Jisc says better AI detectors still do not prove misuse
Jisc's National Centre for AI says commercial detectors rarely flagged human writing in recent tests. It says a flag still does not show whether AI use broke the rules and that assessment must be redesigned.
Jisc's National Centre for AI published a review of AI detection tools on 15 October 2026, saying they are improving while the problem grows harder.
The post by Michael Webb updates his 2025 advice. He says commercial detectors wrongly flagged 0% to 4% of human writing across four studies. He warns that the human texts tested were arguably good writing. They did not include weak first-year essays or heavily proofread work by second-language writers.
Better detection of AI involvement doesn’t necessarily mean better evidence of inappropriate AI use.
Webb cites a preprint, which is not peer reviewed. He says it found that Pangram, the tool he calls the best performing at present, caught 34% of text from GPT 5.4 soon after that model's release. A May update fixed this. Adversarial rewriting in the same paper cut its recall from 94.7% to 24.1%.
Webb cites a study with 30 texts in each of five conditions. Human reviewers scored 19% overall, against 20% for guessing. His advice is not to ask staff to spot AI use as they mark.
Some awarding bodies still require a detector. Webb says there are none Jisc can honestly recommend that suit small cohorts, are reliable enough and have passed data protection checks. He asks colleges in that position to contact Jisc.
For students, the post says a flag is the tool's judgement and not an established fact. It says a flag does not show whether the use was permitted. For lecturers, it calls assessment redesign the only viable way forward and asks that any new measure be reviewed for equality effects.
Webb calls wearables the next challenge for in-person exams. Ofqual said phones and smart devices made up 2,225 malpractice cases in summer 2025. That was 44.3% of student cases in school exams, not university assessment. The post is advice, not a rule. Webb says he cannot find whether UK institutions qualify for access to one AI text watermark checker.
Sources
- Jisc National Centre for AI AI detection and assessment: an update for 2026
- Jisc National Centre for AI AI detection and assessment: an update for 2025
- Ofqual (GOV.UK) Cheating in exams with high-tech smart devices poses growing threat, warns Ofqual chief in new podcast
- arXiv Hitting a moving target: test-time adaptation for AI text detection under continual distribution shift