Fuss over AI detectors in university assessments is unnecessary

It started as a dispute. Plagiarism detector Turnitin flagged an essay by an Adelphi University student in New York as “100 per cent” AI-generated – which the undergraduate promptly denied. Two other software detectors assessed his essay as entirely human-written.

In January 2026, a New York court annulled the Turnitin finding and ordered it removed from the student’s record, citing problems with both the basis of the university’s decision and appeal process.

Singapore’s universities are already stepping away from these tools. The Singapore University of Social Sciences (SUSS) disabled its Turnitin detector in August. Nanyang Technological University will deactivate its AI detector at the end of 2026, describing current tools as fundamentally unreliable.

The National University of Singapore, Singapore University of Technology and Design and Singapore Institute of Technology do not use AI detectors.

Singapore Management University (SMU) continues to use such tools, but they are never the sole basis for determining academic misconduct.

And indeed, removing unreliable software detectors is the right first step. But the maelstrom over the use of AI exposes a larger issue: Are universities assessing students for the capabilities they will actually need in a workplace where AI is increasingly commonplace?

That question cannot be answered by grading essays more harshly or weeding out all AI use entirely. Nor is a wholesale return to pen-and-paper examinations the solution.

Instead, the task before educators is to reconsider the mode of assessment itself.

Singapore’s move to introduce baseline AI competencies from 2027 recognises that graduates will eventually work with AI and must build up AI savviness. Assessments must now catch up.

Is using AI for academic work misuse?
No AI detector is foolproof. Turnitin itself acknowledges that its model may misidentify human and AI writing and should not be the sole basis for academic disciplinary action.

A 2023 study found seven detectors classified, on average, 61.22% of human-written essays by non-native English writers as AI-generated.

More fundamentally, detectors cannot establish whether a student broke the rules. Using AI to translate a passage or improve grammar is not equivalent to asking it to produce an entire answer.

Even a perfect detector would not solve the larger problem. AI can now write a competent e-mail, report or speech in seconds. Workplaces will often expect staff to use such tools.

Permitting AI usage creates another problem: The better prompter may outperform the student with stronger subject knowledge.

On the other hand, pen-and-paper examinations can still verify some foundations, but returning to them wholesale would not prepare students to use AI responsibly at work.

To resolve it, we must first ask what a finished piece of work actually proves.

For a long time, the conventional assumption was that a finished product proves the student mastered necessary cognitive skills.

Generative AI, able to produce an essay, report, presentation or code in seconds, breaks that logic.

Institutions must therefore clarify what capability assessments must certify.

If a task tests professional production with the tools available, AI use may be legitimate.

If it tests legal reasoning, research integrity, clinical judgment or disciplinary knowledge, a polished output cannot be the only evidence that a student has mastered the necessary.

Prompt writing is a useful skill, but it cannot substitute for the competency a degree is meant to signify.

How assessments can be redesigned
In a world denominated by AI, employers will require fresh graduates who can critically evaluate AI output. The World Economic Forum’s Future of Jobs Report 2025 identifies analytical thinking as the leading core skill, considered essential by seven in 10 surveyed employers.

AI literacy demand is also growing. A junior analyst must recognise when an AI-generated forecast rests on weak assumptions. A junior lawyer must check whether its cited cases support the advice. These roles demand verification, judgment and the ability to revise a recommendation when new evidence emerges.

Critical thinking and judgment are longstanding educational goals. What is new is that graduates must exercise them over plausible answers they may not have produced themselves. Assessment must pivot accordingly, testing whether students can interrogate AI output robustly and judge whether it is safe to act on.

Designing such assessments requires faculty who can look well beyond surface-level accuracy.

An AI-generated recommendation may contain correct facts yet rely on an unsuitable method, overlook contradictory evidence or draw wrong-footed inferences.

Academics – whose research is rooted in defining issues, interrogating methodologies and defending conclusions – must design curriculum that enable students to recognise such vulnerabilities.

Consider this practical example: Students are given an AI-generated proposal recommending that an organisation expand a service. Its figures might be accurate, but drawn from a population unrepresentative of the target consumers.

Under a refreshed assessment framework, students would be judged by whether they can identify the sample bias, evaluate whether that flaw invalidates the recommendation or merely reduces confidence in it, and specify what additional evidence would justify proceeding.

The lecturer could subsequently introduce a tighter budget or conflicting findings and ask students to revise their advice. Each stage tests a different capability: detecting a weakness, judging its significance, identifying how to resolve it and adapting a decision.

Such assessments must be expanded across an entire degree programme. Universities must distinguish between capabilities students must demonstrate independently and which they should exercise collaboratively with AI so that high-scoring AI output does not mask deficiencies in essential disciplinary knowledge.

To achieve this balance, universities should consider deploying three types of assessments.

First, brief oral examination following a written assignment can immediately reveal whether students understand their work. Rather than ask them to summarise their work, faculty can change assumptions or introduce contrary evidence. If students cannot defend their conclusion, they have not internalised the material.

Second, simulations of realistic workplace scenarios where AI usage is expected. Students should submit an “audit trail” of which tasks they delegated to AI, what errors they corrected and what suggestions they rejected, supported by evidence. Marks should reflect how well they defined the problem, justified their decisions and produced a reliable result.

Third, longitudinal performance evaluation. Work-attachment models – such as SUSS’ mandatory 24-week internship programme – allow faculty and industry mentors to observe how students frame ambiguous problems, ask sensible questions and respond to feedback.

No single method offers a panacea. Work attachment cannot replace academic assessment, and students’ opportunities will differ. But a refreshed academic assessment framework – comprising work attachments, selective oral checks and AI-enabled tasks – produces a more credible picture than forensic inspection of one essay for AI use.

• Justina Tan is vice-president, strategic partnership and engagement, and vice-provost, global engagement and strategy, at the Singapore University of Social Sciences. Ben Chester Cheong is a law lecturer at the same university.