Lisa Eriksson

Spurningar og svör frá Sli.Do

What is the best way to find out if a student has used AI to do their work?

The honest – and slightly sad – answer is that there is no reliable way to prove it after the fact. AI detectors are unreliable and produce false positives, often disproportionately for second-language writers. So the better question is: how do I design assessment so I don't need to detect it? That means shifting weight from the final product to the visible process: drafts, version history, in-class work, and tasks that reference things only someone in the room would know.

But the method I still find works best is the simplest one: sit down with the student. Ask them to explain once more how they were thinking. Ask follow-up questions about what they actually wrote or handed in. Two things happen in that conversation. The student realises that you know – without any accusation or detector needed. And, more importantly, the student often realises for themselves that they don't actually know this material. That feeling – sitting across from a teacher and not being able to explain your own work – is a far more powerful deterrent than any policy. It may be exactly what makes the student think twice the next time before letting AI do the learning for them. And if they can explain and defend the work, then the question of who typed it matters much less.

This is the "it's still ours to design" point from the lecture: the conversation is the assessment.

When assessing foreign language learning, how can I help the student to accept that the learning I want to see is the process, not a perfect text with no errors

Students accept this when the assessment actually rewards it, not when we just say it. Some concrete moves: assess the draft-to-draft change, not only the final version; ask students to annotate their own texts ("here I tried a new tense, here I'm unsure"); give credit for risk-taking (attempting complex structures) rather than penalising the errors that come with it; and use AI openly as a comparison tool ("ask the AI for a version, then explain what it changed and why – and which changes you'd keep"). When students see that a flawless AI text earns nothing and a messy text with visible reasoning earns something, the incentive shifts on its own.

But the assessment design only works if the classroom climate allows it. I have also worked deliberately on safety in the group, because in a new language every attempt is a risk. Take a simple activity like the one where a student has a word on a note on her forehead and the others have to explain it without saying the word – what I notice is that the safer the group feels, the more students dare to try from wherever they are linguistically. And the mistakes they make in that kind of exercise are the useful kind: spoken, immediate, correctable, and remembered next time.

Alternating a lot between oral exercises where students have to test the language and get it wrong in front of each other, and written work, makes it easier for them to carry the same mindset into writing – trying, getting it wrong, fixing it – instead of reaching for a perfect text they didn't produce. AI fits here too: students can use it during oral practice as well, to get feed forward and coaching on what to try next, rather than only as a text generator.

What signs in the learning process best show that a student is actually thinking for themselves when they are allowed to use AI as a support tool?

Look for the student pushing back on the tool: rejecting a suggestion, changing a wording, saying "that's not what I meant". Look for specificity – references to their own examples, class discussion, personal experience that the model couldn't have produced. Look for good questions to the AI rather than just "write this". Look for the ability to explain the work in their own words without the screen. And look for inconsistency in a positive sense: a student's own voice is uneven; AI output is smoothly uniform. Prompt logs or a short "how I used AI" note can make this visible.

An example from a physics lesson: students were working on kinetic energy and had built a small motorised cyclist that had to get up slopes of different inclines. One student couldn't quite explain in her own words what was happening and asked the AI for help. She described what they had built, what they had observed and what she wanted explained. The AI answered – and she wrote back that it had misunderstood the scenario, and that she also didn't understand the terms it used in its answer.

That exchange is the evidence. She knew her own experiment well enough to notice the AI got it wrong. She was honest about the limits of her own understanding instead of copying an answer that sounded right. And she kept the AI in the role of a tutor she was questioning, not an author she was outsourcing to. The final explanation she produced would probably still be imperfect – but the thinking behind it was clearly hers.

What would you recommend for children with dyslexia who also tire very quickly when writing, so that they can learn through this approach as well?

This is where AI-supported process assessment is a real accessibility gain rather than a threat. Separate the thinking from the transcription: let them dictate (speech-to-text), record voice notes as their "draft", or explain to the AI orally and have it produce a first written version that they then correct and own. Short, frequent bursts rather than long writing sessions. Assess the reasoning through conversation, mind maps, or STEAM-style products (a model, a diagram, a short video) – your lecture's point about multiple ways to show knowledge applies directly here. Keep the goal fixed (show your thinking) but let the medium be flexible, and make the accommodation ordinary rather than special.

One important caveat, though: this applies to the moments when you are assessing. Living with a diagnosis is hard and can feel unfair, but dyslexia is one of the areas where we actually know that continued reading and writing practice pays off, and where I have seen students make real progress precisely because they kept training. That training doesn't have to happen in every task and all the time – but it has to happen somewhere. Removing everything that is hard about dyslexia would not serve the student in the long run. That's my own view, based on my experience and the research I have read: use the tools to let the student show their thinking when it counts, and protect the practice at other times.

How can we make sure the process itself — thinking and reflection — is given enough time in students' learning?

Time for reflection has to be scheduled, not hoped for. Build it into the task structure: a mandatory pause between draft and revision, a short written or spoken reflection as a required part of the submission, "think first, then tool" sequences where students write or sketch their own idea before opening the AI. Reduce the number of products so each one gets more process time – three assignments done deeply beat eight done quickly. And model it yourself: if the teacher visibly pauses, revises and thinks aloud, students learn that this is what the work is.

But there is a second part to this. Reflection is something students have to be taught, just like the multiplication tables or the alphabet. It is a skill, not an attitude we can simply ask for. A group that has learned what reflection actually is, has practised different reflection models and does it as a routine will reflect in a fraction of the time it takes a group that does it sporadically and has never been shown how. So the time question is partly a training question: the more systematically you teach it, the cheaper it becomes.

And, as with the language learning in Q2, safety matters here too. Reflecting honestly about what you didn't understand or where you went wrong is a risk. In a group where students feel safe it goes quickly; in a group where they don't, it takes much longer – or doesn't really happen at all, however much time you give it.