There is a form of academic dishonesty that plagiarism detection software cannot catch, that AI detection tools cannot reliably identify, and that most university policies have not yet found adequate language to describe.
It is not plagiarism in the traditional sense. The words are original. The structure is the student’s own. The argument, such as it is, has not been copied from anywhere. And yet something essential is missing: something that the assignment was designed to develop, and that the submitted work gives no evidence of having developed.
Call it the performance of understanding without its substance. It is, in a different register, what we have elsewhere called the hollow scholar.
What Has Changed
Academic dishonesty is not new. Students have always found ways to produce work that misrepresents their actual engagement with the material. Ghostwriting services, essay mills, collaborative work submitted as individual: these predate AI by decades.
What AI has changed is not the existence of this problem but its accessibility, its scale, and crucially its detectability.
A student who purchased an essay from a writing service in 2015 was taking a meaningful risk. The essay might not match their established writing style. It might contain knowledge or references beyond what the course covered. The transaction itself left traces. Detection, while imperfect, was possible.
A student who uses a well-prompted AI tool in 2026 faces a different risk profile entirely. The output can be calibrated to match their writing level. It can be prompted to produce the kind of prose that reads as engaged without being so, a phenomenon explored in depth in the real reason AI content feels empty. It can be restricted to course material. It can be lightly edited to introduce the kind of minor errors and stylistic inconsistencies that characterise genuine student work. The traces are minimal. Detection, even with the best available tools, is unreliable enough to be legally and ethically problematic as a basis for accusation.
The practical barrier to this form of dishonesty has collapsed. What remains is the ethical one — and ethical barriers without practical reinforcement are historically insufficient.
Why This Form Is Harder to Detect
Traditional plagiarism detection works by matching submitted text against a database of existing text. The logic is straightforward: if your words appeared somewhere else first, the system finds it.
AI-generated text does not appear anywhere else first. It is generated freshly for each prompt. Detection tools that attempt to identify it do so by looking for statistical patterns associated with AI output, including low perplexity, high consistency, and certain syntactic regularities. These tools have documented false positive rates that make them unreliable as evidence in formal proceedings. Several universities have already faced legal challenges after acting on AI detection results that proved incorrect, a dynamic explored in caught in a trap: suspicious minds and AI content.
The result is a practical enforcement gap. Institutions know the problem exists. They cannot reliably prove it in individual cases. Policy without enforcement capacity is at best a statement of values.
The Deeper Problem
The harder question is not whether AI-assisted academic dishonesty is detectable. It is what it costs the student who engages in it.
Academic assignments are not primarily products. They are processes. The value of writing an essay is not the essay — it is what happens cognitively in the student who writes it. The struggle to articulate an argument, the encounter with evidence that complicates your initial position, the discipline of organising complex ideas into coherent prose: these develop capacities that persist beyond the assignment.
A student who bypasses this process with AI assistance does not merely risk academic penalty. They deprive themselves of the development the assignment was designed to produce. They graduate with a credential that may not accurately represent their capabilities, and enter professional or academic contexts where those capabilities are assumed and find themselves underprepared.
This is a cost that falls entirely on the student — and one that no detection system can prevent, because it occurs regardless of whether the dishonesty is ever discovered.
What Institutions Can Do
The response to this challenge cannot rely primarily on detection. It requires a rethinking of assessment design.
Assignments that can be completed adequately by AI with minimal student input, such as the standard argumentative essay on a well-documented topic or the literature review of a canonical field, are increasingly poor instruments for evaluating student understanding. Not because AI completing them is inevitable, but because the possibility of AI completion reveals that these assignments were always measuring a relatively thin slice of what understanding actually involves.
More robust assessment designs share certain features. They require students to engage with specific, recent, or localised material that AI cannot have been trained on. They incorporate iterative development with documented process, including drafts, supervisor feedback and revision histories, that make the final product accountable to a visible process. They include oral components where students are asked to defend, explain, and extend their written arguments in real time.
None of these approaches is new. Many were considered best practice before AI made them urgent. The pressure AI creates on assessment design may, in the long run, produce assessment that more accurately reflects genuine understanding — and that would be a genuine improvement.
A Note on Responsibility
It would be incomplete to discuss this issue without acknowledging the conditions that make AI-assisted dishonesty attractive.
Students under extreme time pressure, managing multiple simultaneous deadlines, working part-time to fund their studies, experiencing mental health difficulties, or writing in a second language under significant linguistic pressure: these students face a different calculation than the question of academic integrity in the abstract.
This does not excuse dishonesty. It does suggest that institutional responses focused exclusively on detection and punishment, rather than on the conditions that produce the behaviour, are likely to be ineffective and unjust in equal measure.
The goal worth pursuing is assessment that students want to engage with honestly, because it is designed well enough, and supported well enough, that honest engagement is the more attractive option.
For students who want to use AI tools responsibly and effectively, supporting genuine thinking rather than replacing it, our guides on best AI tools for academic writing and the problem with relying on AI for academic writing offer a more constructive starting point.
Disclosure: Some links in this article are affiliate links. We only recommend tools we’d genuinely use ourselves.
