How “Thermomix Mode” Can Make Students Stop Learning with AI

How “Thermomix Mode” Can Make Students Stop Learning with AI

Introduction: A kitchen machine that explains a classroom fight

Generative AI in education is having a moment—everyone’s using it, and everyone has an opinion about whether it helps or harms learning. The new research behind this post, from Rummel, Nachtigall, & Panadero (arXiv:2609.09856), tackles that debate using a surprisingly relatable metaphor: the Thermomix.

The core idea is simple: when you use a tool, you don’t just change how you do something—you change what kind of person you become while doing it. And with AI, that question is often missing from public discussions. Instead of asking only “Should students use ChatGPT?”, the authors ask: how does using AI shape learners’ engagement, agency, and skill development?

To make that concrete, the paper compares different Thermomix use styles (guided cooking, tweaking recipes, doing more manual cooking, and building community around it) with different ways students might use generative AI. Then it sorts those use styles using two well-known learning frameworks: ICAP (how cognitively engaged learners are) and SAMR (how technology changes or transforms tasks). The punchline: it’s not AI vs. learning—it’s AI use patterns vs. meaningful learning.

Why This Matters (Right Now): Your classroom may be quietly training “cook assistants”

This research lands especially well today because schools are already moving past the “ban it / allow it” phase. Most teachers aren’t asking whether AI exists—they’re asking how to manage it in realistic student workflows: drafts, feedback, rewriting, brainstorming, citations, and “final submission” practices.

Here’s the uncomfortable part: many classrooms are functionally teaching substitution—the learning equivalent of pressing “start” on a guided cooking program. Students don’t just use AI; they can end up outsourcing cognition while thinking they’re just saving time. And once that becomes normal, you don’t only lose skills like writing fluency—you also risk losing the habit of revising, reasoning, and building understanding from scratch.

A scenario where this applies today: imagine an intro writing course. Students draft essays by prompting AI for an outline, then ask it to “write the full essay,” then submit with minimal edits. The teacher sees a polished final product and may reasonably assume learning happened. But the research lens suggests a different question: what cognitive work did the student actually do? If the student didn’t verify, critique, or restructure the content, the learning process might be mostly passive—while the output looks active.

Also, this paper builds on earlier AI-in-education research that often compares AI vs. no AI using experimental designs. Those studies are important, but the authors argue they can get stuck in simplified “effect” questions that confound lots of factors (like task design, student knowledge, and how teachers guide use). By using the Thermomix analogy, the paper gives a more actionable framework: not “does AI help?” but “when does AI help—and when does it quietly replace the learning work?”

The Thermomix isn’t just a gadget—it reorganizes roles (cook vs. assistant)

The Thermomix is marketed as a “magic pot” style appliance: it chops, mixes, steams, kneads, and even supports guided recipe cooking through a touchscreen and/or app recipes. It can make high-quality food easier—and it can also trigger a cultural backlash: some people claim you’re no longer “really cooking” if you’re just following steps.

That controversy matters for education because it shows how tools reshape identity and competence. If someone uses the Thermomix only in guided mode, they may produce meals without building the underlying judgment skills: tasting, adjusting heat, knowing when texture is “right,” understanding ingredients deeply enough to improvise.

In learning, the same identity shift can happen. Tools don’t simply help students finish—they can redefine what “good performance” looks like, and what learners think they’re responsible for.

Two learning lenses used in the paper: ICAP + SAMR

The authors organize the analogy with two frameworks:

So in this paper, the question “Is this use of AI good?” becomes: Which ICAP mode is students landing in, and what SAMR stage is the AI use actually at?

Scenario 1: Guided Thermomix vs. “AI wrote it, so I’m done” (the passive-substitution trap)

The Thermomix guided mode: you can cook without cooking skills

One popular Thermomix use case is “guided cooking mode,” where the user follows step-by-step instructions. The paper describes how reactions are mixed: some appreciate it because it enables people with low confidence—or limited time—to cook at all. Others argue it can erode real cooking competence because you’re following instructions rather than learning to make decisions.

In ICAP terms, guided cooking is mostly Passive: the user receives a sequence of actions to perform. In SAMR terms, it sits at Substitution when the machine replaces manual cooking steps without fundamentally changing the user’s learning process.

The AI parallel: outsourcing entire assignments without revision

The learning-risk scenario in the paper mirrors that perfectly: students outsource entire assignments to generative AI and then accept the output without critical engagement or meaningful revision.

The authors connect this concern to research showing that when people offload cognitive work, the skills that depend on that work can fade. In the paper’s discussion, they reference examples like:
- Clinicians relying heavily on ChatGPT-like tools potentially experiencing declines in reasoning because they stop weighing alternatives.
- A study (by Radtke & Rummel, 2025 mentioned in the paper) where novice writers revised AI-produced texts more superficially and invested less cognitive effort than when revising their own drafts.

In other words, the output looks like effort; the process might not be.

The trade-off: substitution can still help some students (but only if tasks are chosen carefully)

Importantly, the paper doesn’t treat substitution as universally bad. The authors note that replacing low-level tasks with AI can sometimes support inclusion or free time for deeper work—if teachers and students are reflective about what gets delegated.

But there’s a catch: relying on AI for foundational processes (like composing from scratch) assumes students already have enough prior knowledge to engage critically later. Without that base, AI can reduce the chances to build it.

Practical implication: If your assignment rubric only rewards final quality, students will learn to use AI in the passive-substitution style. If you want a different outcome, you’ll need assessment that forces cognitive involvement—revision logs, source verification, argument restructuring, or oral defenses of decisions.

Scenario 2: Tweaking ingredients vs. cross-checking AI (active augmentation needs prior knowledge)

The Thermomix “I can alter it” mindset

In the Thermomix analogy, the second use case is the moment users shift from following instructions to adapting them. The paper describes testers saying that anyone with some cooking knowledge treats guided recipes as a rough framework: you can skip steps, repeat steps, or change the process based on preferences and experience.

This is more Active in ICAP terms, because you’re manipulating the recipe and thinking about what changes mean. In SAMR terms, it becomes Augmentation: technology adds functionality that supports active adaptation rather than fully replacing the human process.

The AI parallel: verify, refine prompts, revise text

For generative AI, the same pattern shows up when students don’t just accept AI output—they:
- cross-verify claims using credible sources and then revise the AI output accordingly,
- refine prompts when the results don’t match their understanding,
- revise AI-generated drafts more deeply based on disciplinary expectations.

The paper’s discussion points to evidence that experienced writers revise AI-produced text more thoroughly. It also mentions interventions around prompt engineering where students gain (descriptively or significantly) knowledge about AI concepts and self-efficacy.

What ties this together is a prerequisite: students need enough background knowledge to notice what’s wrong and to make meaningful changes. Without that, “active use” turns into confident misuse.

Practical implication: Teachers can structure assignments so verification and refinement are required steps—not optional extras. For example:
- require citations for factual claims,
- require a “what I changed and why” section,
- require at least two independent sources to confirm key statements,
- include prompt/refinement screenshots or logs as part of submission.

Scenario 3: Manual cooking + creativity vs. constructive generation (from help to meaning-making)

When Thermomix becomes a creativity tool

The third Thermomix scenario is where users become more “creative and versatile” by using the device’s manual functions—not just guided mode. The paper highlights that Thermomix can be used in more hands-on ways (stirring, kneading, steaming, chopping, simmering, etc.), and that with more familiarity, manual cooking can become “more interesting.”

This aligns with Constructive engagement: people generate new approaches and expand their repertoire. In SAMR terms, it becomes Modification: the task changes—users don’t simply replicate recipes in a digital format; they redesign how they cook complex dishes.

The AI parallel: AI as a sparring partner, not a ghostwriter

For generative AI, the “constructive” style is when students use AI to develop their own ideas through:
- brainstorming and outlining,
- evaluating and reflecting on information,
- generating drafts that students then reshape into a meaning-making process.

The paper describes research where AI can act like a “sparring partner,” stimulating critical engagement rather than replacing thinking. The key is that AI is treated as a catalyst for learning activities, not as a generator of the final truth.

The assessment twist: redesign the assignment so AI can’t do all the thinking

This is a big “teacher lever.” If you keep the assignment as “produce an answer,” AI will likely land students in passive or active-but-superficial modes.

But if you redesign tasks so students must:
- argue with AI output,
- justify edits,
- reflect on reasoning,
- critique assumptions,
then you’re pushing toward constructive/modified learning.

Practical implication: Consider prompts like:
- “Use AI to generate three perspectives—then identify which perspective you disagree with and why.”
- “Produce a draft outline with AI, but then rewrite the thesis and reorganize evidence based on your sources.”
- “Write an argument, then use AI to challenge it; revise only after you can explain why your original reasoning stands (or doesn’t).”

Scenario 4: Online communities + dialogue vs. interactive AI feedback (dialogue enables redefinition)

Thermomix communities: knowledge co-construction

The final Thermomix scenario moves beyond the individual user. The paper discusses how Thermomix apps and recipe databases helped create communities where users share recipes, test each other’s creations, and give feedback.

This transforms cooking from solitary execution into something closer to Interactive learning. People ask questions, negotiate solutions, and build shared knowledge—exactly the kind of dialogue that ICAP associates with strong learning outcomes. In SAMR terms, this becomes Redefinition because the practice becomes something new compared to traditional cooking with only pots and cookbooks.

Generative AI as a dialogue partner that provides real-time feedback

The generative AI analog is interaction: chat-based AI that supports ongoing conversation, feedback, questioning, and adaptive responses.

The paper emphasizes that such dialogue can help personalize feedback at scale—something many classrooms struggle to provide. It also references models of AI as:
- a dialogic sparring partner,
- a Socratic “friend” that encourages critical thinking through questions,
- a partner that supports creative problem-solving dialogue.

In this scenario, the learner isn’t simply consuming output; they’re negotiating meaning in real time with the system. That aligns with ICAP’s interactive mode and SAMR’s redefinition stage.

Practical implication: If you want this benefit, make interaction part of the task. Don’t just ask for “a final answer.” Ask for:
- an exchange log (“What I asked, what it suggested, what I challenged”),
- iterative refinement steps,
- evidence of feedback-driven reasoning changes.

What it means for identity, authorship, and evaluation: “Who counts as the cook?”

The final section of the paper asks a philosophical but highly practical question: What makes someone a learner—or a cook?

The authors argue that competence doesn’t only mean storing knowledge in memory. It can also include the ability to mobilize tools, judge information quality, and adapt strategies. That matters because generative AI can shift students from “makers” into “managers of knowledge”—but managing knowledge still requires judgment and cognitive skills.

Then comes the assessment problem. Teachers (and examiners) often evaluate only the final product. They rarely know how much human reasoning produced the output. That raises questions about:
- authorship,
- trust,
- and whether grading should look at process, not just outcomes.

The paper references the idea that assessment should be integrated into learning with generative AI—aiming to evaluate critical thinking, communication, and collaboration skills, not just polished text.

Practical implication: To keep learning authentic, assessment must be aligned with the cognitive mode you want:
- If you want interactive learning, assess dialogue-driven reasoning.
- If you want constructive learning, assess revision depth and meaning-making steps.
- If you only assess final submission, you’ll incentivize substitution.

Key Takeaways

• AI isn’t the deciding factor—AI use style is. The Thermomix analogy shows a continuum from passive following to interactive co-construction.
• Using generative AI in a passive, substitution-like way (“AI wrote it, I’m done”) risks skill erosion because learners offload core cognitive work.
• Active augmentation (cross-verifying, refining prompts, revising with knowledge) can help, but it depends on students having enough background knowledge to judge quality.
• Constructive/modified use (brainstorming, critiquing, redesigning tasks) is where AI becomes a creativity and meaning-making catalyst.
• Interactive/redefined use (dialogue with AI + feedback loops) can transform learning into a conversation—closer to how effective learning often works.
• For educators: redesign assignments and assessment so AI can’t complete the “learning work” by itself—require verification, revision rationale, and interaction logs.
• For the future: research and practice should move from “Does AI improve learning?” to “Under what conditions, for whom, and with what guidance does AI produce real engagement?”—exactly the kind of question this paper pushes you to ask.

If you want, tell me what subject/grade level you’re thinking about (e.g., high school English, first-year university writing, teacher training). I can translate these scenarios into a few concrete assignment templates and rubrics that push students toward constructive or interactive learning—not Thermomix-guided substitution.