AI can't make you care enough to learn
I ran an experiment at home that taught me something about work.
Picked three gamified learning apps for my kids. Research-backed. AI-enabled. Adaptive difficulty. Two weeks in, I reassessed. Retention had dropped. Not stayed flat — dropped.
They’d learned to game the system. Skip between apps when I wasn’t watching. Hit the daily streak requirement, then bail to the games they actually wanted to play. The tools worked exactly as designed. The learning didn’t happen.
I started asking around at work. Found out people are using AI to break down our LMS content into simpler language — then still asking humans to explain it. One SE told me he feeds every technical doc through internal AI tool first “to make it readable,” then messages me when he still doesn’t get it.
We have more tools than we’ve ever had. More formats. More interactivity. More adaptive difficulty. The threshold to actual absorption hasn’t moved.
Learning still requires crossing a specific threshold. That moment where it clicks. Where you can apply the knowledge without scaffolding. AI can reword the content. It can generate a podcast. It can role-play scenarios. It can’t make you care enough to cross that threshold.
My kids had every advantage — adaptive algorithms, instant feedback loops, progress tracking. They still needed me in the room explaining why the concept mattered. Not what it meant. Why it mattered to them specifically.
There’s research here that maps directly to what I’m seeing in the field. Kirschner and Sweller’s work on cognitive load theory shows that learning happens when working memory successfully encodes information into long-term memory — but working memory is extremely limited. You can hold about four chunks of information at once. Adding more formats doesn’t expand working memory. It just adds more things competing for those four slots.
The Cognitive Theory of Multimedia Learning, developed by Richard Mayer, found that people learn better from words and pictures than from words alone — but only when the pictures directly support the words, not when they’re decorative. Every extra element that doesn’t directly serve the learning goal increases cognitive load and reduces retention.
Albert Bandura’s social learning theory demonstrated that people learn most effectively through observation, imitation, and modeling — especially when they trust the model. His research showed that learning isn’t just about information transfer. It’s about seeing someone you trust demonstrate that the knowledge matters and works.
The teams I’ve seen get this right aren’t the ones with the most AI-enabled content. They’re the ones who’ve protected the social learning infrastructure. Weekly office hours where an SA walks through a specific customer scenario. Slack channels where people post “here’s what I tried today and why it worked.” Peer shadowing where a junior SE watches a senior run discovery, then they debrief in the car afterward.
That last one — the debrief in the car — that’s where the tribal knowledge transfers. Not in the recorded demo. In the ten-minute conversation after where the senior explains, “I asked that specific question because I noticed his body language shift when we talked about timeline.”
AI can’t pick up on that. Can’t tell you why it mattered. Can’t model what caring about the outcome looks like.
I’ve been predicting this since ChatGPT launched three years ago. As much as AI replaces human roles — and it’s replacing plenty — the jobs of teachers, enablers, and coaches will be the last to go. Not because the technology can’t get there eventually. Because learning is fundamentally social.
We’ve been learning in circles for thousands of years. Direct instruction. Knowledge sharing. Discussion. Peer influence. That’s the original format. Everything since then has been trying to scale it.
AI-enabled enablement isn’t a magic wand where if we deploy the tool, we can replace all the human trainers. Before we had these tools, my kids couldn’t learn independently without my guidance. Same thing is true at work.
The only way I can imagine AI getting close to that level — and I doubt it will — is if AI accumulates so much context about the user that the user feels like learning from it is reliable. But we know that’s a machine. That human touch won’t go away.
The gap isn’t in the content anymore. It’s in the care. Whether someone cares enough to cross the threshold from “I read it” to “I can use it.”
You can’t automate caring.
What this means right now.
This week: Audit your learning stack. If you removed all the async content tomorrow, what would break? That’s your real learning infrastructure. The parts that only work when humans show up.
This month: Pick one piece of tribal knowledge that only transfers peer-to-peer. Record it happening. Not the demo — the debrief after. That’s what you’re trying to scale. Capture the “here’s why I made that choice” moment.
This quarter: Build one social learning ritual that can’t be replaced by a tool. Weekly live Q&A. Peer shadowing with mandatory debrief. Slack thread where people post one thing they learned by watching someone else this week. Small. Repeatable. Human.
The goal isn’t to eliminate AI-enabled tools. It’s to protect the moments where learning actually transfers. Those moments are always social. Always specific. Always human.
What’s the one piece of tribal knowledge in your team that only transfers when humans are in the room together?