Assessing theses: plagiarism, citations and guidelines

Academic work follows rules: quote correctly, cite sources traceably, mark others' ideas, make one's own contribution recognisable. Many universities have set out these rules in guidelines or statutes on good research practice. For assessment to be fair, the rules must be known in advance, clearly worded and the same for everyone.

A good guideline numbers its rules so that every objection in the assessment report can refer to a specific rule. It specifies which version applies to which piece of work — sensibly the one in force when the topic was issued — and which deviations can be agreed with the supervisor. And it states openly which rules can be checked by machine and which only a human can judge.

Text-matching software compares a piece of work with other texts and shows overlaps. That is useful but not a verdict: an overlap may be a correctly marked quotation, a technical formula or a standard phrase. Only review in context shows whether impermissible copying has occurred.

Tools that check quotations against their sources go further: does the quotation appear like this in the cited source? Is the page reference right? Does the source exist at all? Now that texts can be written with language models, invented sources are a real problem. Such a check finds them and justifies the finding with location and rule.

Tools claiming to recognise AI-written text deliver probabilities, not proof. They err in both directions and demonstrably disadvantage people writing in a language other than their first. An accusation of cheating based solely on such a score hardly holds up. Concrete findings are robust: non-existent sources, false quotations, demonstrable copying.

Legally, an examination grade is a decision with legal effect. Under Article 22 GDPR it must not be based solely on automated processing. And AI systems intended to evaluate learning outcomes are high-risk systems under Annex III point 3 of the AI Act. This implies a clear architecture: the machine produces findings and suggestions, the human reviews every finding and sets the grade.

A fair procedure checks all works in a cohort with the same tools, records what was and was not checked, and gives authors the chance to comment before a breach is established. The result is an assessment report that supports every finding with rule, location and source and justifies the examiner's grade.

Most rule breaches stem from ignorance, not intent. Universities that teach academic writing early and practically, make their guideline understandable and state clearly how AI tools may be used in writing reduce conflicts. An open rule on AI use — for example allowed for research and language editing, with disclosure in an appendix — is fairer than a blanket ban that cannot be enforced.

Assessment formats help too: a short oral defence of the work quickly shows whether someone has mastered their topic. For students: document sources carefully, check quotations word for word and know your own university's rules.

An accusation of impermissible copying has consequences up to failing the work or revocation of a degree. A fair procedure therefore lets the author learn of the findings and comment before a decision. Much is often cleared up this way: a forgotten footnote, a different edition of the source, an exception agreed with the supervisor.

Decisions on consequences are not made by the software, and usually not by the examiner alone, but by the body competent under the examination regulations, such as the examination board. The complete file — work, findings, statement, log and wording of the rules cited — makes this decision traceable and able to withstand judicial review.