Generative AI is weakening a long-standing assumption of education: that the quality of submitted work is reliable evidence of the learner’s own capability.
For a long time, much of education has relied on a fairly simple inference.
A student produces an essay, solves a problem or completes an assignment. The quality of that work is then used as evidence of what the student knows and can do.
Generative AI complicates that relationship.
A student can now produce work that is clearer, more structured and more sophisticated than they could produce independently. This does not make the work worthless. AI can support learning, explanation and feedback. But it does make the output increasingly ambiguous as evidence of capability.
The OECD’s Digital Education Outlook 2026 draws precisely this distinction. Generative AI can improve learning when it is used with clear pedagogical objectives. But when it removes the intellectual effort involved in performing a task, students may achieve better immediate results without developing correspondingly stronger understanding.
The emerging problem is therefore larger than cheating.
It concerns what educational assessment can reliably measure.
From producing work to demonstrating capability
Before generative AI, a reasonably strong assumption could often be made that producing a competent piece of work required possession of many of the capabilities embodied in it.
Writing a persuasive essay required some combination of comprehension, argumentation, synthesis and writing ability. Producing a functioning piece of code generally required at least some programming capability. Translating a text required knowledge of both languages.
The relationship was never perfect. Students received help from teachers, parents, editors, calculators, search engines and other tools.
But generative AI changes the degree of separation possible between the quality of the output and the capability of the person submitting it.
That distinction matters because educational institutions do more than facilitate learning. They also certify it.
A diploma, examination result or degree communicates something to universities, employers and society about what its holder is presumed to know or be capable of doing.
If the evidence underlying that certification becomes less reliable, the problem eventually moves from classroom practice to the credibility of the credential itself.
Institutions are already responding
Some responses are becoming visible.
In August 2026, Denmark’s Ministry of Education announced measures addressing inappropriate AI use in upper-secondary education. Among them are oral defences of certain take-home examination assignments, greater use of controlled conditions for written work and monitoring tools during examinations. The ministry explicitly linked the measures to the need to protect students’ competencies, academic standards and capacity for independent thought.
These measures can be interpreted simply as anti-cheating controls.
There is another interpretation.
They represent a shift in how capability is verified.
An oral defence asks whether the student can explain and defend the work. Controlled writing makes the production process more observable. Interactive assessment allows an examiner to vary the question and observe reasoning in real time.
The common feature is that assessment moves away from relying exclusively on the finished product.
That may prove more significant than any particular rule concerning AI.
The assessment problem may become a credential problem
One response would be to prohibit AI wherever independent capability needs to be measured.
That will sometimes be appropriate.
But prohibition becomes increasingly difficult as AI is integrated into ordinary software, workplaces and professional practice. Education also has a legitimate reason to teach students how to use tools they are likely to encounter outside school.
This creates a tension.
Institutions need to assess what a person can do without assistance, while also preparing them to perform effectively with assistance.
Those are no longer necessarily the same capability.
A student might need to demonstrate independent mathematical reasoning while also learning to use AI to explore complex problems. A writer might need to demonstrate command of language while also becoming competent at editing machine-generated text. A future analyst might need both independent judgment and the ability to work effectively with increasingly capable analytical systems.
Assessment therefore may need to become more explicit about what exactly is being measured.
Not:
Can this student produce a good result?
But:
Which capabilities does this result demonstrate?
That is a harder institutional question.
What may change
If the distinction between output and capability continues to widen, assessment could gradually become more varied.
Oral examination may become more important. So may supervised tasks, iterative work, demonstrations, interviews, portfolios that reveal process, and assessments in which students must respond to unfamiliar situations.
There may also be greater separation between learning activities, where extensive AI use is acceptable or encouraged, and verification activities, where assistance is deliberately restricted.
The underlying purpose would not necessarily be to make education less technological.
It would be to preserve the informational value of assessment.
A qualification has value partly because others believe it communicates something meaningful about its holder. Maintaining that trust may require assessment systems to adapt as the technologies surrounding them change.
What would weaken this signal?
The direction is not predetermined.
This signal would weaken if evidence showed that unrestricted AI-assisted performance consistently produces durable independent learning, or if reliable methods emerged for inferring individual capability from AI-assisted work.
It would also weaken if employers and other institutions stopped caring about independently demonstrated capability and increasingly assessed people entirely through observed performance in AI-enabled environments.
Both are plausible possibilities.
For now, however, the institutional response appears to be moving in the opposite direction: toward making capability more directly observable when the authenticity of the finished product becomes harder to establish. Denmark’s measures are one early example rather than proof of a general transition.
Watch next
The important signals will not be the number of schools banning ChatGPT.
Watch instead for changes in how competence is demonstrated: expansion of oral examination, supervised assessment and interactive testing; explicit separation between AI-assisted and independent work; changes in university assessment rules; and employers placing greater weight on direct capability tests rather than qualifications alone.
The deeper question is not whether students will use AI.
They will.
It is whether institutions can still determine what the person behind the work can actually do.
Leave a Reply