Hello World

GenAI products continue to be asocial. That should come as no surprise, given that such products are developed by Silicon Valley companies whose CEOs act in almost every way imaginable except prosocially.¹

That doesn’t mean, however, that we have to use the underlying GenAI technologies in this prescribed way. Talking to a realistic chatbot and creating half-working game demos was an amazing showcase of language model capabilities in 2023. Not so much in 2026, where the third wave of AI hype is becoming more of a nuisance and hindrance, especially in the context of education. Can we develop EdTech platforms so that students use language models as collaboration partners when needed, instead of being forced to use them all the time?

Even though ‘Connecting People’ is a cliché, can GenAI enable new approaches to education that actively support its social aspect? Image source: Picryl

Motivation

Students already have social networks for socializing online. Why would you want to integrate social elements into an EdTech product, isn’t that superfluous?

Li and Liu (2023) analyzed design elements that influence learning outcomes, and social interaction is by far the most impactful one. Indeed, this is further supported by Self-Determination Theory (SDT), with relatedness being one of its three core elements. SDT has been the foundational framework behind many successful educational policies, as discussed in the previous article.

Are we using generative AI in EdTech to actually foster sociability or to rather suppress and diminish it? What are the trends?

Even though compelling efforts are being made to create collaborative EdTech products, gamblification seems to prevail over gamification. In other words, extrinsic rewards such as badges or daily streaks dominate to the exclusion of intrinsic rewards like active content creation, open and nonlinear exploration, or collaborative problem-solving.²

Admittedly, building intrinsically motivating EdTech platforms is several orders of magnitude more difficult than building extrinsically oriented apps. The reason is prosaic: it is far more challenging and expensive to pull off correctly.³ Does that imply such obstacles can be tackled much more easily with the presence of GenAI?

User Respect and Agency

Is it a good idea to mix human-created content with machine-generated content?

GenAI products built as solution dispensers have one indisputable cognitive advantage: you get clanker content only. This makes it much more straightforward to consume and reason about, since you know where you stand. Another characteristic such products frequently share is prioritizing output quality and consistency over the specific needs of individual users. Even though AI content detection or flagging is a useful feature, it is far from sufficient for providing credibility and trustworthiness.

Abhari et al. (2026) propose designing AI products that are coherent and consistent as a whole, rather than relying on consistency within the black box of language model inference. This holistic integrity aims to align with user expectations and the outputs of non-deterministic parts of the system.⁴

If we don’t want to just raise another generation of advanced keyboard shortcut users, we should show students what goes on behind the scenes. Image source: Picryl

Designers of EdTech products should make it obvious what the system is or is not doing. Apart from being compliant with “damage control” acts like the EU AI Act, we should go the extra mile and stop assuming what a user needs and wants at any given time. By introducing transparency, any student or teacher should:

• have control over the settings used for any inference they are presented with, rather than inspecting the LLM opaquely from the outside,
• choose which style or persona they prefer at the moment,
• be informed when the language model is uncertain or when there are multiple interpretations of their input,
• not be coerced into any behavior by the system, irrespective of its determinism.

Of course, there are already theoretical models proposing how to effectively solve problems, the Practical Inquiry model (2001) being one of them. Joy (2025) adapted this model to the context of collaborating with GenAI. Let’s demonstrate it with a specific scenario. Imagine a student is learning how to compute the area of a regular polygon:

Let’s consider a shape with an edge length of 8 decimeters and a central angle of 45°. What is the area of this shape?

1. Ask

The student is dealing with a problem they don’t know how to tackle. However, they have a knowledgeable other they can brainstorm ideas with. In this case, they use a specialist Socratic agent designed to provide gradually decreasing supervision. To start off, the student simply asks:

🧑 How do I compute the area of a polygon?

to which the agent responds:

🤖 To answer your question accurately, I need a few clarifications:

🤖 1. Are we calculating the area of a general polygon, or are we dealing with a special case, such as a regular polygon? This might help us simplify the derivation of the formula.

🤖 2. Do we have any information about the polygon? For example, the length of its sides or the measures of its interior angles.

🤖 3. Do you have any ideas of your own on how to approach the problem? We can refine or develop them further.

2. Engage

The student reviews the initial AI output and asks follow-ups, gaining a clearer picture of how to tackle the problem. They still do not consider this as an authoritative answer, but they are slowly getting their bearings.

🧑 I think the polygon is symmetric, as it has sides of the same length. I am given a=8dm, so I guess that’s a side length?

🤖 You’re most likely dealing with a regular polygon with a side length of 8 decimeters. However, there is still one missing bit of information: is there any indication as to how many sides the polygon has?

🧑 No. I mean, the only thing that it says here is that a central angle (whatever that means) has size 45°.

🤖 As a matter of fact, that’s more important than it seems at first glance. A central angle is …

🧑 … so the polygon has eight sides?

🤖 Precisely so: 360° divided by 45° is 8, which means the full angle is split into 8 equal parts. [verified by a script you can see here] Notice that this angle sits at the top of an isosceles triangle, of which there are 8 in the polygon.

3. Evaluate

Now that the student has an idea of what they’re dealing with, it’s time to compare it with their own knowledge and other credible resources, confronting their understanding with the generated output. This is where the student slowly takes the reins and the agent steps into the background.

🤖 Now you have all the information you need. What are your next steps?

🧑 That’s easy. I calculate the area of one of the triangles and multiply by 8.

🤖 That’s one of the most straightforward ways to solve it. [Glossary source: regular polygons]

Møller et al. (2026) suggest that users prefer feedback on their own reasoning and output over a pre-generated answer. Furthermore, agents in EdTech products should act as specialists rather than generalists. In other words, the output should be tailored to the context, not pretend to be a buddy with immediate answers for every imaginable scenario.

For example, if a student is prepping for a school leaving exam, they have no time for lengthy explanations. They want to follow along and get things done quickly. In that setting, it makes sense to design the agent to be a bit more directive than shown here. But “getting things done quickly” shouldn’t be the default modus operandi.

4. Reflect

The student formulates a solution plan and executes it. The agent should only verify and discuss the student’s argumentation and result against credible resources.

🧑 I got 309 dm^2, did I get that right?

🤖 That depends on what precision is required. It’s roughly correct [verified by a script you can see here]. Are there precision requirements in the assignment, or does your teacher usually require a specific level of precision?

🧑 There is no precision specified in the assignment; the teacher usually wants us to round results to two decimal places.

🤖 In that case, your result is incorrect because you’ve rounded too much. You’re very close, however! [✋ I need a hand with rounding] [✋ I need a hand with inputting the area calculation into my calculator]

🧑 You’re right, it should be 309.02 dm^2. I originally rounded to a whole number.

🤖 309.02 dm^2 is the correct answer. Do you want to go through the whole solution again so it sticks better?

Collaborative AI use is the mindset we should apply in EdTech. I firmly believe this should be the default modus operandi. We should stop striving for GenAI output excellence and learning how and to whom we can delegate all of our work.

At the same time, generating content must be an innate part of student discovery. We just need to teach students (and teachers) what to do with generated content, such as how to stay curious even when the vast majority of the job is seemingly done (but in reality isn’t). We also should avoid the cardinal sin of any UX-friendly product: ignoring users’ needs and their real struggles and instead assuming we know what they are.

Too many times, I’ve encountered online materials dwelling on technicalities most students would understand pretty well or don’t even need. To give an example from my own tutoring practice, students often struggle with the formalism of mathematics, that is, how the answer should be written so it gets accepted. We’re not grappling with principles, but with how to put them on paper, which is an incredibly useful skill to have at any time. For example, having a simple button like [Would this answer be accepted the way it is written?] might be far more valuable to students than we might initially expect.⁵

So far, we have mostly focused on individual effort. Is there also a way to design EdTech so that it brings students and teachers together, instead of locking them into the digital confinement of endless prompting with a clanker that is oblivious to human memory and experience?

Relatability of Student Effort

Emara et al. (2021) studied what happens when a group of students faces difficult problems. It turns out that the longer students remained in a group setting, the more likely they were to systematically collaborate on the problem as a whole, rather than letting a single dominant peer do the work or relying on trial-and-error.

Apart from socialization, effort-intensive teamwork relies on the ability to make compromises.⁶ Instead of insisting on quantifying every aspect of human experience, it is far more worthwhile to create equally viable learning pathways. How about creating interest-based modules that are cheaper to build than ever before: mathematics applied to forensics, historical events, music, sports, gaming, etc.? The possibilities were always endless, but now they are genuinely achievable. Once we make it ridiculously easy to add and verify new content, students might engage with topics on their own terms.

Polygons, revisited

To revisit polygons, consider a concrete scenario. A teacher or tutor prepares three viable projects, and students choose the context that appeals to them most:

🕵️ Investigators are analyzing a crime scene where blood spatter indicates an origin point, but high-impact spatter has created a multi-point perimeter boundary around the impact zone.

🤌 Artisans during the Italian Renaissance are laying a new tile floor in a merchant’s main hall, using alternating geometric shapes.

⚽ Groundskeepers are marking lines on a school sports complex for soccer and baseball fields.

Once students select their preferred topic, they join their project team. The team is presented with an interactive canvas containing guidance from the teacher. Workload scales with the number of students so that the assignment remains too challenging for a solo student to complete alone. Because each task within the project varies in nature, team members are forced to communicate and strategize.

At the center of the workspace is a shared AI toolbar that enables students to generate interactive elements, data structures, or accompanying multimedia directly onto the canvas.

Class concludes with each team presenting their solution. Every group shares core foundational knowledge with the entire class—such as calculating the area of a regular polygon or the radius of an inscribed circle—but every individual student walks away with a specific, context-rich connection to the material that is far more likely to stick in memory.

Schola ludus! Image source: Picryl

Agency Over GenAI

I personally think we can also leverage language models to deromanticize mathematics, which has traditionally been viewed as a discipline reserved only for genius mathematicians. We can start by noting how mathematics itself is dramatically changing due to GenAI. The year 2026 already marks an incredible milestone in mathematical research, as many breakthroughs in hard problems have been assisted by GenAI.

Learning by teaching is another promising approach, even if it requires prudence. We have strong indications that students themselves are willing to take part: Zimmermann and Tomczyk (2025) conducted research in which adolescents suggested that taking an active role in social media interaction makes it more engaging and valuable.

When executed correctly, enabling students to create their own materials might be one of the most creative ways to use LLMs. Compare these two scenarios:

Explain why the Pythagorean theorem holds true. You are not allowed to use generative AI; you should be able to defend your ideas on your own.

versus

Here is a challenge for you: find an unconventional proof of the Pythagorean theorem using any means necessary, understand it, and build a presentation around it. Once you have your material ready, upload it. You will defend the proof in front of your peers.

Which one of them feels more future-proof to you?

To sum up:

1. We should let students embrace generative AI and use it to its fullest potential. Of course, this necessarily means we will encounter obstacles, but that’s okay.
2. We should collaborate with language models, not act as their middle managers. This requires new ways of learning, but we shouldn’t let a few geeks from California decide the worldwide future of the next generation.
3. We should let students know when they’re working with generated content and why it was generated that way. That way, they will be able to critically assess the source of information in contexts outside of school.

Happy learning!

---

1. Arguably, asociability (or anti-sociability, even) has economic incentives as well. If GenAI were used primarily in group settings, it could lead to fewer inference calls, and therefore a smaller volume of data harvested by GenAI providers. ↩︎

2. Interestingly, peer reviews often seem to be a waste of time, as they rarely lead to meaningful knowledge transfer. ↩︎

3. Outside the context of EdTech, FigJam is frequently cited by designers as an ideal model for collaborative design prototyping. ↩︎

4. One of the most intriguing passages in this paper notes: For example, while humans can readily detect visual anomalies, they often struggle to recognize errors in textual outputs without careful verification. This leads me to a simple observation: could we supply alternative visual formats of the same concept (such as illustrations, GIFs, etc.) to make errors and mishaps easier to spot? ↩︎

5. Admittedly, many test frameworks sidestep this frustration by relying on multiple-choice formats. That doesn’t mean, however, that things will stay that way, especially in the presence of toxic EdTech products that promise zero effort while expecting top results. ↩︎

6. My observation is that the concept of a compromise itself is slowly becoming a sin (or, in modern terminology, an inefficiency). ↩︎