There's been another round of "students caught using AI" headlines this week. For a community that builds intelligent systems, the interesting part isn't the moral panic, it's the design failure underneath it.
Most AI tools are optimised to do the task. Paste the question, get the answer. That's great for productivity and terrible for learning, because understanding is the one thing you can't outsource and still keep.
So a question worth chewing on as builders: what does an AI system look like when its goal is to make the human more capable, not less?
Some design principles we keep coming back to:
- Withhold the final answer by default. Ask the next question instead.
- Make reasoning visible. Show the path, not just the destination.
- Measure success by what the user can do without the tool afterwards, not by how much the tool did for them.
- Treat "the user learned nothing but got the output" as a failure state, not a win.
This cuts across education, but also onboarding, clinical decision support, technical training, anywhere a shortcut quietly erodes capability.
Where have you seen AI make people sharper instead of more dependent? And how would you actually measure that?