effect makes the entire question quiet. In CiteDash, papers live in your Library with sentence-level provenance, so claims trace to the passages behind them. The Fact Checker verifies cited claims against held full-text sources, and citations resolve to real papers by construction, so an unresolvable reference cannot enter your bibliography in the first place.

Before submission, the platform generates an AI-use disclosure reflecting how you actually worked, runs an originality pre-check, and applies submission readiness checks. For that final pass, see how to check your thesis before you submit.

Integrity is a record, not a vibe

Stripped of the anxiety, integrity processes for AI ask an old question in a new form: is this your work? The answer that convinces is never a detector score, and never an indignant assertion. It is a trail: sources read, drafts accumulated, assistance disclosed, claims that check out when someone follows them to the paper. Build the trail as you go, and the question largely answers itself.

That is also the honest reassurance in all of this. The habits that make a case survivable, reading your sources, keeping your drafts, disclosing your tools, checking your citations, are the same habits that make a thesis good. Nothing here asks you to work differently for a board's sake. It asks you to keep evidence of working the way you were supposed to work anyway.