AI and academic malpractice: how to spot and prevent AI written work

Every assessor has had that moment: a piece of coursework lands that reads far too smoothly for the learner who wrote it, or uses vocabulary and structure that does not match anything they have seen from that person before. The question of what to do next is one FE providers are still working out, and it is worth being honest that there is no perfect answer yet.

Why this is harder than plagiarism used to be

Traditional plagiarism left a paper trail, a matching source you could point to. AI-generated text does not copy from a specific source in the same way, so text-matching software often finds nothing even when a piece of work was clearly not written unaided. That has forced providers to rethink what evidence of malpractice actually looks like.

Signs worth investigating, not proving

None of the following prove AI use on their own, but together they justify a conversation with the learner:

• A sudden jump in vocabulary, sentence structure or argument sophistication compared with previous work
• Generic phrasing that could apply to almost any topic, with no specific detail tied to the actual task or context given
• Perfectly balanced paragraphs that hedge every point without committing to a view
• References or examples that do not check out, including citations that do not exist
• Formatting quirks common in AI output, such as repeated three-point lists or a habit of starting sections with "In conclusion" or "It is important to note"
• A learner who cannot explain or expand on their own submitted work when asked

Why AI detection tools are not the answer

It is tempting to run suspect work through a detection tool and treat the score as proof. Do not do this. Detection tools produce false positives against learners who write in a formal, structured style, including many ESOL learners and neurodivergent learners, and false negatives against AI text that has been lightly edited. No awarding body accepts a detector score alone as evidence of malpractice, and using one as your sole method leaves your decision open to appeal. This is covered in more depth in Is AI marking safe? Assessment integrity in the age of generative AI.

What to use instead

Build verification into the assessment process rather than trying to catch problems after the fact.

Drafting history and version control

Ask learners to submit work through a platform that logs edit history, or to keep and submit earlier drafts alongside the final piece. Genuine drafting shows revision, false starts and development over time. A single, polished document with no history is not proof of malpractice, but it removes one of the easiest ways a learner has to demonstrate authenticity.

Professional discussion

A short conversation where the learner talks through their own work, in their own words, is one of the most reliable ways to confirm understanding. If they cannot explain a term they used or a decision they made, that tells you something a detector cannot.

Staged and in-class submissions

Breaking a larger piece of work into stages, with some elements completed under supervision, limits how much of the final grade rests on unsupervised, unverifiable work.

In-person verification

For vocational and practical qualifications, observed practice and oral questioning remain harder to fake than written coursework, which is one reason many awarding bodies are leaning further into these methods.

Following JCQ and awarding body guidance

JCQ guidance on the use of AI treats undeclared AI-generated work submitted as a learner's own as malpractice, in the same category as plagiarism or collusion. Awarding organisations expect providers to:

• Have a clear, published policy on acceptable AI use, referenced in Writing an AI policy for your college or training provider
• Make learners declare where and how they used AI tools in producing a piece of work
• Investigate suspected malpractice through a fair, documented process rather than an informal accusation
• Apply consistent penalties in line with the relevant awarding body's malpractice policy

Running a fair investigation

1. Do not accuse. Raise the concern factually: "I noticed this section reads differently to your other work, can you talk me through how you approached it."
2. Gather evidence beyond a detector score: drafting history, previous work samples, the learner's own account.
3. Give the learner a chance to respond before any decision is recorded.
4. Record the outcome and reasoning, whether or not malpractice is confirmed, in case of appeal.
5. Apply the penalty set out in your provider's malpractice policy, not an ad hoc one.

Prevention beats detection

The providers managing this best are not the ones with the most sophisticated detection software, they are the ones who have redesigned assessment tasks to make undeclared AI use less useful. Highly specific, context-dependent tasks tied to a particular workplace, learner experience or in-class discussion are much harder to outsource wholesale to a generic AI tool than an open-ended essay question.

Supporting staff to make these calls

Assessors and IQAs need confidence and a shared understanding to handle these situations consistently, which is exactly the kind of thing that structured training addresses well; providers looking to build this across a team can look at our qualifications and CPD courses for assessor and IQA routes that now build AI and assessment integrity into their content.

A fair balance

The aim is not to treat every learner as a suspect. Most learners who use AI are doing so for legitimate support, drafting help or accessibility reasons, not to cheat. A workable approach protects assessment integrity without turning every marking session into an investigation, and that balance comes from clear policy, sensible task design and evidence-based conversations rather than blanket suspicion or an over-reliance on software that was never built for the job.