Post

Ml 04

Posted
Ml 04

1. Respect of other people

Surveillance capitalism1

Heuri adapts to every individual user and their needs. Heuri must minimize Time-to-Mastery. Namely the way information is presented to them: not only mood and characters, but in general: - Does user prefer cramming for exams or are they willing to take time to understand things? - Does user prefer gentler, more patient approach? - Are they curious how stuff works?

Gemini answer:

  • Tone Switching: The AI decides whether to use the “Patient Tutor” prompt or the “Challenging Peer” prompt.
  • Content Routing: The AI decides whether to serve a direct equation or an analogy.

  • Notification Timing: The AI decides when to send a push notification to maximize open rates.

  • How I engage the user with learning mathematics. No matter the user profile, we should slowly converge towards a “mathematical world” that the student builds and owns and is able to use outside of Heuri. Two examples of such paths:
    • User is used to mind-numbing solving of equations. We let them do that for a while (to show them the app is useful to them RIGHT NOW) but here and then, we will poke them: listen, you’ve been doing quite great at those quadratic equations. Don’t you want to learn more about what they’re capable of? I have an amazing analogy for you.
    • User is inquisite and knows the technicalities. Instead of boring them with solving equations, we will directly leverage their positive attitude: look, I won’t bore you to death with solving quadratic equations. We can already do better than this! Let me give you a challenge: if you were to explain to your peer what quadratic inequality represents, how would you go about this?

You cannot improve what you do not measure. Track adoption, churn and override rates, and ask users if they understand decisions. Watch support tickets to spot trust issues. Continuously evaluate bias and drift, set thresholds for error and fairness, and adjust as needed.

2. Excitement, novelty, challenge in life

What inputs are NOT in my control?

  • Content itself. There is no denying users will use the app mainly because they learn math at school.

  • Short attention span of users. There CANNOT be walls of text, the interactions must be quick, brisk, to the point.

  • We should “sense” when we’re drifting too much from the user needs. That’s where ML can shine as well. https://www.saqibsafdar.com/the-cognitive-debt-crisis-designing-ethical-ai-for-deeper-learning/
  • We ban “frictionless efficiency”
  • No “banking model” of education (teachers “deposit” knowledge into students), leading to loss of student agency

In 1984, educational psychologist Benjamin Bloom identified what he called the “2-sigma problem”.Through his research, Bloom demonstrated that students who received one-to-one tutoring consistently performed two standard deviations better than students in a traditional classroom setting; a massive improvement equivalent to raising the median student to the 98th percentile. Bloom attributed these remarkable gains to the core components of effective tutoring: personalised feedback, mastery-based sequencing, and continuous corrective guidance.

But AI can solve this -^

AI can help construct a “personalized Zone of Proximal Development (Vygotsky)”! Not only as a scaffolding but as an active element in helping students picking WHAT to pick next!

Gradual help, not direct answers immediately (issue with many tutors and online websites that I try to stay away from).

So-called ‘study modes’ are merely behavioural overlays, telling AI to act like a tutor without embedding actual pedagogical understanding. If we bolt on “learning modes,” we still risk cognitive debt because the system isn’t pedagogically aligned. This is just a way of GenAI companies to feel they’re doing something against intellectual atrophy / cognitive debt.

As soon as a student stops asking question, we know we’ve screwed up big time: either they’ve completely disociated from the topic or want to understand but are missing a key point.

First, proper sequencing that respects cognitive development. Second, rich context awareness. i.e. understanding the learner’s prior knowledge, emotional state, and goals. Third prioritise productive friction. It should be designed to support the learner within their ZPD, providing hints, prompts, and Socratic questioning rather than direct answers.

spacing out study sessions, interleaving different topics, and requiring retrieval practice

work on “productive failure” demonstrates that allowing learners to grapple with complex problems and even fail before receiving instruction can lead to deeper conceptual understanding than direct instruction alone

Instead of optimising for engagement metrics like clicks or time-on-task, which can incentivise shallow, game-like interactions, the system’s success would be benchmarked against genuine learning outcomes, such as improved performance on transfer tasks or demonstrated mastery of concepts.

Serendipity is a big part of this

RS (recommendation system based on ranking “next most attention-grabbing content” by interestingness, diversity and unexpectedness) https://www.researchgate.net/publication/389157870_Design_of_a_Serendipity-Incorporated_Recommender_System

3. Creativity

mention constructivism as a backing pedagogical philosophy that’s aligned with this

https://trainingindustry.com/magazine/winter-2026/your-brain-on-ai-the-surprising-science-of-cognitive-atrophy/

One study demonstrated that generative AI enhances individual creativity but reduces the collective diversity of novel content.

4. Honesty, Openness and Transparency

https://www.parallelhq.com/blog/designing-ai-transparency-trust https://www.parallelhq.com/blog/ethical-considerations-in-ai-design

  • How I process user behavior. I would like to track their mistakes and their strengths, not only for adjusting the content for them but also to effectively motivate them: “good job!” is cheap compared to: “look, we can work on this later. I’ve already seen how you aced X, so the problem isn’t you can’t get the hang of X; we just need a step back, what do you reckon? (an interesting idea just occurred to me: could we structure the learning in the way dialogues are done in Fallout: New Vegas? user can also have “perks”, hm…)”

Make Process Visible

Show Uncertainty

  • Doing user behavior without using third party service will be most definitely tough but I don’t want a spying system for arguably the most vulnerable online demographic. This can be our strength: “Look, the rules are simple: what happens in Heuri, stays in Heuri, if you need to. Since we’re doing this from scratch, things won’t be perfect: but together, we can make it perfect, for you. We’re not sending your behavioral patterns to third parties because we appreciate the trust you put into Heuri and believe you should be in control over your data.”

Gemini had a brilliant point: You don’t have to compromise your morals to save time. In 2026, there are phenomenal open-source, self-hosted analytics platforms (like PostHog).

McKinsey notes that over 40 % of business leaders see lack of explainability as a key risk of artificial intelligence yet only 17 % are addressing it.

Interpretability

LIME and SHAP for xAI

Mind the Product advises designers to clearly communicate how results are ranked and let users adjust inputs to reduce bias.

So, something like “I didn’t like this answer” or “Report this answer”

Decision logic must be shown

Commit to bias mitigation techniques like balanced datasets, adversarial de-biasing and fairness metrics.

Keep a record of model updates and communication logs.

Provide “why this recommendation” links or tooltips summarising the reasons behind an output. Design interactive explanations that allow curious users to drill down for more detail. Avoid dark patterns. Let users undo or override automated actions when appropriate. By designing for AI transparency and trust, your interface becomes a guide rather than a black box.

honesty about weaknesses builds credibility

If users cannot question or override an artificial intelligence output, trust erodes. Offer ways to seek clarification or opt for human review.

https://www.topbots.com/data-capitalism/ No digital feudalism: services either cost money or your privacy (and extraction of whatever is digitally extractable). It’s your choice. We offer you privacy-first service for an exchange of money.

Commercial AI products are typically designed for a general market, with opaque algorithms and objectives optimised for engagement and scalability, not for the specific pedagogical values of an educational institution.

https://standardbeagle.com/designing-trust-in-ai-products/ Trust doesn’t come from AI being “smart.” or behaving like a person. It comes from UX that’s transparent, predictable, and grounded in real human needs.

Designers should treat every system failure as a moment to demonstrate integrity. A well-designed fallback, such as “I didn’t understand, can you rephrase?” keeps users in control. The system’s errors aren’t hidden. They are acknowledged and recoverable.

https://ojs.aaai.org/index.php/AIES/article/view/31613

Anthropomorphistic elements cannot be fully evaded but we must be explicit in that students are not interacting with a human: we want to discourage the student from developing a social and emotional bond to prevent possibly many harmful effects this bond can have (overreliance, manipulation, emotional overtrust and dislosure, false notion of responsibility towards an AI bot).

https://arxiv.org/html/2502.14019v1

Possible interventions in text models:

  • remove “I” and “we” in responses -> “I’m sorry to hear that you’re having problems understanding this. Do you want me to explain the topic differently?” -> “You seem to have problems understanding this. …”
  • remove “you” to address the user -> “I’m sorry to hear that you’re having problems understanding this.” -> “There are also other options to explain this topic. [interactive dropdown of choices]”
  • explicitly disclose non-humanness -> “Can you help me understand quadratic equations? I like the way you explain things.” -> “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]”
  • the output should fit a template/predictable format
  • avoid cognitive verbs and prefer using domain-precise terminology (infer, process, compute, sort, retrieve, …) -> e.g. spinning wheel (“inferring response…”, “processing user input…”, “searching for the best follow-up topic…”)
  • do not pretend to be in a body: “tell, speak, communicate, go with, …” -> (chatbot response) “I have to tell you that you’re making huge progress with quadratics. How about we look at quadratic inequalities first?” -> “This concludes quadratics. Exploration of this topic is not over, though! [dropdown of follow-up choices, followed by detailed explanation why]”
  • do not pretend to have a history or personal story -> (user query) “Can you tell me a bit about yourself?” -> “Heuri is a platform where its students learn how to create and build their own mathematical worlds.”
  • remove self-evaluations -> “I am not designed to answer general questions about mathematicians.” -> “Even though lives of mathematicians are an amazing resource to discover, Heuri’s purpose is to actively build mathematical worlds. [list of alternative actions/reformulated questions]”
  • make it explicit when we’re not sure: -> “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 -> internal Heuri references or external links
  • remove value judgments and prescriptive statements -> NO “You should seek a help from a doctor.” -> NO “It’s important to note here…” -> NO “Before talking about functions, one crucial detail must be explained here…” -> NO “You shouldn’t solve it this way.” / “It’s always better to solve it this way.”
  • use neutral language and refrain from using slang, aggressive and swearwords, don’t use emojis
  • remove customer service language: “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…”
  • perfect grammar, no typos, no inaccuracies

exceptions for Heuri

  • we should be interested in user’s views
  • we must refer to the past

5. Independence of thought and action-choosing

Create Dialogue, Not Delivery

Support Thinking, Don’t Replace It

Build Toward Independence

  • we should collect as much structured data as possible about what keeps any user from progressing towards their goal
  • what mistakes any user does
  • what successes the user has already achieved
  • why users leave the app; using an annoying exit questionairre might not be it, we should be able to figure out ourselves from the user footprint and only confirm that our guess was correct

https://www.cambridge.org/core/journals/memory-mind-and-media/article/wherever-there-is-ai-there-is-memory-ai-as-the-agency-of-the-synthesized-past/C9D626E1F920D54DAC7E4C1D2DB40933

AI is a temporal mirror, acting as us in the past.

In law: The point is that the best decisions must come from the present generation of judges, not from a loop where the past interprets itself. -> in the context of education: we shouldn’t build EdTech products that are a perfect assembly of the past (didactics, processes, …) because they will reinterpret (and reproject) past as the only way to view today.

https://blogs.worldbank.org/en/education/is-ai-making-us-smarter-or-just-making-us-look-smart-

Research has shown that poor use of AI can lead to even worse performance before using it. We cannot force students to juggle and enhance their short-term memory, their mathematical understanding must be sound and solid (refer to Hejný).

  1. A lot has been said about surveilance capitalism elsewhere: Zuboff, 2019, Zuboff, 2022, Harvard Kennedy School, 2025, Lipartito 2025. ↩︎

Music Fun Fact

Loading a fun music fact...