[Opinion] What can we do instead, after moving away from AI detectors?

“ChatGPT is really good at bullshitting!” a student said in my genomics class in the first semester of 2023. It was the first time since ChatGPT’s sudden emergence in late November 2022 that I had invited students to use it to answer some test questions in class, then check the answers, identify absurdities and discuss ways to improve them. At the time, ChatGPT’s answers were often riddled with fabrications and exaggerations.

University teachers worldwide soon recognised an enormous problem: how could universities uphold academic integrity when generative artificial intelligence (AI) could complete students’ assignments in seconds? University leaders were divided over AI. Oxford, Cambridge and some other UK universities banned ChatGPT at the time, while most other universities welcomed it and adapted. Universities in the US and Australia were much more open, leaving professors to set their own classroom rules.

At Monash University, for example, teachers have been able to decide how students may use various AI tools in assignments ever since ChatGPT was launched. Many have opted for “unrestricted use”, subject to academic integrity and verification requirements, or “guided use”, allowing AI at specific stages, such as brainstorming or grammar checks. Students must also disclose the tools and prompts they used in an “Acknowledgements and Declaration” section.

At the same time, the university has advised against using AI detection tools to determine whether assignments were AI-generated – even though Turnitin, the leading plagiarism-detection platform, claimed at the time to have improved its system to catch dishonest students.

This policy of embracing ChatGPT has certainly delighted students, but it has made life difficult for teachers. We know that most students use AI for their assignments, but which of them are working honestly and conscientiously? And how are those using AI doing so?

We are all at a loss.

A survey released this March by the UK’s Higher Education Policy Institute (HEPI) found that 94% of university students, out of a sample of 1,054, said they used AI to help with assignments. The problem is that AI tools are suitable only for people who already possess subject expertise and critical thinking skills. Beginners lack the knowledge to distinguish fact from fiction, making these tools unsuitable for them. Moreover, students who rely on AI will stagnate in their analytical abilities, critical thinking and subject knowledge, leaving them helpless without it.

A bigger problem is that, more than three years on, commercially available AI detection tools remain unreliable. Classroom experience tells us that AI detectors rely too heavily on unverifiable probabilistic estimates, unlike plagiarism detectors. Manufacturers also measure accuracy under controlled conditions. In practice, detection results often have very high false-positive and false-negative rates, and different tools produce inconsistent findings.

Some recent studies suggest that AI detectors may be gradually improving. A report this June from the Free University of Brussels in Belgium, for example, found that false positives were rare with GPTZero, Pangram, Copyleaks and Turnitin. A team at Brazil’s University of São Paulo also published a study this May finding relatively strong performance from QuillBot and from Sapling used alongside Copyleaks. An analysis this February by Sultan Qaboos University in Oman found that Originality outperformed Turnitin in overall accuracy, at 0.69 versus 0.61. Yet all these authors issued the same warning: AI detectors should not be the sole basis for determining academic misconduct.

Meanwhile, a growing number of recent studies have found that AI detection does more harm than good. A February 2025 study by a team at Canada’s University of Lethbridge showed that even using non-generative writing aids such as Grammarly and EditPad to improve grammar could trigger false positives. A paper published this August by a City University of Hong Kong team analysed 13 AI detectors, found their assessments unreliable, and called for a move away from reliance on detection software and a redesign of assessment methods.

In short, an AI detector’s estimate of the “extent of AI involvement” cannot be equated with evidence that a student has cheated.

This September, the New South Wales Education Standards Authority announced a ban on schools using AI detection software to determine whether students had cheated. Several Singapore universities also decided in August to reassess the reliability of AI detection. The Singapore University of Social Sciences has stopped using Turnitin’s AI detection feature, while Nanyang Technological University announced that it would stop using it from 2027. The National University of Singapore, Singapore University of Technology and Design, and Singapore Institute of Technology do not use such detectors.

Echoing the concern about “unverifiable probabilistic estimates”, a study by a team at Australia’s Charles Sturt University, published in the Journal of Higher Education Policy and Management in January 2026, argued that AI detection tools are conceptually unsound. It cited methodological flaws, breaches of procedural fairness, unverifiable outputs and a failure to account for AI-assisted rather than AI-authored work. The tools cannot meet the evidential threshold required for academic integrity investigations and therefore should not be used in education, the study concluded.

Most of the world’s leading universities apply the common law standard of the “balance of probabilities” when determining academic misconduct, rather than the criminal law standard of proof beyond “reasonable doubt”. Monash University’s Student Academic Integrity Procedure, for example, explicitly states that decision-makers must be satisfied that the alleged conduct was “more likely than not” to have occurred, and that “the required standard of proof is the balance of probabilities, not beyond reasonable doubt”. AI detection cannot meet the “more likely than not” standard and therefore cannot serve as evidence of student dishonesty.

In sum, it currently appears impossible to use detection tools to prove beyond reasonable doubt that someone used AI to “produce an assignment”. More seriously, building academic integrity procedures around these tools could harm honest students without deterring dishonest ones. Universities have a responsibility to uphold academic integrity, but they must also recognise the problems and limitations of AI detection tools, rather than treating them as a solution to cheating.

Over nearly a century of change in educational technology, computers, mobile phones, the internet, social media, Wikipedia and other new technologies and platforms have all disrupted education. Yet educators have eventually learnt to integrate them into teaching and learning.

First, as with these earlier developments, teachers must accept AI and allow students to use it legitimately under clear, simple guidelines – to help them understand material, improve grammar, brainstorm and complete assignments or reports.

Second, assignments and assessment methods must be redesigned. Teachers must learn to devise assignments or examination questions that AI cannot easily answer, helping students develop their skills through, for example, in-class writing, oral defences, personalised tasks and questions demanding a high level of critical thinking.

Third, universities must place much greater emphasis on AI literacy and skills. Generative AI cannot produce innovative concepts or improve students’ cognitive abilities, but rather than banning students from using it, we should guide them in using it properly.

Fourth, determined cheats will always find ways around the rules; banning AI will not stop them. Universities cannot bar the many diligent students who use AI ethically simply because some people might misuse it.

For now, the right approach is to acknowledge AI’s benefits for learning, provide introductory guidance for students and teachers, and encourage students to use AI appropriately in assignments, freeing up time for deeper learning and critical thinking.

(The author is a visiting associate professor at Monash University Malaysia.)