When the Artifact Gets Cheap, Where Does Assessment Move — Why Denmark Chose the Oral Defense

• When the Submission Alone Stopped Telling You Anything
• What Denmark Actually Announced
• The Three Measures Target Different Things
• The Center of Gravity Moves from the Artifact to Authorship
• What an Oral Defense Catches and What It Does Not
• From Here On This Is My Extrapolation
• If You Move It to Hiring Assignments and Code Review
• The Cost of Solving It with Surveillance
• Wrap-Up and Sources

When the Submission Alone Stopped Telling You Anything

Any team that has run take-home hiring assignments has had the same experience over the past few years. The average quality of submissions has visibly risen. The structure is clean, there are tests, the README is well written. And yet the discriminating power has gone down. Everything is bunched at the top, so you cannot choose.

This is not a problem of candidates getting worse. It is a revelation of what we were measuring. The quality of the submission was a proxy metric for that person's understanding. Because the cost of producing it was high, quality was itself evidence of effort and understanding. When the cost falls, that link breaks. The metric stays the same while the thing it measured disappears.

In education this problem arrived much bigger and much faster. And one country has issued an official answer to it.

What Denmark Actually Announced

The Danish Ministry of Education announced a package of immediate measures against AI cheating at the upper secondary level. I write only what I confirmed in the source.

There are three measures. First, an oral defense becomes mandatory for exam assignments written at home. In particular, roughly nine thousand HF programme students who submit a large written assignment every year will defend that assignment orally. Second, tools that monitor students' screens during written exams are recommended, along with firewalls that filter reachable content not only during exams but also during class time. Third, written assignments are encouraged to be carried out inside the school, a controlled environment.

There are attached requirements as well. On large written assignments students must clearly declare whether they used AI, and there must be no AI access during preparation for oral exams. The ministry stated it would consult with the relevant school organizations on how to implement this, and added that it would separately prepare a national AI strategy covering the education sector as a whole.

The Three Measures Target Different Things

What makes this package interesting is that the three are entirely different in character.

• The oral defense changes what is assessed. It does not assess the writing; it assesses the understanding of the person who wrote it.
• Screen monitoring and firewalls block the cheating itself. They leave the assessment method alone and tighten control.
• Doing the work inside the school changes the conditions. It changes neither what is assessed nor what is blocked; it only moves the location.

On the third, local organizations have already attached conditions. A joint statement from the schools organization, the teachers organization, and the students organization welcomes the package as a necessary and constructive first step, while noting that it addresses only the short term. And the students organization side stated that any increase in writing done at school must be designed to genuinely strengthen learning rather than merely relocating the work from home to school. That point is exact. Moving the location alone does not solve the proxy metric problem.

The Center of Gravity Moves from the Artifact to Authorship

Of the three measures, the structurally largest change is the first. The other two are a defense of the existing assessment; only the first changes the object of assessment itself.

The logic is simple. When the artifact gets cheap, the quality of the artifact and authorship come apart. Good writing used to be evidence of a person capable of writing it; now it is not. So authorship has to be confirmed separately, and the oldest and simplest way to confirm it is to ask the person to their face.

Spoken assessment has the property that it has no cost of duplication. Writing can be done by someone else, copied, prepared in advance. An answer given on the spot cannot. This is why the oral defense has suddenly reappeared. Not because it is a new technology but because it is an old technology that cannot be duplicated.

What an Oral Defense Catches and What It Does Not

For balance the limits have to be written down too.

What an oral defense catches well is depth of understanding. Why this approach was chosen, why the alternatives were discarded, what happens if you change a condition here — only the person who actually made those decisions can answer questions like these. Someone who memorized the surface stops at the second question.

What it does not catch is equally clear. Speaking ability and understanding are not the same thing. People who get nervous, people not working in their native language, and people who need more time to organize their thoughts into speech are disadvantaged. That is to say, the oral defense filters out cheating and in exchange brings in a different kind of bias. And unlike surveillance tools, this bias is not easy to see.

From Here On This Is My Extrapolation

Everything up to this point is what the Danish Ministry of Education announced; from here it is my own thinking about moving that logic over to software work. The source material is about education policy, not about hiring or code review.

The part that carries over is the logic of the proxy metric. We too have used the quality of an artifact as a proxy for a person's ability, and the cost of producing that artifact has fallen sharply. So the same conclusion follows. Assessment that looks only at the artifact loses discriminating power, and authorship or understanding has to be confirmed separately.

The part that does not carry over is equally clear. The purpose of school is for the student to acquire the ability; the purpose of a company is to get the work done. Using a tool and producing an answer at school circumvents the learning objective, but using a tool and solving the problem at a company is usually just doing the job well. So the question a company has to ask is not whether tools were used but whether the person can be accountable for the outcome.

If You Move It to Hiring Assignments and Code Review

Keeping that distinction, here is the form the move can take.

For a take-home assignment, shrink the assignment itself and expand the conversation that follows it. Instead of scoring the submission, use the submission purely as material for a conversation. What to ask: which alternatives were considered here and why were they dropped, what failure modes does this part have, where would you change things if the requirements changed like this. These questions reveal understanding without asking whether tools were used.

For code review the same device already exists. We just are not using it. Most reviews end with remarks about the code and never ask about the author's judgment. Instead of adding one more review comment, insert one question. Why this approach? A change for which that answer does not come out is dangerous regardless of its size. The same holds no matter who wrote it.

I will note the side effect of this approach too. Confirming authorship consumes human time. It is more expensive than screening documents. So it cannot go into every stage, and you have to decide which one you make the final gate.

The Cost of Solving It with Surveillance

Finally we have to look again at the second measure. Screen monitoring and firewalls take effect immediately and are quick to deploy. Which is also why they are the approach organizations adopt first.

The real cost of surveillance is not the price of the equipment but two other things. One is the inconvenience the honest majority bears every single time. The other is the signal the system sends. Repeat the signal that we do not trust you at every assessment and the attitude of the people being assessed changes. And surveillance invites new circumvention, which invites more surveillance.

In the Danish package, what I think is worth watching is how the weight between the first and second measures gets adjusted from here. If the weight lands on changing what is assessed, it is sustainable; if it lands on tightening control, it becomes an arms race. Organizations handling the same problem stand at the same fork.

Wrap-Up and Sources

When the cost of making something falls, you can no longer judge a person by what they made. At that point there are two things you can do. Control the conditions of making, or change the object of judgment. The latter is harder and lasts longer.

• Danish Ministry of Education press release — an immediate package of measures against AI cheating in upper secondary schools — announced on August 6, 2026. I confirmed from this document the content of the three measures, the number of students affected, the consultation plan, and the announced national AI strategy.
• CNN coverage — I confirmed here that it takes effect immediately, that it applies to upper secondary education, the requirement to declare AI use, and the blocking of AI during preparation for oral exams.
• Joint statement by Danske Gymnasier and two other organizations — I confirmed from this document the assessment that it is a necessary and constructive first step, the caveat that it addresses only the short term, and the students organization condition on expanding writing inside schools.
• The later part about hiring processes and code review is my extrapolation and is not in the source material. The Danish measures are education policy and are not a recommendation about employment assessment.