By the Feedback Flows education team

Evidence Over AI Detection Scores

A single AI percentage can start an argument. Evidence of how a learner writes helps you have a fair conversation. Institutions that are rethinking detector-only workflows usually want the second of those.

Why a lone score is weak

Detector scores can shift with genre, subject, EAL writing, shorter scripts and formulaic academic prose. Without quoted passages and context, a percentage invites accusation rather than enquiry. Several universities have already warned staff that AI-detection software should not be treated as proof on its own.

What better evidence looks like

  1. Writing history. Compare a new piece with the learner’s own past work, not only with a global model of “AI style.”
  2. Quoted reasoning. If a tool flags a passage, it should be able to show the exact text it is talking about.
  3. References. Check whether citations resolve and whether they support the claim the learner made.
  4. A recorded outcome. A teacher or tutor notes what was discussed and what was decided. The profile builds over time.

From detecting AI to evidencing learning

Integrity tooling is most useful when it sits beside the mark and helps you understand the learner’s process: consistency with earlier drafts, strength of sources, and whether the writing voice matches what you already know. The goal is not a courtroom verdict from software. It is enough evidence to decide whether a conversation is needed, and to keep that conversation fair.

How Feedback Flows approaches this

Three checks sit on the learner profile:

  • Writing fingerprint compares new work with that learner’s own baseline.
  • AI writing detector adjusts for writer context and only highlights passages it can quote back exactly.
  • Reference checker verifies citations against Crossref and the claim in the text.

A person records the outcome. Nothing publishes as a final integrity finding without human judgement. For the wider data picture, including that learner work is not used to train or build AI models, see how we handle learner work.

If you are evaluating providers more broadly, use the companion checklist: questions to ask about student data and AI. Universities and colleges can also start from for universities.