Rico Eberle – Learning Nuggets

Learning styles are a myth. Here's what works.

Thu, 01 Oct 2026 00:00:00 GMT

This post comes with a video (1:09 minutes) and transcript: watch it in the post.

“What’s your learning style?” The idea behind this question is widespread among educators, including in adult education. Visual, auditory or kinaesthetic, and if you’re taught in your style, you learn better. The idea is intuitive, widespread and well meant. It just isn’t supported by evidence.

What the research says

In 2008, Harold Pashler, Mark McDaniel, Doug Rohrer and Robert Bjork reviewed the evidence for the Association for Psychological Science. Their question was precise: do people learn better when teaching matches their preferred learning style?

Preferring a format does not mean you learn better with it.

Testing this properly requires a specific design: learners with different styles are randomly assigned to different teaching methods, and everyone then takes the same test. The researchers found only a few such studies. Those studies didn’t show the expected effect, and several flatly contradicted it. Their conclusion: there is no adequate evidence for tailoring teaching to learning styles.

A more recent meta-analysis from 2024 finds small effects under certain conditions. That changes little in practice: the effort spent on learning style tests and tailored materials is out of proportion to any possible benefit.

Why the myth persists

Most educators still believe in it. In 2020, Philip Newton and Atharva Salvi analysed 37 studies covering more than 15,000 teachers and trainers in 18 countries. Weighted, around 89 percent said they believe in matching instruction to learning styles, roughly nine out of ten.

No evidence for learning styles, yet around 89 percent of teachers believe in them.

One reason: the idea contains a grain of truth. We all have preferences. Some people prefer reading, others prefer listening. But a preference doesn’t mean we learn better that way. Rogowsky and colleagues tested exactly this in 2015: adults who described themselves as readers or listeners did not learn better with their matching format.

What works instead

Match the format to the content, not the person. A process needs a diagram, pronunciation needs sound, a hands-on skill needs practice. That choice is the same for every participant.

The format follows the content, not a learning style.

Active recall and spaced repetition. Both are among the best-supported learning techniques there are. More on this in the post on the forgetting curve.

My suggestion for your next course

Drop the learning style test at the start. Use the time for a short quiz on what participants already know. It tells you more about the group than any typology, and it doubles as the first recall exercise.

And when you plan materials, don’t ask “What do participants like?” Ask “What does this content need?”

Mix, don't block: three times better on the test

Thu, 01 Oct 2026 00:00:00 GMT

This post comes with a video (1:17 minutes) and transcript: watch it in the post.

Two groups practise for the same amount of time, with the same problems. A week later, one group scores three times higher than the other. The only difference is the order of the exercises. The result runs against what feels right while practising.

The experiment

In 2007, Doug Rohrer and Kelli Taylor had university students learn how to calculate the volume of different solids. Everyone got the same problems and the same amount of time. Only the order differed:

Blocked: first all problems for the first solid, then all for the second, and so on.
Mixed: the same problems, but shuffled.

One week later, the mixed group scored 63 percent, well ahead of the blocked group at 20 percent.

A week later, students in the mixed group solved 63 percent of the test problems correctly, students in the blocked group only 20 percent.

The twist

During practice it was the other way round. The blocked group was clearly ahead, with 89 percent correct compared with 60 percent in the mixed group. When you practise in blocks, you apply the same procedure over and over and quickly feel confident. When you practise mixed, you first have to work out which procedure fits each problem. That takes more effort, and it is exactly the skill the test requires later.

During practice, the blocked group was ahead, 89 to 60 percent.

Mixing feels harder, but it lasts longer.

Adults misjudge this too

In 2008, Nate Kornell and Robert Bjork had adults learn the painting styles of different artists, once blocked and once mixed. Most learned more from the mixed version. Even so, 78 percent rated blocking as equally good or better, a little over three quarters. How practice feels is a poor guide to what sticks.

78 percent rated blocking at least as good, even though they learned more from mixing. More precisely than the slide: according to the follow-up study, older adults benefit much like younger ones.

A follow-up study by Kornell, Castel, Eich and Bjork in 2010 also showed that older adults benefit from mixing much like younger ones. Good news for continuing education.

What this means for continuing education

Mix your exercises. This works especially well for problem types that look similar and are easily confused. In a bookkeeping course, for example, mix journal entries from different types of transactions instead of going chapter by chapter. That example is my own transfer; the studies come from mathematics and art.

Explain why it feels harder. Mixed practice leads to more mistakes at first. Without an explanation, that looks like a bad course. With one, it becomes a useful signal: this is what learning feels like.

My suggestion

Take your next worksheet and check whether the problems are sorted by topic. If they are, shuffle the second half. The first half can stay blocked so participants get to know each procedure. Then comes telling them apart.

Hard doesn’t mean bad.

Model, LLM, chat, agent: four terms, one picture

Thu, 01 Oct 2026 00:00:00 GMT

This post comes with a video (1:33 minutes) and transcript: watch it in the post.

Model, LLM, chat, agent: these four terms come up in almost every conversation about AI, and two people using them often don’t mean the same thing. Yet they fit together quite simply, like layers, one around the other. The picture is a teaching simplification, but it helps you place tools and ask the right questions.

At the core: the model

At the very centre is the model. A model is the result of training: a huge file with billions of numbers, known as parameters. An LLM, a large language model, is one kind of model. It was trained on a vast amount of text and predicts, piece by piece, how a text continues. GPT, Claude, Gemini or the Swiss model Apertus are such language models.

The model sits at the core. An LLM is one kind of model, not a separate layer.

Apertus shows how large such models are: EPFL, ETH Zurich and the Swiss National Supercomputing Centre CSCS released it in 2025 in two sizes, with 8 and 70 billion parameters.

The “pieces” mentioned in the video are called tokens: words, parts of words or punctuation marks. The video deliberately says “piece” rather than “word”, because a model doesn’t work word by word. Why this prediction can sound fluent and still be wrong is the subject of the post AI doesn’t lie. It guesses.

The first shell: the chat

On its own, a model does nothing. It needs a shell through which you can use it. The first shell is the chat: an input field, a history, maybe a few files. You ask, the model answers.

A well-known name makes the difference tangible: ChatGPT is the chat, and the model inside it is called GPT. So when a colleague says “ChatGPT” got something wrong, it’s worth asking whether they mean the model’s answer or a feature of the interface.

The second shell: the agent

The agent gives the model a goal and tools: searching the web, opening files, running programs. And it lets the model work in a loop: plan, act, check, continue, until the task is done.

The agent gives the model a goal and tools and lets it work in a loop.

In a 2024 guide, Anthropic describes agents as systems in which the language model directs its own process and tool use. The guide distinguishes them from workflows, where predefined code decides when the model uses which tool. It also names the downside: agents cost more, and errors can compound over several steps. It therefore recommends extensive testing in sandboxed environments and appropriate guardrails.

The lines are blurring

Many chats can now search the web. Then they already work a little like an agent: the model decides for itself whether and what to search for, and works the results into its answer. The picture of layers is therefore not a sharp division but a way to get your bearings. The further out, the more the AI does on its own.

What this means for teaching

The line to remember from the video sums up the picture: the model computes, the chat answers, the agent acts.

The model computes, the chat answers, the agent acts.

For course leaders, this leads to two practical points:

Say exactly what you mean. If you set rules for AI in your course, make clear whether you mean a chat or also tools that search, edit files or run programs on their own.

Look more closely the more the AI does on its own. With a chat answer, you check a text. With an agent, you also check which steps it took, which sources it opened and what it changed. That’s my own view; it fits Anthropic’s warning about compounding errors and the need for testing and guardrails.

How an AI assistant works with your own documents, and what the language model really sees, is explained in RAG explained: your AI never read the handbook.

My suggestion

In your next training session, draw three nested circles on a flip chart: model, chat, agent. Ask participants to write the AI tools they know or use on sticky notes and place each one in a circle. The borderline cases are the interesting ones, such as a chat with web search. Ask there: what does the tool do on its own, and what do I need to check afterwards?

If you keep the layers apart, you know where to look more closely.

RAG explained: your AI never read the handbook

Thu, 01 Oct 2026 00:00:00 GMT

This post comes with a video (1:44 minutes) and transcript: watch it in the post.

Many AI assistants that work with your own documents have never read the whole handbook. They only see a few snippets of it, picked by a search. The method is called RAG. Once you understand it, you can better judge when such assistants answer reliably and when they don’t.

How RAG works

RAG stands for retrieval augmented generation: text generation supported by a search. Patrick Lewis and colleagues described the method in 2020. It starts with preparation: all documents are cut into small snippets. Each snippet gets a numeric code for its meaning and goes into a database. Every question then goes through three steps:

Translate the question: your question is also turned into such a numeric code.
Search: the database looks for the snippets with the most similar code, usually just a handful.
Answer: only now does the language model come in. It gets your question and these few snippets and writes the answer from them.

Preparation plus three steps per question: the language model only comes in at the end.

The video simplifies in a few places. In technical terms, the snippets are called chunks, the numeric code an embedding and the database a vector database. The embedding is also produced by a neural network, a so-called embedding model, but it isn’t a language model that writes answers. Some systems also use a language model to rephrase the question or to re-rank the snippets found. And how many snippets the model gets depends on the system.

What the language model sees

The language model never sees the whole handbook, only what the search picked beforehand. If the key snippet is missing, the model has nothing to stand on. Then it guesses, or says it can’t find anything. And connections across chapters easily get lost, because each snippet stands on its own.

Anthropic described the problem in 2024: traditional RAG systems remove context when they split documents, which often means they fail to retrieve the relevant information. A sentence about revenue growth doesn’t say on its own which company or which quarter it refers to. In Anthropic’s tests, adding context to each snippet reduced such failures considerably, but didn’t eliminate them.

The alternative: everything in the context window

The context window is everything a model can take into account when answering, a kind of working memory. According to Anthropic, current Claude models hold a million tokens. Tokens are the pieces a language model breaks text into: words, parts of words or punctuation marks. How many pages that is depends on the text. Anthropic equates 200,000 tokens with about 500 pages; scaled up, a million tokens comes to around 2,500 pages. That’s a rough guide, not an exact conversion.

If you put everything in directly, the model really has it all in front of it. Studies show that this is often better. In 2024, Zhuowan Li and colleagues compared both approaches on several public datasets: given enough resources, long-context models performed better on average than RAG. Anthropic, too, recommends simply including the entire knowledge base in the prompt if it is smaller than 200,000 tokens.

The downside of a large window

The large window has three limits:

It only scales up to the limit of the window. A whole library won’t fit.
Every question costs more and takes longer. Billing is per input token, and all the material goes in with every question. Prompt caching lowers the cost, but not to the level of RAG.
The fuller the window, the more likely the model misses details. In 2023, Nelson Liu and colleagues showed that models make worse use of information in the middle of a long text than at the beginning or end (“lost in the middle”). In 2025, Chroma tested eighteen models and found that performance drops as input grows, even on simple tasks. Anthropic’s own documentation calls this “context rot”.

Both approaches have strengths: RAG is cheap and fast, the full window runs into limits.

RAG, on the other hand, has real advantages: it is cheap and fast. And because it is known which snippets were used, it can show where an answer comes from. Lewis and colleagues already named the lack of provenance as a weakness of language models without retrieval.

The rule of thumb

If it fits in the window, put it all in. If it’s a whole library, you need RAG. Or an agent that searches on its own.

The rule of thumb from the video: the window, RAG or a searching agent.

An agent doesn’t pick snippets in advance. It searches with tools itself, checks what it finds and keeps searching if needed. Boris Cherny from the team behind Claude Code wrote in 2026 that early versions of the tool used RAG with a local vector database, but that agentic search generally works better and is simpler.

What this means for continuing education

If you use AI assistants for your own material, such as course dossiers or regulations, there are three things to take away:

Check answers against the source. If an assistant shows which document an answer comes from, it’s worth taking a look. That is one of RAG’s strengths.

“Nothing found” doesn’t mean “doesn’t exist”. The search may simply have missed the right snippet. Try asking again with different words.

Be careful with questions that span several chapters. Comparisons, summaries of entire documents or contradictions between sections easily get lost with RAG. If the document fits in the window, it is often better to provide all of it.

What exactly separates a model, a language model, a chat and an agent is explained in Model, LLM, chat, agent: four terms, one picture.

My suggestion

Test the AI assistant that you or your participants use for your own documents with three questions about a document you know well: a detail question answered in a single section, a question that links two chapters, and a question whose answer isn’t in the document at all. Compare the answers with the original. You’ll quickly see where the assistant is reliable, and you can go through the results with your participants.

If you know what the model sees, you also know what it can’t know.

Three time sinks you can hand over to AI

Tue, 29 Sep 2026 00:00:00 GMT

This post comes with a video (1:00 minutes) and transcript: watch it in the post.

When people talk about AI in schools, the conversation often turns to exams and homework. Less often does it turn to the question many teachers face on a Sunday evening: how do I get through all this preparation? On that question, there is now a solid measurement.

What the trial measured

In 2024, the Education Endowment Foundation (EEF) ran a trial in England with 259 teachers from 68 schools, evaluated by the research institute NFER. Schools were randomly assigned to two groups: one used ChatGPT for lesson preparation, the other prepared without AI.

With ChatGPT, teachers spent 31 percent less time on lesson preparation, with no drop in quality.

The result: the ChatGPT group spent an average of 56.2 instead of 81.5 minutes a week, a little over 25 minutes or 31 percent less. An expert panel rated the materials from both groups without knowing where they came from. It found no difference in quality.

What the trial doesn’t tell you

The trial looked at secondary school science (Years 7 and 8 in the English system), and the teachers had a guide on how to use the tool. Whether the figure carries over to other subjects, other age groups or use without guidance is not something the trial answers.

The three time sinks below don’t come from the trial either. They are my recommendation for where the same idea applies: the AI delivers a draft, and the teacher turns it into something usable. Course leaders in continuing education can apply this too, for example to course announcements, messages to participants or rubrics for course assignments. The trial did not measure that.

Three time sinks to hand over

Letters to parents and announcements. The AI writes the first draft, you adjust tone and content.
Differentiation. The same task at three levels of difficulty, or a text in plain language.
Grading rubrics. The AI suggests criteria and levels, you sharpen them.

Three tasks where the AI can take on the first draft.

The boundary stays clear

The AI delivers drafts, you make the decisions. That holds for all three tasks: whatever goes out is your responsibility. And personal data stays out. The public schools of the Canton of Zurich make the same point in their guidance on AI. My tip: draft a letter to parents with placeholders such as “[name]” or “[date]” and fill them in only at the very end.

The AI delivers drafts, you make the decisions, and personal data stays out.

My suggestion to try

Pick a single time sink for next week, not all three. For differentiation, for example: give the AI a task from your current teaching, without any details about your class, and ask for three versions, an easier one, the original and a more demanding one. Then check: is the content correct? Does the level suit your learners? What would you change?

Make a rough note of how long it would have taken you and how long it took with a draft. After a week you will know from your own experience whether it pays off for you, rather than relying on an average from England.

Time saved is time for relationships and good feedback.

Making assignments AI-robust

Sat, 26 Sep 2026 00:00:00 GMT

This post comes with a video (1:04 minutes) and transcript: watch it in the post.

“Summarise the text and answer the questions.” An AI can finish homework like this in a few seconds. The obvious reflex is to ban AI. That’s understandable, but it achieves little: what happens at home can hardly be controlled. A different question is more useful: how does an assignment need to be built so that the thinking still stays with the learners?

The problem is the assignment, not the tool

An assignment that only asks for a finished product ends up mostly checking whether a text was handed in. You can’t tell from the text who wrote it. So AI-robust doesn’t mean locking AI out. It means designing assignments in a way that makes real understanding visible.

Four building blocks for redesigning an assignment to be AI-robust.

Four building blocks

1. Process over product. Assess not only the result, but also drafts, interim versions and notes. The ZHAW (Zurich University of Applied Sciences) states this explicitly in its guideline on AI in assessments: not just the end product can be graded, but also the working phase, in other words a process assessment.

The ZHAW guideline also names conversations and assessing the work process.

2. Local context. Connect the task to your class, your town or an experience from the lesson. What your class discussed last week is something no AI knows. It isn’t watertight: anyone who feeds the AI the context will still get an answer. But to do that, learners have to engage with what they experienced.

3. Oral defence. A short conversation quickly shows whether someone understands their own text. The ZHAW guideline mentions graded oral check-ins and defences. According to Swiss broadcaster SRF, oral formats are gaining weight at Swiss universities.

4. Let learners critique AI output. Learners check an AI answer, find mistakes and improve it. That turns AI from a shortcut into something to learn from.

Building blocks 2 and 4 are my own recommendations from teaching practice. Building blocks 1 and 3 are directly backed by the ZHAW guideline. The examples come from school, but the building blocks work just as well in continuing education courses: the class becomes the course group, and the classroom becomes the training room or the participants’ own workplace.

My suggestion: redesign one assignment

Take the homework from the opening and redesign it in three steps:

Instead of a summary, have an AI write one and give it to your learners with the task: mark what’s missing, what’s wrong and what you would weight differently.
Instead of generic questions, ask for a link to your lessons: where does the text contradict what you discussed in the last lesson?
Instead of just handing it in, have a conversation of a few minutes with some of the learners about their corrections. The notes from steps 1 and 2 count towards the grade.

That puts all four building blocks into a single assignment. You control the effort yourself: you decide how many learners you talk to.

AI-robust doesn’t mean AI-free

None of these building blocks bans AI. It can help with research, with wording, and can even be the subject of the assignment. What matters is that the thinking visibly stays with the learners: in the process, in the link to their own surroundings, in conversation and in critical review.

AI may help, the thinking stays with the learners.

Less on the slide, more room to think

Thu, 24 Sep 2026 00:00:00 GMT

This post comes with a video (1:04 minutes) and transcript: watch it in the post.

You have probably seen slides like the one in the video in training sessions, meetings and courses: a title, six bullet points, 94 words. Well meant, because everything important is on it. The catch: while you speak, your audience reads. And doing both at once works worse than we think.

The bottleneck is working memory

Cognitive load theory goes back to John Sweller. Its core idea: our working memory is tightly limited. Anything we want to understand has to pass through this bottleneck.

Listening and reading at the same time sends two streams of language through the same narrow working memory.

If you listen and read a full slide at the same time, you push two streams of language through it. Mayer and Moreno described in 2003 how exactly this duplication creates an additional, unnecessary load. That costs attention, which is then missing for understanding.

Redundancy and coherence

In 2017, Richard Mayer translated the research into design guidelines for e-learning. Two of them apply directly to slides:

Redundancy principle: people learn better from pictures and spoken words than from pictures, spoken words and the same text on screen. Mayer puts it as a rule: do not add on-screen text to narrated graphics.
Coherence principle: anything that doesn’t serve the learning goal distracts. So leave out extraneous material.

Redundancy and coherence principles: the same text on top of the voice does not help learning.

This doesn’t mean text on slides is forbidden. It means that text which only repeats what you are saying anyway doesn’t help. It gets in the way.

What about the subtitles in this video?

Fair question, since the nuggets also show text alongside the voice. The redundancy principle refers to learning situations with sound, such as slides being read aloud. I still chose to use subtitles: they make a short video understandable without sound, and they help when watching in a foreign language and for people with hearing impairments.

Three things for your next presentation

One message per slide, as a short sentence.
A picture or diagram instead of bullet points.
What you say doesn’t need to be on the slide. The details belong in the handout.

My exercise for you

A suggestion you can try in half an hour: take your last presentation and find the fullest slide. Count the words. Then write the one message that should stick as a single sentence on a new slide. Think of a picture or diagram that shows this message. Everything else goes into your speaker notes or the handout.

Less text on the slide leaves more capacity for understanding.

The next time you give the talk, pay attention to where your audience is looking: at you or at the screen.

Less text doesn’t mean less content, it means more room to think.

The data traffic light: what can go into AI?

Tue, 22 Sep 2026 00:00:00 GMT

This post comes with a video (1:10 minutes) and transcript: watch it in the post.

Writing a student report is quick. Finding the right words takes longer. So it is tempting to paste the text into ChatGPT and ask for better sentences. But that is exactly the problem: a student report contains a child’s personal data, and that does not belong in an AI tool.

In everyday school life, the decision often happens in seconds. That is where a simple traffic light helps. It is my own teaching simplification and does not replace the rules of your canton or your school. For a quick assessment, though, it usually does the job. The examples come from primary and lower secondary school; course leaders in adult and continuing education can apply the traffic light in the same way, for example to participant lists, feedback or exam results.

Green: public and your own material

Worksheets you created yourself, curriculum texts, a poem: content like this contains no personal data and can go into the AI. One side note: with texts written by others, copyright is a separate question. That is a topic of its own and has nothing to do with the traffic light.

Green: your own and public material without personal data can go into the AI.

Yellow: internal, but no link to individuals

A letter to parents without names or anonymised feedback is possible, but only if nothing can be traced back to individual children. That is harder than it sounds. Even without names, the class, the place or a particular event can reveal who it is about. The Zurich school authority explicitly mentions references to names and locations, too. The Data Protection Commissioner of the Canton of Zurich recommends anonymising data, ideally before a provider can see it.

Yellow: possible, but only if nothing points to individual children any more.

Red: personal data of learners

Names, grades, reports, and above all sensitive data, for example about health, do not belong in the AI. Under Zurich data protection law, health data counts as a special category of personal data. The Data Protection Commissioner also makes two points: personal data stays worth protecting even when it is public, and responsibility stays with the school, not with the AI provider.

Red: learners' personal data does not belong in the AI.

Cantonal rules: different, but pointing the same way

The Zurich school authority is clear: no personal data of students, above all no special categories of personal data. Other cantons handle the details differently. The guidance from the Canton of Solothurn, for example, addresses parental consent for children under 16. The direction is the same everywhere. If your school uses its own AI solution with a vetted contract, different rules may apply. When in doubt, ask your school leadership. The school authorities’ rules apply to compulsory schooling; for continuing education they are transferable, but the rules of your own institution are what count there.

My suggestion for the student report

Instead of pasting the report into the AI, ask it for general phrasing help with no link to the specific child. For example: “Give me five appreciative ways to describe progress in reading.” You then put the fitting sentences into your report yourself. The report stays with you, and the AI still helps.

And a small exercise: take three documents you worked on this week and assign each one a colour. Yellow deserves a second look: would a colleague at your school recognise who it is about? Then it is red.

The rule of thumb to take away: if someone could tell who it is about, it doesn’t belong in the AI. The video puts it more briefly, “what you wouldn’t post publicly”; that means content linked to individuals (red), not internal content with no link to individuals (yellow).

Online is not the same as online

Sat, 19 Sep 2026 00:00:00 GMT

This post comes with a video (1:32 minutes) and transcript: watch it in the post.

“This course is also available online.” It sounds like a clear answer, yet it tells you surprisingly little. It might mean a live lesson in a video call, a classroom course backed by a digital learning platform, or a learning path you work through entirely on your own. All three carry the same label. For learners, they feel completely different.

Why “classroom or online?” falls short

The usual question only looks at location. But location is not what defines a learning offer from a teaching perspective. Teaching with fixed times is still a timetable, even in a video call. Learners save the commute, but the school still decides when and at what pace they learn.

Three very different offerings, all with the same “online” label.

Three dimensions instead of one

That is why I describe learning offers along three dimensions:

Flexibility: Who decides place, time and pace?
Guidance and commitment: How much does the school structure and accompany learning?
Self-regulation: How much planning and stamina does the model demand from learners?

Together they give four types: the online classroom, the school with a digital learning ecosystem, the digital self-study model, and the modular system that offers several formats side by side. The typology is a conceptual model for commercial continuing education, not a rating of individual schools.

These dimensions did not come out of nowhere. The Community of Inquiry model by Garrison, Anderson and Archer (2000) describes, alongside social and cognitive presence, a teaching presence in online learning: the way teachers design and guide the learning process. That is close to my second dimension.

Flexibility comes at a price

More flexibility shifts responsibility to the learners. If you choose your own place, time and pace, you also have to plan, keep going and track your own progress. A review by Broadbent and Poon (2015) found that such self-regulation strategies, time management among them, are linked to achievement in online learning.

The more flexible the format, the more responsibility shifts to the learners.

So the most flexible model is not automatically the best. What matters is the fit with the target group.

An exercise for your team

Here is my suggestion for your next team meeting: take one of your own courses and rate it on the three dimensions, from “low” to “high”. Then ask three questions:

Is the description accurate? Would someone from the target group place the course the same way based on the brochure?
Does the demand for self-regulation match the target group? People learning alongside work and family may want flexibility, but also clear milestones.
Where is support missing? If you offer a lot of flexibility, what do you give learners so they can handle it?

The better question: which balance fits which target group?

The answers usually lead to the real question: do we need more flexible offers, or better support for handling flexibility? Not classroom or online, but the balance of flexibility, guidance and self-responsibility that fits your learners.

Learning isn't an event. It's a rhythm.

Thu, 17 Sep 2026 00:00:00 GMT

This post comes with a video (1:12 minutes) and transcript: watch it in the post.

When we plan a course, we tend to focus on the moment of teaching: good slides, a clear explanation, a fitting exercise. What happens to the material afterwards is often left to chance. Yet it is exactly this afterwards that decides how much sticks. Hermann Ebbinghaus showed this as early as 1885.

A self-experiment with nonsense syllables

Ebbinghaus was both experimenter and subject. He memorised lists of nonsense syllables and tested how much remained after intervals ranging from 20 minutes to 31 days. His measure was the savings method: how much less effort does it take to learn a list again?

The result is the forgetting curve. It drops steeply at first and then levels off. In 2015, Murre and Dros repeated the experiment, again with a single subject, over roughly 70 hours. Their results were similar to Ebbinghaus’ original data.

What the curve doesn’t tell you

Be careful with precise percentages that circulate online, such as how much you supposedly forget after a week. Ebbinghaus’ figures apply to nonsense syllables in a self-experiment and don’t transfer neatly to course content. That’s why the video shows the curve as a schematic, without measured values.

Schematic: meaningful material is retained better than nonsense syllables.

Meaningful material is retained better. Ebbinghaus himself found that stanzas from Byron’s “Don Juan” needed only a fraction of the repetitions. But the principle holds: without repetition, meaningful material fades too.

Repetition changes the curve

The good news comes from Ebbinghaus as well. When he relearned the same list on several days, his savings grew. Put simply: every repetition pulls the curve back up, and afterwards it falls more gently than before.

Each repetition pulls the curve back up, and afterwards it falls more gently.

Current learning research points the same way. In 2013, Dunlosky and colleagues rated common learning techniques by how useful they are. Distributed practice and practice testing came out on top, while rereading was rated low. As Dunlosky puts it, almost any kind of practice test helps.

Three things for your courses

Plan repetition, spaced over days and weeks, instead of cramming everything into one day.
Test instead of rereading. Actively recalling something sticks better than going over the text again.
Build in short moments of recall, for example a quiz at the start of the next lesson.

My suggestion for your next lesson

Open the lesson with three to five short questions: two on the previous lesson, one on a topic from a few weeks back. No grades, and the answers follow straight away. It takes a few minutes and combines exactly the two techniques that scored best in Dunlosky’s review: testing and spacing.

Repetition works as a rhythm, not as a one-off event.

And an exercise for you as the course designer: take your current course plan and mark, for every key topic, when it comes up a second time after it was first taught. Wherever there is no mark, the course is quietly planning for forgetting.

Learning is not a one-time event. It’s a rhythm.

AI detectors are not proof

Tue, 15 Sep 2026 00:00:00 GMT

This post comes with a video (1:08 minutes) and transcript: watch it in the post.

Ever since language models started writing entire essays, schools have been looking for a tool that settles the question at the push of a button: was this written by AI? Detector vendors promise exactly that, often backed by impressive percentages. The trouble is not only that those numbers rarely hold up in practice. It starts with the maths itself.

One percent sounds small until you scale it

Say a detector is right 99 percent of the time. Your school runs 1000 essays through it, and every single one was written honestly. A one percent false alarm rate then means ten students fall under suspicion, even though they did nothing wrong.

Even at 99 percent accuracy, ten out of 1000 honest essays come under suspicion.

The key point: false alarms grow with volume. The more texts you check, the more innocent students get flagged, even if the percentage stays the same. Each case is a difficult conversation with trust at stake. And 99 percent is a deliberately optimistic figure.

What the research shows

Real-world results are considerably worse. In 2023, a research team led by Weber-Wulff tested fourteen detection tools. Their verdict: the tools are neither accurate nor reliable.

Three research findings on AI detectors.

A second study, by Liang and colleagues, points to a fairness problem. English texts by people writing in a language that isn’t their first are flagged as AI especially often. The study looked at English essays, including TOEFL essays; it says nothing about other languages. The researchers suggest that detectors may unintentionally penalise writers with more constrained linguistic expression, and they explicitly caution against using these tools in evaluative or educational settings.

Even OpenAI, the company behind ChatGPT, shut down its own detector in July 2023, citing its low rate of accuracy.

What works better

A detector result is not proof. At most, it is a reason to look more closely. In everyday teaching, these three recommendations from the video go further:

Make the writing process visible. Drafts, notes and intermediate versions become part of the submission. Students who developed a text themselves can show how they got there.
Talk with the student about the text. Not as an interrogation, but as part of the assessment: Why this structure? What was hard?
Design tasks that require their own thinking. For example, by connecting them to your own lessons, personal experience or a decision that has to be justified.

Three approaches that achieve more than a detector score.

An exercise for your staff meeting

My suggestion: estimate together how many written assignments are handed in at your school each semester. Then apply a one percent false alarm rate, the optimistic figure. The number you get is how many conversations you would have to have with honest students. After that, ask a better question: which tasks in our own teaching would make a detector unnecessary?

Trust doesn’t come from software. It comes from good tasks.

AI doesn't lie. It guesses.

Sat, 12 Sep 2026 00:00:00 GMT

This post comes with a video (1:18 minutes) and transcript: watch it in the post.

Anyone who works with AI tools knows the moment: the answer is fluent, well structured and sounds competent. And yet it is wrong. A source that does not exist. A year that is slightly off. A quote nobody ever said. This is called hallucination. The word suggests a malfunction. In fact, it follows directly from how language models work.

A language model knows probabilities, not truth

A language model does not look things up. Put simply, it calculates which word is most likely to come next, word by word. Because it was trained on enormous amounts of text, the most likely word is very often the right one.

The most likely word is not always the right one: Sydney instead of Canberra.

When the model does not know something, that process does not change. It still picks the most likely next word. The result is a sentence that sounds just as confident as a correct answer. Hesitation, a “maybe” or a hint of uncertainty does not appear on its own.

Why the AI doesn’t simply say “I don’t know”

In 2025, OpenAI published an explanation, together with a research paper. The core idea: language models are scored much like students on a multiple-choice test. If you guess, you sometimes get points. If you leave it blank, you certainly get none.

As in a multiple-choice exam, guessing earns more points on average than staying silent.

When scoring rewards guessing and penalises restraint, a model learns exactly that: guessing pays off. Hallucination is therefore also a result of how models are measured and improved.

What this means for teaching

The key insight for learners: a confident tone says nothing about accuracy. An AI answer is a draft, not a reference work.

Three checks for every AI answer.

Three questions help in everyday use:

Can I find this in a second, independent source? Not in a second AI, but in a source that is itself backed up.
Can names, numbers and quotes be verified? These details are especially vulnerable because they have to sound plausible but should be exact.
What happens if I ask the same question differently? If the answer changes with a slightly different wording, that is a warning sign.

An exercise for your next lesson

Have learners work in groups and ask an AI about a topic they already know well. Each group uses the three questions to find a statement that is wrong and backs up the error with a source. The debrief usually shows quickly how convincing wrong answers can be. And it turns distrust into a technique that can be practised.

AI is not an encyclopaedia, it’s a very good guesser. Knowing that is what makes it useful.

The 4 Ds: turning an AI tool into a competence

Thu, 03 Sep 2026 00:00:00 GMT

This post comes with a video (1:36 minutes) and transcript: watch it in the post.

AI can write you an entire exam in ten seconds. Impressive, but it leaves the more important question open: should you use it as is? Many of us work with AI on the principle of hope. Prompt in, result out, done. That saves time, right up until the first wrong answer key is in circulation.

What’s missing is rarely a better tool. It’s a competency model: a shared idea of what good work with AI actually looks like. One such model is the four Ds.

Where the four Ds come from

The four Ds are part of the AI Fluency Framework by Rick Dakan (Ringling College of Art and Design) and Joseph Feller (University College Cork), developed in collaboration with Anthropic. The framework describes AI fluency as the ability to work with AI effectively, efficiently, ethically and safely. It comes with free courses, including courses for educators.

The four Ds accompany every use of AI: before, during and after.

The four Ds at a glance

Delegation: hand over consciously. Decide what the AI takes over and what stays with you. Having it suggest practice questions for a text? Sure. Grading a final thesis stays with you.

Delegation: the AI may suggest practice questions, grading a final thesis stays with you.

Description: describe clearly. The AI doesn’t know your class. Tell it who is learning, what the goal is and what the result should look like. The clearer the brief, the more useful the answer.

Discernment: check critically. Is the content right? Is the level right? AI doesn’t become dangerous when it’s often wrong. It becomes dangerous when it’s almost always right and we stop checking.

Diligence: take responsibility. You are responsible for what you use. That means no personal data of learners in third-party tools, and being open about where AI was involved.

A worked example: questions on a factual text

Here is the example from the video, played through all four Ds. How each step is filled in is my suggestion:

Delegation: the AI drafts practice questions. You decide which ones make it into the lesson.
Description: “Upper secondary, year 2, second-language learners. Goal: identify the arguments in a factual text. Result: five questions with an answer key, one page maximum.”
Discernment: you go through every answer in the key. Is it correct? Does the language suit this class?
Diligence: the brief contains no names or data about your learners, and the worksheet states that the questions were created with AI support.

A clear brief names the learners, the learning goal and the expected result.

An exercise for you

My suggestion for your next planning session: take the last piece of material you created with AI and run it through four questions.

What did I hand over? Should any of it have stayed with me?
Did the AI know enough? Who is learning, what is the goal, what should the result look like?
Did I check, or just skim? Content and level, line by line.
Where does my responsibility lie? Did learner data end up in the tool? Is it clear where AI was involved?

The D that makes you hesitate longest is usually the one worth practising.

Four questions before, during and after every use of AI. That’s how a tool becomes a competence, for you and for your learners.