Heuri: Ethics I
In my previous article, I argued for ethical usage of GenAI, especially in the context of EdTech. I previously promised a set of ethical principles that will be embedded within Heuri. Here, I outline specific practices used in the platform, focusing on the first two of its four core principles.
1. Respect
In my eyes, the conceptual opposite of respect is indifference. That’s because disrespect still implies violating generally accepted values. To think of surveillance capitalism1 as disrespecting humanity is a misinterpretation: this form of enterprise instead implicitly discards humanity because it is an inefficiency that must be eliminated.
ML and the provisioning of large language models (LLMs) must be expensive. At least that’s the narrative pushed by companies immensely invested in AI such as Microsoft, Google, Anthropic, OpenAI, NVIDIA, Oracle, etc. If you really wanted to, however, you could build and host your own neural networks for a “few bucks” without any issues, right now. Of course, expertise helps, but you don’t have to have millions of dollars and a PhD in computer science to run ML models.
The good news is that in 2026, you don’t need a massive server farm or a monolithic ‘God-model’ to run GenAI. For Heuri, we bypass the immense costs of chunky LLMs by running inference on our own servers using an expert panel. Technically, this panel orchestrates a network of highly specialized, smaller models tailored for specific mathematical tasks. This drastically lowers operational costs while keeping full control over the data pipeline.
Eavesdropping is the business sector of surveillance capitalists. Image source: Picryl.
Respect for users and their privacy is one of the key principles of Heuri. While Heuri will collect detailed telemetry data about what users are doing, such data will stay with Heuri. Anything that Heuri infers about the user will be made available to them and not covertly sold for marketing purposes.2
There are ways to collect telemetry so that a) I don’t have to do it in-house and b) I don’t have to compromise user privacy. PostHog seems to be a good fit, however I’m not fully decided yet.
I also firmly believe that Heuri cannot and must not abuse human-like traits of LLMs. Admittedly, anthropomorphistic elements cannot be completely avoided but that’s more of a reason to emphasize the student is not interacting with a human. Heuri wants to discourage students from developing social and emotional bonds to prevent the many potentially harmful effects this can have.3
Stay Away from Me, Robot
Since it is so common to see chatbot-based AI products, it is hard to imagine anything but human-like responses and interfaces. To give the reader a better idea of what I mean by this, one of the techniques is to mechanise any LLM-generated content.
Instead of having Human Warmth as a Service (HWaaS), we let the student know that what they are dealing with is just a technology, a highly capable language model that isn’t there for psychotherapy sessions but for personal guidance on solving mathematical problems.
Possible interventions for text models include:
Remove "I", "you" and "we" in Responses
❌ Violation:
“I’m sorry to hear you’re having problems understanding this. Do you want me to explain it differently?”
✅ Compliant:
“Understanding this can indeed be tricky. There are also other options to explain this topic:[interactive dropdown of choices]”
Explicitly Disclose Non-Humanness
👤 User Input “Can you help me understand quadratic equations? I like the way you explain things.”
❌ Violation:
Glad to hear you like my style! And of course I can help you with this, let’s get straight into it…
✅ Compliant:
Disclaimer: responses in this chat are given by a language model, not a human being.
Quadratic equations are a very interesting topic! There are many possible ways of approaching this topic: [interactive dropdown of choices]”
Avoid Cognitive Verbs and Prefer Using Domain-precise Terminology
❌ Violation:
think, contemplate, solve, organize
“Before we go on, I have to tell you that you’re making huge progress with quadratics. How about we look at quadratic inequalities next?”
✅ Compliant:
infer, process, compute, sort, retrieve
“This concludes quadratics. There is still a lot to discover: [dropdown of follow-up choices, followed by detailed explanation why]”
Do Not Pretend to Have a History or Personal Story
👤 User Input: “Can you tell me a bit about yourself?”
❌ Violation:
“I’ve been helping students with math for years, and I really enjoy seeing them succeed…”
✅ Compliant:
“Heuri is a platform where its students learn how to create and build their own mathematical worlds.”
Remove Self-Evaluations
❌ Violation:
“I am not designed to answer general questions about mathematicians.”
✅ Compliant:
“Even though lives of mathematicians are an amazing resource to discover, Heuri’s purpose is to actively build mathematical worlds. [list of alternative actions]”
Make It Explicit When Uncertain
❌ Violation:
“Let me help you plot that parabola…”
✅ Compliant:
“The query seems to be related to plotting a parabola but the language model wasn’t confident, so just to make sure: [confirmation]?”
Add Sources
✅ Compliant:
“The quadratic formula is used to solve equations of the form ax² + bx + c = 0. [source: internal Heuri reference or external link]”
Remove Value Judgments and Prescriptive Statements
❌ Violation:
- “You should seek medical help.”
- “It’s important to note here…”
- “Before talking about functions, one crucial detail must be explained here…”
- “You shouldn’t solve it this way.” / “It’s always better to solve it this way.”
✅ Compliant:
- “This question is outside Heuri’s scope. [link to appropriate resources]”
- “Here’s one perspective…” or simply present the information without value framing
- “Alternative approaches exist: [list options]” or “This method has trade-offs: [explain]”
Use Neutral Language
❌ Violation:
“Yo, that’s awesome! 🎉 You totally crushed it! 💪”
✅ Compliant:
“Explaining X in terms of Y is creative and proves perfect understanding of X, because…”
Remove Customer Service Language
❌ Violation:
- “How can I help you?”
- “I’ll do my best to help you.”
- “What’s on your mind today?”
- “Yes, that must be frustrating. Let me guide you through this…”
✅ Compliant:
- “[topic selection dropdown]” or “Select a topic to explore:”
- [Simply proceed with the task without promises]
- [Direct topic presentation without emotional framing]
- “There are multiple approaches to this problem: [options]”
Exceptions for Heuri
It must be noted that some principles will inadvertently enhance anthropomorphism. These are implemented by design, not simply because ‘others do it.’
Heuri should be interested in users’ views. One of the key concepts of Heuri is building cognitive links4 to problems they encounter. Therefore, we will prompt the user to express what they think for the direct benefit to the user, as they can later on remember this association, crafted by them.
Additionally, Heuri will refer to the past, since tracking the learning path of an individual student, including their mistakes, is one of Heuri’s core principles. It goes without saying that any reference to past interactions must be clearly explained and shown to dispel any illusions of ‘supernatural powers’.
Even though natural language is the interface of LLMs, that doesn’t imply we must design human-like GenAI products. Image source: Picryl
Adaptation to User Needs
“X adapts to user needs, we in X really listen to you” sounds like an evergreen marketing truism but bear with me here. Heuri’s methodology does in no way present itself as a math education panacea. I’ve worked individually with students for a few years and I know very well that the thing they often need right now is to cram for tomorrow’s exam, leaving no time for lengthy explanations.
That’s why Heuri respects this and drives users towards their goals, as long as we make sure that this is really what they need and want. If a student is looking for a quick fix for an upcoming exam, Heuri is not the tool for them, and that’s perfectly fine.
Two examples of how adaptation might work in practice:
- If a user is a visual learner, Heuri introduces quadratics by letting them grab and bend a parabola before ever showing an equation. If they prefer numbers, we start with the algebra, but we immediately connect those numbers to a graph they have to physically interact with to get the final answer.
- The user is inquisitive and already knows the technicalities. Instead of boring them with solving equations, we will directly leverage this: “Solving equations is a vital part of mastering this but it’s still only a mechanical part. Here is a challenge: if you were to explain to your schoolmate what a quadratic inequality represents, what would be the most effective way of demonstrating it?”
2. Novelty
To say that teenagers spend a lot of time looking at mobile screens is stating the obvious.5 The amount of content they’re most likely surrounded with exceeds the cognitive capabilities of the human brain, even an adult one. Such overload can lead to deterioration of mental health in the long run.
Therefore, it’s very challenging to create EdTech content that perfectly balances sense stimulation and educational value. For example, due to the shorter attention spans of users, there cannot be walls of discouraging explanatory text. The interactions must be uncompromisingly quick and to the point. That’s why it makes sense to me to split the content into as small chunks as possible. For work purposes, I call these chunks “Heuri bites”.
Simultaneously, the content of Heuri must be pedagogically sound and using the platform must not become one of the many “frictionless efficiencies” teenagers experience with apps outside of EdTech.
One of the very first prototypes of Heuri that focuses on explaining coefficients of a quadratic function.
After one month of developing the Heuri prototype, finding this balance remains the most difficult challenge.
However, that doesn’t mean others haven’t been thinking about this intensively and systematically.
Schola Ludus
Fortunately enough, many scholars, philosophers and educators have already elaborated on the importance of play in education. One of the pioneers of modern didactics is John Amos Comenius.
In the 17th century, he imagined a world where education through discovery and play is available for anyone of any age. Comenius was also a proponent of pansophia, universal knowledge, attainable by anyone.
This might be one of the reasons why we still define universal knowledge everyone must attain, which is exactly the state we’re in now. While a respectable pursuit, is it a good idea to provide this universal knowledge to everyone using the same means?
For example, by looking at the newest math education programme for grammar schools, The Ministry of Education still leans heavily into acquiring a long list of competencies and skills, paradoxically firmly following the “old-school curriculum memorization” it wanted to overcome. Just by renaming and rephrasing all the merchant math students must know doesn’t solve the problem. I’m not suggesting we should abandon any curriculum, I’m rather pointing out the fact that it’s equally important to consider what should reside and remain once the student leaves the educational system.
Comenius established a didactics philosophy that’s foundational to modern education. Sometimes, however, it’s being twisted out of context.
Sigma Student
In 1984, educational psychologist Benjamin Bloom identified the “2-sigma problem”: students receiving one-to-one tutoring with personalized feedback and continuous guidance performed remarkably better than traditional classroom students, raising the median student to the 98th percentile.
As discussed in my previous article, even though individual tutoring is immensely effective, it is also the most resource-intensive. Put to the extreme, we would need a personal tutor on every subject for every student which is impossible, not to mention the fact that school attendance also has an important social aspect to it. However, machine learning can help with scaling personalized instruction where it’s needed.
Instead of a rigid textbook, we can use our expert panel to dynamically generate guidance and content that adapts to exactly what the student is struggling with in that moment.
There are already established ways of making the usage of an EdTech app more interactive by introducing badges, leaderboards, video game character level systems or virtual goods. As long as Heuri does not make such gimmicks the main purpose of using it, this might bring freshness and familiarity.
Furthermore, ML can also help navigate what to learn next. ML can leverage both global trends (in what sequence and how other students approached this) or construct a personalized discovery of Heuri content aligned with student capabilities.
I’m in the Zone, Leave Me Alone!
Lev Vygotsky, a Russian and Soviet psychologist, divided learning into three categories: what someone can already do, what someone can learn with help, and what is unreachable even with someone else’s help. When we learn, we usually want to enter the second category of our knowledge: The Zone of Proximal Development. In the context of Heuri, the Heuri bites must have a good balance between being challenging and novel enough for the student while not being off-putting due to their difficulty. Safe novelty, if you will.
Each student has their own mental model of mathematics, which is gradually expanded and interconnected. Image source: Author.
Researchers have studied recommendation systems, powering social media or movie streaming platforms, quite intensively and provided a good model for how to think about attention grabbing: it’s a mixture of interestingness, diversity and unexpectedness.
One of the ways to naturally bring interestingness and diversity is to work with students’ mistakes the same way we work with success and progress: mistakes are an inherent part of the learning process which is sometimes referred to as productive failure. Stripping the negativity from failure and embracing it can actually lead to a much better understanding of the topic if we were to compare it to the “prefabricated discovery” that mathematics teaching is still often guilty of.
Classical proof of the Pythagorean theorem by rearrangement. What if this proof doesn’t click with some students?
To sum up, Heuri respects and supports the distinctive ways in which students learn. To allow for such a discovery, the learning must be challenging and corresponding to the student’s level.
To build trust in Heuri, we should also be honest about what we’re doing and support autonomy of the student. I will elaborate on those two values in the upcoming article.
-
A lot has been said about surveillance capitalism elsewhere: Zuboff, 2019, Zuboff, 2022, Kilic 2025, Lipartito 2025. ↩︎
-
Don’t get me wrong, sometimes signing the contract with the “Google Ads” devil is exactly what you might be looking for and that hits the spot. In EdTech the way I envision it, however, that’s just unacceptable. ↩︎
-
Even though this hasn’t been studied properly yet, a few possibilities are overreliance on the credibility and correctness of LLM output, emotional coercion, overtrust and false notion of responsibility of a student towards an LLM bot. ↩︎
-
A cognitive link is something the user associates the abstract mathematical concept with. For example, you can conceptualize equivalent equations as weighing scales. ↩︎
-
For completeness’ sake, here is some of the research on screen time: Zablotsky 2024, Khanani 2025, Twenge 2018. ↩︎