Saturday, November 7, 2026

Denmark's Defend-Your-Work Approach Is a Canary in the Coal Mine for Take-Homes in the Private Sector

Josh Priollaud

A prediction. Denmark kept the home assignment and added a defense. I think hiring ends up in the same place.

In August, Denmark decided that an exam written at home only counts if the student can defend it out loud. I think hiring is about two years behind. By the end of 2027, a take-home assignment with no defense attached won't be treated as evidence by anyone serious about hiring. Here's why I'm confident enough to put a date on it.

What Denmark actually did

On 6 August, Denmark's education minister, Magnus Heunicke, announced an emergency package against AI cheating in upper-secondary schools. The headline measure: every exam written at home must be defended orally. CNN picked it up the next day, which is how most of us heard about it.

What struck me is the choice. Denmark kept the work and changed what makes the work count.

The follow-through was fast. By 25 September, six parties from left to right had signed a deal with money attached: DKK 5 million a year, oral exams from January 2027. Eighty-seven days from a minister floating an idea to a funded, cross-party deal. Education policy moves that fast when everyone in the room already knows the old test is broken.

They had reason to know. In a survey of 647 Danish teachers, 81% said AI weakens their students' learning, and the Danish Evaluation Institute found two in three students admit to having cheated with AI on written work at least once. Then, when the ministry published this year's exam grades in October, written Danish on the main academic track had dropped to its lowest level since 2012. The students who started high school the year ChatGPT launched did worse the moment someone was watching.

My favorite detail is from Copenhagen Business School, where external examiners reported students whose written theses were headed for top marks and who then couldn't account for large parts of them when asked. Every hiring manager has met that candidate.

Hiring has the same problem, and the same three doors

A take-home assignment is a home-written exam. It has always been faked a little, and it's now fakeable completely, in minutes, for free. If you send one out today and grade the artifact, you are grading the artifact. You have no idea who made it.

The mapping is almost one to one.

What Denmark did in schools	The hiring equivalent
Major written assignment done at home	Take-home assignment
Oral exam on that assignment	Live walkthrough of the take-home
Screen monitoring during written exams	Proctored or locked-down tests
More writing done at school	Live or onsite exercises
Teachers' proposal: students declare their AI use	Candidate shows their chat history

There are three ways out, and they're the same three Denmark looked at.

Door one: stop setting the assignment. Four days after Denmark's announcement, New South Wales moved this way, and by September it had capped take-home assessments for its school-leaving exam. In hiring, this looks like scrapping the take-home and putting everything into live rounds.

Door two: detect the AI. Nobody serious believes this works anymore. Norway's expert committee on AI in universities told institutions to stop relying on detectors. The Danish package doesn't mention them.

Door three: keep the work, add a defense. Denmark's choice.

Why door one doesn't work for hiring

Live assessment is the best evidence there is. Watching someone think beats reading what they produced. It also has two problems that grow with every candidate you add.

The first is arithmetic. A live working session occupies your people for as many hours as the candidate works. Run that for eight semifinalists and you've spent a working week of senior staff time before anyone gets an offer. We've measured how much signal each stage of the process gives you, and there's a reason panels sit at the end of the funnel. (Working trials also raise pay and classification questions; I've written about those elsewhere.)

The second problem is the one people under-weight. Working under a microscope doesn't mirror any real job. Nobody at your company writes code or builds a deck with three people silently watching. What you measure in that room is composure under observation, and you override the thing you actually wanted to see, which is the person's ability, with artificial nervousness. Plenty of good candidates will simply decline. Danish students raised the same worry within days of the announcement, and an Australian education researcher laid out the fairness problems with high-pressure oral formats in detail. They're right that pressure distorts.

So the take-home survives. For the five to eight people who've cleared your screens and would otherwise be heading for a panel, letting them do real work on their own schedule, with their own tools, is still the most practical and most humane thing you can ask. It just can't be the whole thing anymore.

Door three, done the Danish way

Here's the part of the Danish model that I think hiring will borrow, with one change. A Danish oral exam has two people on the other side of the table: the examiner, who asks the questions, and an external censor, there to keep the grading honest. Asking and judging are separate roles. In Denmark they happen in the same room at the same time, and the DKK 5 million a year pays for both people's hours, roughly 9,000 times over.

Hiring can pull the two roles apart in time. The candidate explains their work live, to someone who asks unscripted follow-ups, because a rehearsed monologue is just another artifact. The person who decides reviews the recording later, at double speed, skipping to the parts that matter. The asking still has to be live. The judging can wait. NYU Stern ran an AI-voiced oral final this spring with a voice agent asking the questions, so the examiner's chair can now be filled for nearly nothing.

Live defense, async review.

There's a third piece the Danish teachers' union asked for and the government hasn't adopted yet: students declaring how they used AI. In hiring this has a cleaner version. In a perfect world the candidate opens their chat history during the walkthrough and talks through it. How they framed the problem, what they asked for, what they rejected. The reviewer can check it against the work. Since the defense runs on a screenshare, that history is already on the screen, under the candidate's own name, and the examiner can ask about any of it. Chat history is the new scratch paper, and good candidates will want to show it.

As of November 2026, sharing an AI conversation means one link per chat, on whichever platform it lives. Real take-home work spreads across several AI chats, the everyday software of the trade, working documents, drafts and the finished artifact. There is no way to hand that trail over as one thing. On a screenshare, the candidate opens all of it in a couple of minutes and narrates as they go: here's where the brief got ambiguous, here's the approach I threw out. A pile of links gives a reviewer text to read. A walkthrough gives them the candidate explaining the trail while they look at it.

Explaining work you did is also the most natural thing in professional life. It's every standup, every design review, every client call.

The prediction, so you can hold me to it

By the end of 2027, hiring teams that take their evidence seriously will treat an undefended take-home the way Denmark now treats an undefended home exam: as not evidence. Nobody will mandate it. People will simply stop paying for artifacts they can't attribute.

The one thing that could slow this down is that hiring has no scoreboard. Denmark acted on what teachers and examiners were telling it, and two months later the national exam results confirmed them in a single number. A company that keeps sending undefended take-homes gets no such confirmation; the AI-written submissions come back looking fine, and the cost shows up months later as a hire who can't do the job. Hiring teams will have to notice without the number.

Common questions

What does an oral defense measure? Whether the person understands the work in front of them, which is what you need to know anyway. People who made the work can talk about it in detail and handle the second follow-up question. That holds whether or not a machine was involved.

How do nervous candidates do in a defense? Better than in a live working session. Explaining your own work is routine professional behaviour, and a self-scheduled session removes the panel. Every candidate gets the same brief and the same examiner; the follow-ups differ because the work does.

Why keep the take-home at all? Interviews are talk about work; a take-home is work. Together they're the best evidence you can get for the money, and the defense is what joins them.

Will Denmark's rule spread? No government has copied it yet. Norway's expert committee recommended the same pairing for universities, and professors at Cornell, Penn and NYU have adopted it on their own. I'd watch what happens when the first Danish defenses run in early 2027.

Pudding runs paid work demonstrations where candidates defend their work live on a screenshare, chat history included, to an agent that asks unscripted questions, and reviewers judge the recording on their own time. See how it works.

Appendix: the record

The tables below carry the facts behind the argument, with dates and links, for anyone who wants to check the claim or watch it play out. Updated 7 November 2026.

Denmark, from idea to law

Date (2026)	What happened
30 Jun	Heunicke, about a month into the job, tells Berlingske some exam formats are obviously being gamed and more examiners will cost money (via Ritzau).
21–24 Jul	Berlingske obtains 40 external-examiner reports from Copenhagen Business School: students with top-mark written theses fail the oral (TV 2).
6 Aug	Emergency package: oral defence of home-written exams, screen monitoring, more writing at school.
13 Aug	Survey of 647 teachers: 81% say AI weakens learning. Same report cites the Danish Evaluation Institute: two in three students admit AI cheating on written work.
14 Sep	Heunicke says a second package is possible.
25 Sep	Six-party deal: oral exam on the HF major assignment from January 2027, DKK 5 million a year through 2031.
28 Sep	Education and higher-education ministers meet student groups to start a national AI strategy covering universities too.
29 Oct	2026 exam grades: written Danish on the main track at 6.2, down from a 6.8 peak in 2018, lowest since 2012. Minister promises more.

Who else hit the same wall, and which door they took

Where	What happened	When	Door
Norway, universities	Government expert committee advises against home exams as the only assessment and recommends adding oral questioning where they stay; advises against AI detectors (Utdanningsnytt). Final report NOU 2026:14 scheduled for delivery 6 Nov 2026.	Dec 2025	Defend (advice, not rule)
Germany, Saxony	Students declare AI tools used; a text that departs from a student's usual style triggers an exam conversation (state tracker).	Feb 2026	Defend, on suspicion
United States	Cornell, Penn and NYU Stern professors adopt oral defences of written work; NYU's runs with an AI examiner (AP via Fortune).	Mar 2026	Defend, bottom-up
England	Exam regulator Ofqual says dropping extended-writing coursework is "never off the table" and orders closer scrutiny of coursework in new qualifications (Apr, Jul).	Apr–Jul 2026	Leaning restrict
Australia, New South Wales	Four days after Denmark, the state moves to pause unsupervised take-home tasks; final rule allows one, worth at most 15%, for the Class of 2027 (NSW Government, Information Age).	Aug–Sep 2026	Restrict
Sweden	Government orders the Schools Inspectorate to review how schools handle AI cheating; report due April 2027 (Regeringen).	Jan 2026	Undecided
Ireland	Reformed Leaving Cert gives coursework at least 40% of the grade; teachers' union says members cannot verify AI use (Irish Examiner).	2025–2031 rollout	More take-home, no defence

What employers are doing

Finding	Source	Door
65% of US hiring managers have caught applicants using AI deceptively; 39% are doing more in-person interviews to verify authenticity. 4,136 respondents, four countries.	Greenhouse, Nov 2025	Restrict
Google, Cisco and McKinsey bring back in-person rounds; 72.4% of recruiting leaders use them against fraud. Gartner's "fraud" here means impersonation.	Gartner via Computerworld, Aug 2025	Restrict
Startup Passion.io drops some take-homes after receiving AI-generated work, moves to screen-shared live writing.	Wall Street Journal, Sep 2026, via FavTutor	Restrict
Keep the take-home, allow any tools, add a 30-minute live debrief.	Dover, Aug 2026	Defend
"Reverse proof of work": polished submissions are cheap, so ask candidates to defend a recommendation and explain what they rejected.	Antoine Vastel, Castle, via Security Boulevard, Oct 2026	Defend

Sources

Josh Priollaud

Founder @ Pudding

Josh Priollaud is the founder and CEO of Pudding, a Siligo Ventures portfolio company. Josh has worked in business development roles at Cloud9 IDE, Webgility, MuleSoft and Adobe. Before his Silicon Valley career, Josh served as an Economic Development Adviser in the Peace Corps. Josh holds an MBA from the Leavey School of Business at Santa Clara University and a BS in Business Administration from Portland State University. Josh has been a full stack developer (Ruby on Rails, React) since 2020.

LinkedIn siligo.tech

Back to blog