Same Standards, a Different Road In

In this paper

1. 1. Two failures
2. 2. Flavouring and restructuring
3. 3. What the research shows
4. 4. A record of one student
5. 5. The loop
6. 6. What MalachAI refuses
7. 7. One system, every screen
8. 8. The adults around the student
9. 9. Beyond high school
10. 10. How we’ll measure it
11. References

1. Two failures

The one teacher who ever looked at me and saw someone worth taking seriously was my high school philosophy teacher. No one managing thirty students at once can give any of them what mine gave me. That isn’t a complaint about teachers. It’s arithmetic.

The first failure is one of meaning. Reported student engagement falls from 76% in grade 5 to 44% in grade 12.^[1] The students who disengage are often not unmotivated. They are unseen. Many of them spend hours on something difficult outside school, and nothing connects that effort to what they’re asked to do in a classroom.

The second failure is newer. A general-purpose AI will now do most of a student’s homework on request. In a study of about 1,000 high school students, those given open access to GPT-4 did 17% worse once it was taken away, while a version with guardrails largely avoided the harm.^[2] Used carelessly, the tool that could give every student a patient teacher can instead remove the effort that makes learning happen.

These failures are related. A student who doesn’t see why something matters has every reason to outsource it. So any serious answer has to address both at once: why this matters to you, and why you still have to do it yourself.

2. Flavouring and restructuring

Personalization in education usually means flavouring. Putting dinosaurs in a word problem is flavouring, and anyone can do it. The requirement and the path stay the same, and the student sees through it within a minute.

Restructuring is different. A student who wants to design snowboards can meet 1800s wheat disputes through materials, tariffs and supply chains. The outcome hasn’t moved. The reason to care has. We call this a bend, and it’s the core of what CourseBend does: same standards, same school, a different road in.

Two constraints keep a bend honest. First, CourseBend is an overlay, not a school. The accredited school keeps the teacher, the grades and the diploma. Second, curriculum is sourced, never invented. Requirements are retrieved from the school of record and published standards, and cited. CourseBend writes lessons, never requirements. Anything that would change what must be learned or how it’s graded is out of scope for the product.

David Agahchen’s Intention Is All It Takes gives me a useful vocabulary for why this matters.^[3] Agahchen argues that large language models are fast pattern engines and that reasoning systems built on top of them pursue static, externally supplied goals, so neither is tethered to what a person actually wants. The human’s intentionality is intrinsic. The machine’s is derived. Most educational AI has this exact shape: it optimizes against a prompt and a syllabus, with no idea who the student is or what they’re reaching for. A passion is about as intrinsic as a teenager’s intentions get. Our bet is that it’s the right thing to tether to.

3. What the research shows

CourseBend stands on research from several fields. Taken together, it points one way: learning takes hold when it connects to what a student cares about, and when the student still does the work.

• Interest in the moment matters. In a classroom study of 14-year-olds, situational interest during a task predicted knowledge gain, while existing individual interest did not predict it directly.^[4] Lab work finds that curiosity before information improves recall of it.^[5] The lever isn’t simply having a passion. It’s using one to create the state in which learning happens.
• Interests in the structure of the problem help. In ninth-grade algebra tutoring, matching problems to students’ out-of-school interests improved problem solving and efficiency, and carried through to later exams.^[6][7]
• Connecting a lesson to your own life helps those struggling most. In ninth-grade science, students who expected to do poorly gained nearly two-thirds of a letter grade over controls.^[8] In a college psychology course, low-performing men finished about a C-to-B step ahead of controls.^[9] Parents can deliver a version of this too, with lasting effects on course-taking.^[10]
• Motivation predicts engagement far better than achievement. Across a large meta-analysis, intrinsic motivation correlates strongly with engagement and only weakly with objective achievement.^[11] Motivation gets a student to the work, and the design has to make sure the learning follows.
• AI tutoring isn’t automatically good. A 2026 randomized trial with university students found that access to a course-integrated AI tutor lowered final performance.^[12] That’s why MalachAI is designed around what the student does, not what the AI does.

Most of this work stops at ninth grade or at a single course. CourseBend is built to carry it across every required course, through all of high school.

4. A record of one student

Agahchen’s central technical move is to give the machine a persistent, evolving model of the person’s beliefs, desires and intentions, because stateless systems can’t hold what a person is trying to do.^[3] I agree, and a school subject makes the case concrete. A teacher who has known a student for three years teaches differently from one who met them this morning.

We call this record the Student Model. It’s built from observations, a word we use deliberately in place of data: something noticed, in context, about one student, for that student’s benefit. It holds a student’s ambition in their own words, a passions map that distinguishes interest from passion from obsession, a proficiency record with evidence and the date each skill was last demonstrated, signs of where they’re stuck, and their cited requirements and progress. Threads are kept in the moment, whenever a lesson bends through something they love.

Three properties set it apart from most learner profiles:

• It belongs to one student. Nothing learned about a student generalizes to anyone else. If a feature’s output has value for another student, it fails our test.
• It assumes you forget. Knowledge decays. After a gap, MalachAI refreshes at the level a student last reached, and the student proves it again. Bobby Motamed’s image is a locked box: old knowledge is still in there, and the refresher is the key.
• Dreams can move. When a student’s ambition changes, MalachAI re-asks. The old dream stays in the story, because the change is part of the story.

5. The loop

Agahchen proposes a Think, Do, Reflect loop that runs at both small and large scales.^[3] Ours is narrower and more specific to a required outcome. It has five moves.

The Check step rests on a five-level scale: encountered it, repeats it with support, says it in their own words, applies it to a new problem, and finally explains it in their own words and connects it to their passion. Parroting doesn’t count. Only the last level earns a badge, because that’s the level at which the knowledge has become theirs.

6. What MalachAI refuses

This is where I part ways with some of the language around personal AI. Agahchen describes personal AI as a “cognitive exoskeleton,” and for an adult with a life to run, that’s a good image. A fifteen-year-old has a different job. The point of a math class is not to get the problem done. It’s to grow the muscle. An exoskeleton that carries the load is exactly what a developing mind doesn’t need. What it needs is augmentive AI: technology that adds to what a student can do without doing it for them.

So the most important feature of MalachAI, our AI, is what it won’t do. It won’t produce finished graded work. It refreshes, checks and guides a student to write their own. Bobby’s phrase for it is mirrored, not offloaded. This can’t just be a matter of tone in a prompt, so we’re specifying it as a check in the code, tested against students trying to get around it.

The same restraint applies to engagement. No streak guilt, no punishment for breaks, no compulsion loops. Progress is measured against the student’s own goal, not time spent in the product.

MalachAI mentors when the moment calls for it, and otherwise stays in the background as a quiet companion to the work. I’m careful about that distinction. It’s software, it says so plainly, and it doesn’t present itself as a friend. Its feedback is honest, not flattering. A learning tool for a developing mind shouldn’t be designed to be needed, and the people who matter in a student’s life should stay their teachers, their family and their friends.

7. One system, every screen

Students don’t live in one app. So MalachAI is designed as one system that takes a native form on the screens a student already has: a phone for chat, cards and voice, a desktop beside the homework, and a camera for paper work, which it can check but never grade. Most of what sparks a student’s curiosity happens away from a screen, so wearables are a natural next frontier: tap to keep a moment from class or practice, and it becomes an observation. Capture is opt-in and student-controlled, with clear rules for schools, parents and classmates before anything records in a classroom.

The same discipline governs when MalachAI speaks. We design for positive interruptions, never negative ones. A negative interruption pulls a student back to an app for the app’s sake, breaks their focus, or trades on guilt. A positive one starts from what they care about, connects to something real that’s coming up, and leaves them more able to do it themselves. From wake-up to wind-down, every interruption has to earn its place.

On every surface the job is the same: serve the work in front of the student, then get out of the way. We’re also working toward limits such as session caps and quiet hours, because sleep and time away from screens aren’t things a learning tool should compete with.

8. The adults around the student

A teenager isn’t an autonomous adult, so the question of who sees what can’t be avoided. Our planned approach gives each adult a different view. Teachers see progress and friction, never personal content. Parents get a privacy dial, because privacy is a culture each family chooses, not a default a company imposes, and parents decide where that line sits until their child turns 18. Interesting moments reach a parent as dinner-table conversation starters, not surveillance reports. These portals come after our first version.

CourseBend holds itself to a stricter standard still. Our commitment is that CourseBend can’t read a student’s personal chats, while academic observations stay readable so the record can work. We’re engineering our first version around that line.

9. Beyond high school

Nobody stops being curious at graduation. A student who starts at 14 could carry their record into university, apprenticeships, work and beyond, and we think of that as CourseBend, then CourseBend College, then CourseBend Beyond. At each stage the work gets harder and the support steps back. What carries through is a student’s own story.

Agahchen extends the argument from persons to groups, imagining many intention-aligned systems supporting choices that people make together.^[3] We have a smaller version of the same intuition: nobody falls in love with a subject alone. A community around CourseBend, for students, parents and teachers, is part of where this goes, designed with the care minors deserve.

We’re building high school first, because that’s where the required road is narrowest.

10. How we’ll measure it

Our first pilot, with my son, his friends and others, starts in the new year, building on our onboarding work with BC students. These are the measures we care about most.

• Retention. Engagement is the easiest thing to move and the easiest to mistake for learning. The measure that matters most is delayed retention: does the student still have it, later, without help?
• Independence. Success means a student more able to think without MalachAI, not less. Developmental and ethical advisors will be part of this from the start.
• Family trust. Every family sets its own privacy line, and a student’s privacy and a parent’s visibility both have to hold.
• Reach across the curriculum. We’ll track how well the bend carries through the hardest required units, for every kind of passion.

As AI reshapes work, job preparation alone can no longer be the purpose of education. Students will need to know what they care about, and they’ll need to be able to think without a machine doing it for them. I think both can be taught through the courses they’re already required to take. That’s what we’re building.

Acknowledgments

Many of the ideas here came out of working with Bobby Motamed, who shapes how CourseBend teaches. Any errors are mine.

References

1. Gallup student poll figures, as reported by Education Evolving.
2. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS.
3. Agahchen, D. (2025). Intention Is All It Takes: Binding AI to Human Intention Through Extended Conscious Systems. Etcera. etcera.com/docs/intention-is-all-it-takes
4. Rotgans, J. I., & Schmidt, H. G. (2018). How individual interest influences situational interest and how both are related to knowledge acquisition. Journal of Educational Research.
5. Gruber, M. J., Gelman, B. D., & Ranganath, C. (2014). States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit. Neuron.
6. Walkington, C. (2013). Using adaptive learning technologies to personalize instruction to student interests. Journal of Educational Psychology.
7. Bernacki, M. L., & Walkington, C. (2018). The role of situational interest in personalized learning. Journal of Educational Psychology.
8. Hulleman, C. S., & Harackiewicz, J. M. (2009). Promoting interest and performance in high school science classes. Science, 326.
9. Hulleman, C. S., Kosovich, J. J., Barron, K. E., & Daniel, D. B. (2017). Making connections: Replicating and extending the utility value intervention in the classroom. Journal of Educational Psychology.
10. Rozek, C. S., et al. (2017). Utility-value intervention with parents increases students’ STEM preparation and career pursuit. PNAS.
11. Howard, J. L., et al. (2021). Student motivation and associated outcomes: A meta-analysis from self-determination theory. Perspectives on Psychological Science.
12. Liu et al. (2026). The effects of course-integrated AI tutoring on student performance and engagement: A randomized university trial. EdWorkingPaper 26-1598. Not yet peer reviewed.