Using AI to Personalize Learning for Every Student

piloting personalized ai learning paths with 30 students right now and the early results are genuinely interesting – not perfect, not magic, but interesting enough that i’m going to keep going. wanted to share what’s actually happening in the classroom before this either works or blows up in my face.

## what i set up and why

the basic idea was simple: instead of everyone getting the same assignment sequence, students interact with an ai to diagnose where they are, then get a slightly different path through the material based on those gaps. i teach mixed-ability groups so the range in any given class is pretty wide. some kids are two years ahead, some are still missing foundational stuff. one lesson plan has never actually served all of them at the same time.

i’m running this across two class sections, about 30 students total, ages 14-16. the subject is math – algebra and early functions mostly.

## early results – the actual numbers

after about six weeks here’s where things stand:

– 8 students who were chronically behind have completed more practice problems voluntarily than in the entire previous semester
– 4 students who were already advanced have gone significantly deeper into extension material
– the middle group is roughly the same, which honestly surprised me
– teacher time spent on re-explaining basics dropped noticeably – i’d estimate 30-40% of those intervention conversations just aren’t happening anymore because the ai caught the gap first

the thing that’s working best is the feedback loop speed. a student gets something wrong, the ai identifies the specific misconception (not just “wrong answer”), and adjusts the next three problems accordingly. i can’t do that 30 times simultaneously.

## the real challenges nobody talks about

here’s where it gets messy though:

– some students are gaming it. they figured out that certain types of wrong answers loop them back to easier content, so they’re deliberately tanking diagnostic questions
– students with lower reading ability are struggling with ai-generated explanations that assume too much vocabulary
– one student just stopped engaging entirely – turns out she found the whole thing isolating and missed the classroom discussion piece
– my school’s data policy is a gray area and i’ve had to have awkward conversations with admin about where student inputs are going and who can see them

the isolation thing is the one i’m most concerned about. the personalization can make each student feel like they’re on their own island. i’ve started building in more structured moments where the class compares notes on what paths they ended up on, which helps, but it’s a design challenge i didn’t anticipate going in.

also the equity issue is real. students without reliable home devices are effectively getting a different version of the pilot than everyone else, which is the opposite of what personalization is supposed to do.

## where i’m taking this next

thinking about adding a peer component where students in similar learning paths occasionally work together – kind of like dynamic grouping based on where the ai has placed them rather than fixed ability tracks. also want to tighten up the gaming problem by making the diagnostic questions less predictable.

privacy-wise i’m moving toward running more of this locally where i can. the privacy benefits alone make local worth it for my use case once you get past the setup friction.

anyone else running something similar at scale, even small scale like this. specifically curious whether others have cracked the isolation problem, and what you’re doing about students who disengage from the ai interaction entirely rather than leaning into it.