Denmark wants classroom AI to add friction. Measure whether students actually think

THE SIGNAL IN ONE SENTENCE

Most classroom AI is sold as a solvent. It removes the stubborn parts of a task, smooths the prose, finds the summary and produces an answer before a student has quite finished asking the question. Researchers at the IT University of Copenhagen want to test the opposite move. On November 9, the university announced a five-year research project called Friction by Ambiguity: Improving Engagement and Critical Thinking in LLM-Supported Learning. Independent Research Fund Denmark is backing it with DKK 7.1 million. The project is led by Professor Jichen Zhu with Associate Professor Christian Hardmeier as co-principal investigator. Associate Professor Pawel Grabarczyk, from the university's Center for Computing Education Research, will contribute expertise in ethics and responsible AI. The work combines human-computer interaction, artificial intelligence and learning science. Its planned test bed is an educational chatbot called Diffrakt. Instead of always delivering a polished final answer, the system is supposed to use controlled ambiguity, open questions and a Socratic style to make students compare interpretations, weigh arguments and form their own positions. The first domain is AI ethics. That is a sensible place to begin because ethical questions rarely arrive with one clean answer tucked beneath the page. A student considering facial recognition, automated hiring or an autonomous weapon has to identify competing values, test assumptions and explain why one tradeoff deserves more weight than another. A chatbot that instantly supplies a confident little essay can short-circuit that work. It can also create the comforting illusion that an unsettled problem has been settled by a machine with excellent punctuation. The Danish project is asking whether a little resistance can help. That question is genuinely interesting. It is also very much still a question. The university announcement describes the beginning of a five-year project. Diffrakt is planned, not a finished classroom product. The researchers have not reported learning gains, retention results, accessibility findings or evidence that an uncertain model can reliably produce the right amount of ambiguity for the right student. So nobody should file this under proven better teaching. The useful response is to build a test that can tell the difference between productive struggle and ordinary confusion. Productive struggle is the mental work that helps a learner connect ideas, notice a contradiction or repair a weak explanation. It is not the same thing as being lost. If a student has enough background knowledge, a carefully chosen question can make the next insight stick. If the student lacks the vocabulary, context or confidence to take the next step, the same question can feel like a locked door with a motivational poster taped to it. That distinction is where the research earns its keep. The first comparison should be brutally simple. Give similar students the same learning objective through several conditions: a direct-answer chatbot, a Socratic chatbot, a calibrated-ambiguity chatbot and a version of the lesson without an AI assistant. Measure what students know before the lesson, immediately after it and again later. Then give them a new problem that requires the same kind of reasoning. The last part matters. A student who repeats the chatbot's vocabulary on the original case may look impressive without having learned how to reason. Transfer to a new case is much harder to fake. Do not rely on satisfaction surveys alone. A direct answer can feel wonderfully efficient. A confusing answer can feel deep. Neither feeling proves learning. Score the reasoning itself. Did the student identify relevant facts? Did they consider more than one plausible interpretation? Did they test a claim against evidence? Could they explain a tradeoff, revise a position and recognize what remained uncertain? Have independent reviewers grade those explanations without knowing which tool the student used. Then measure the human cost. How often did students ask for a hint, request the answer, abandon the task or call a teacher? How long were they stuck before help arrived? Did the system create more useful classroom discussion or simply move support work onto an already busy instructor? Did students with different language backgrounds, disabilities, prior knowledge or confidence levels experience the same intervention differently? That is not a side question. Ambiguity is an access decision. A learner using a screen reader may need a different interaction pattern. A student working in a second language may be reasoning well while struggling with the chatbot's phrasing. A neurodivergent student may benefit from explicit structure instead of hints. Someone with an accommodation may need the direct explanation immediately. A system designed around one idealized student's appetite for puzzles can punish everybody else for failing to enjoy the maze. The safest design therefore needs an exit path. Students should be able to ask for a concrete example, reveal the next hint, request the direct explanation or hand the exchange to a teacher. The system should explain why it is asking another question. A teacher should be able to override the level of friction for a person, a lesson or an entire class. Nobody should have to perform confusion for a chatbot to prove that they tried. The project also plans to investigate technical methods for generating calibrated ambiguity based on model uncertainty. That phrase needs careful handling. A language model's uncertainty signal is not a truth meter. A model can be confident and wrong, uncertain for irrelevant reasons or inconsistent across phrasings. An ethical question can have several defensible answers even when the facts are clear. If the system treats internal model uncertainty as a precise measure of what the student needs, it may produce hesitation where clarity is required and certainty where caution belongs. Calibration must therefore include the learner, the task and the stakes, not just the model. For a low-stakes classroom discussion, an incomplete answer may invite useful exploration. For a safety instruction, an accessibility request or a student asking whether a claim is factually true, withholding a clear answer can be irresponsible. The intervention should be bounded by a teaching objective and tested against explicit stop conditions. The project should also preserve an experiment receipt. Record the model, prompt, system rules, retrieval sources, software version and uncertainty method used in each study. Log which ambiguity strategy appeared, when the student requested help and which answer or hint eventually resolved the exchange. Preserve privacy by collecting only what the research truly needs, limiting retention and telling students how their data will be used. Without that receipt, a future result may be impossible to interpret. The chatbot could improve during the study, the model provider could change a hidden system behavior or one version could produce more factual errors than another. Educational evidence cannot rest on a moving product with no version label. Ethics education adds another hazard: the model can invent the premise of the debate. If Diffrakt presents a fabricated law, misstates a research result or invents a stakeholder, students may reason beautifully about a case that does not exist. Every exercise needs verified source material and a visible way to separate established facts from the values under discussion. A teacher should be able to inspect and correct the case before students encounter it. The chatbot should not grade the same reasoning that it provoked without independent checks. Otherwise the system gets to write the riddle, judge the answer and congratulate itself on the learning gain. This is why Denmark's experiment matters beyond one university. Schools everywhere are deciding whether educational AI should make work easier, make thinking deeper or attempt both. The market naturally rewards speed, completion and pleasant user feedback. Those are easy to demonstrate in a product video. Durable understanding, intellectual independence and the confidence to challenge a machine take longer to measure. The ITU project points toward a better standard. Ask whether an AI tool leaves a student more capable after the chat window closes. That means measuring delayed learning, transfer to new problems, reasoning quality, student agency, subgroup outcomes, teacher workload and the frequency with which friction became mere obstruction. Publish failures as carefully as successes. State which subjects and students were tested. Do not stretch a result from university AI ethics courses into a claim about primary mathematics, language learning or vocational training. There is a lovely irony here. The project asks students to resist the temptation of an easy answer, and the project itself deserves the same treatment. Adding ambiguity is not automatically thoughtful. A vague bot is still a vague bot. Socratic questioning is not a magic incantation. Friction becomes educational only when it serves a clear objective, respects the learner and ends before curiosity turns into avoidable defeat. Diffrakt may eventually show that a well-designed chatbot can slow students down in exactly the right way. Or it may show that the most valuable feature is the moment a human teacher steps in. Either result would be useful, provided the study measures thinking rather than mistaking difficulty for depth.