Q&A

I used to rely on essays written at home for course grades, but now with AI provided even by the university (and knowing that AI detectors do not work) this does not work any more. How should I redesign my assessment?

AI literacy is harder than we thought. Research published this week (Ahmed et al., 2026) from Stanford & Yale shows that LLMs can extract near-verbatim copyrighted texts. Claude 3.7 Sonnet achieved 95.8% recall for some books. Gemini and Grok extracted over 70% of Harry Potter, without jailbreaking. This phenomenon is known as memorization: the encoding of specific training data in a model’s weights such that it can later be extracted in outputs. Why this matters for education. When students ask an LLM to “explain photosynthesis” or “summarize the themes in 1984,” they assume synthesis across many sources. In most cases, that's true. The problem is that they cannot tell when it is not. Research on memorization shows that, under some conditions, LLMs can reproduce long, near-verbatim passages from specific copyrighted texts. There is no signal indicating whether a response is synthesized or recalled. Under these conditions of opaque generation, attribution-based academic integrity frameworks become difficult to apply meaningfully. Consider this scenario. A student genuinely trying to learn uses AI for help. The AI produces an argument that is largely verbatim from a book the student has never seen. The student internalizes it and later writes an assignment in their own words. The work may be original in expression, but its intellectual provenance is unknowable. Who plagiarized? The student didn't know. The AI can't “know”. Intent, visibility, and traceability have collapsed. This is what students need to understand: LLMs may reproduce specific source material rather than synthesize across sources in some cases You cannot reliably tell whether output is memorized or transformed Attribution becomes structurally impossible when sources are hidden AI obscures the learner’s ability to distinguish synthesis from reproduction, challenging traditional academic integrity frameworks. “Teach them to cite AI” can't be the solution. So what helps? There are no foolproof answers, but one response is to make research accountability explicit. That means teaching the slow, often invisible, work of evidence building so students can verify and defend where their ideas come from. In practice, this means introducing traceability at the level of student process: Annotated bibliographies Research logs documenting search strategies and decisions Evidence tables mapping claims to sources Explicit documentation of rejected sources Treating process documentation as seriously as the final product Follow-up oral presentations and peer questioning This is not a complete solution. Determined students can still game process documentation. But it shifts assessment toward intellectual work we can actually verify: source evaluation, evidence quality, the evolution of thinking across drafts, and the ability to defend choices under questioning. Read more: Ahmed, A., Cooper, A. F., Koyejo, S., & Liang, P. (2026). Extracting books from production language models. arXiv preprint arXiv:2601.02671. Reisner, A. (2026, January 9). AI’s memorization crisis. The Atlantic. https://www.theatlantic.com/technology/2026/01/ai-memorization-research/685552/ This post was first published on LinkedIn by Charlotte von Essen in February 2026 and is reposted here with permission as an example for possible content on this blog. https://www.linkedin.com/posts/charlottevonessen_ai-literacy-is-harder-than-we-thought-research-activity-7416120957548871680-gw3o

AI is not socratic. We often hear AI described as a "socratic tutor" but there are fundamental differences between how AI supports learning and what Socrates actually did (well, at least what we know from Plato’s Dialogues). The Socratic method is not asking generic questions instead of giving answers. It's about a teacher who asks deliberate questions because they know where the conversation is going. They know the intellectual territory. And they have a pedagogical goal. Socrates knew what misconceptions his students held and could anticipate their likely responses. Each question was then strategically designed to guide the student toward a specific realization. It was carefully orchestrated discovery. Plato portrays Socrates' teaching method as a form of intellectual "midwifery," which helped students give birth to their own ideas. AI is not directional in this way. It can't anticipate specific misconceptions or craft a sequence of questions tailored to dismantle them systematically. This matters because effective teaching often requires: Deep understanding of common misconceptions in a domain A clear diagnosis of where a particular student is stuck A strategic sequence of questions designed to bridge that gap The ability to improvise when student responses take an unexpected path (while still knowing how to redirect toward the learning goal) Socratic teaching requires a level of intentionality and strategic foresight that's fundamentally different from pattern-matching responses. This post was first published on LinkedIn by Charlotte von Essen in December 2025 and is reposted here with permission as an example for possible content on this blog: https://www.linkedin.com/feed/update/urn:li:activity:7396435289897844737/

We would love to be able to tell our students exactly what they can and cannot do regarding using AI, and we know that our students also really want clear guidelines so they do not accidentally cheat. But, as Corbin et al. (2025) write, “The more detailed and specific our instructions become about ‘acceptable’ AI use, the more we highlight the gap between what we can specify and what we can verify”. When we try to code everything a student could do in relation to learning into the traffic light of red – forbidden, green – ok, and yellow – it depends, we run into several problems: The underlying assumption — that students will understand what exactly is permitted and what is not — is not valid. The lines between, for example, drafting and editing, are blurry: what if a suggested edit brings out an argument that we hadn’t thought about explicitly before but now we think we should stress it more? Where is the boundary between having a question explained and getting hints for how to approach it? What counts as “refining” an argument: correcting grammar but not alterning meaning, finding errors but not helping towards a solution? Students need to voluntarily follow the rules, even when there are disagreements (or misunderstanding, as above) between teachers and students about what is legitimate use of GenAI for learning or in assessment. Compliance cannot be enforced because there is no way to reliably detect GenAI use (other than leaving in parts of a response like “Sure, here is the response to the question …”, the teacher directly observing a student using GenAI, or the student admitting to it). So while the traffic light analogy seemed appealing at first for how easily it signals to students what’s ok and what is not, traffic lights in real traffic are substantially different. They exist in much larger structures with speed cameras, police patrols, fines, etc, so rule violations can actually be detected and compliance enforced. To deal with the challenge, Corbin et al. (2025) suggest a “two-lane approach”, where lane 1 is controlled, in-person assessment, and lane 2 is open in the sense that there is no attempt to control how students use GenAI. But lane 2 can of course be structured in a way to discourage misuse, like having someone sign off on data produced in a lab so that students have to continue with that data set rather than generating synthetic data; or checkpoints where students discuss their work and then asynchronously build on the ideas discussed with the instructor. Corbin et al. (2025) conclude that “The path forward through this increasingly challenging terrain [of increasing AI capabilities] lies not in more sophisticated rules about AI use, but in fundamentally redesigning how we structure assessments to demonstrate student capability. This will require significant effort and creativity from educators but has the advantage of allowing for genuine solutions to maintaining assessment validity in an AI-enabled world. These must be solutions based not on what we say, but on what we do.” So – how do we design cheat-proof assessment? There are of course attempts like going back to pen-and-paper (but can we detect AI glasses? And is writing by hand under time pressure really an inclusive and accessible way to assess learning?) or oral assessments (what about test anxiety or teacher biases?). But it is a slippery slope where we might slide from teaching into policing. Read more: Corbin, T., Dawson, P., & Liu, D. (2025). Talk is cheap: why structural assessment changes are needed for a time of GenAI. Assessment & Evaluation in Higher Education, 50(7), 1087–1097. https://doi.org/10.1080/02602938.2025.2503964 An earlier version of this post was first published by Mirjam Glessmer on https://mirjamglessmer.com/2026/04/13/currently-reading-corbin-et-al-2025-on-talk-is-cheap-why-structural-assessment-changes-are-needed-for-a-time-of-genai/ and is reposted here with permission as an example for possible content on this blog.

The argument that “AI is here to stay” and therefore we have to use it is not a valid one. The future is always undetermined and we do have influence on how it turns out. And that includes whether AI is staying -- or not. So what if the critical AI skill of our era, as Karen Costa puts it, was not how to use it, but how to resist it? There are five main reasons for AI resistance, according to Şimşek and Yasar (2025): Socio-economic concerns, like AI taking people’s jobs, Ethical issues, since AI systems are opaque and biased, Safety risks, when AI gets to make decisions, Threats to democracy and sovereignty, “including the use of AI for large-scale societal manipulation“, and Environmental impact. All of these are serious and worth exploring in detail. So if we wanted to resist AI, for these reasons or others, fully or just inducing enough friction in the adoption to make sure we don’t move faster than we have the chance to safely assess what we are doing, how would we do that? Drimmer & Nygren (2025) suggest four small acts of friction: Centering students in our teaching. Not AI, not preventing cheat with AI, just students. They write that “[d]efensive maneuvers, like in-class essay writing exclusively, are acts of deprivation. They deprive students of the opportunity to reflect and refine away from the pressures of the classroom clock.” Instead, we should keep writing assignments and not police how they were generated, but rather make them thought-provoking and give thoughtful feedback on the thoughts reported in them. Not optimizing everything. Making space for things that are not on the curriculum and that are not credited: “Reading groups, lightning round presentations, unambitious programs of being in simple, un-CV-able conversations” to focus back on humanity and community, and make space for skepticism. Taking some things offline again: Sharing printouts, using leaflets, not making everything be online all the time (but careful that we do not roll back all efforts on inclusion that have been made thanks to assistive technologies!). Ask questions — in person, in established relationships — about what problem we are really trying to fix by adopting AI, and whether we are really working on the right problem. What do you think? Is resisting AI the real AI skill of the era? If you are resisting or introducing friction, how are you doing it? --- Read more: Drimmer & Nygren (2025). Four Frictions: or, How to Resist AI in Education. https://www.publicbooks.org/four-frictions-or-how-to-resist-ai-in-education/ Şimşek and Yasar (2025). From Rejection to Regulation: Mapping the Landscape of AI Resistance. Available here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5287068 --- An earlier version of this post was first published by Mirjam Glessmer on https://mirjamglessmer.com/2026/03/11/thinking-about-ai-resistance/ and is reposted here with permission as an example for possible content on this blog.

My student submitted a text and I am 99% sure that it is written by AI. It uses all those words that are all over AI-generated texts, like “thus”, and “leverage”, and I just feel it in my gut. So I uploaded the student’s text to ChatGPT and asked “did you write this?” and it confirmed that it had written it. I even uploaded it to one of those AI checkers, and it gave me a score of 83% AI generated. I confronted the student, but they deny having used ChatGPT or any other AI at all. What can I do? I don’t want to escalate this if I am not 100% sure.

A master student wants to use AI to generate an image in their thesis. The publisher of the original image does not respond to attempts to contact them, and the student really wants to use this image and says they need it. The image is a photograph of some structure they are discussing in the thesis through a microscope. I do not think generating and using an AI image is right thing to do. What should I do?