New research confirms what many of us suspected: AI assistance comes at a cost. The question is whether that cost matters.
The Research That Made Me Think
Anthropic recently published research on AI assistance and coding skills that put numbers to something I've been mulling over for months.
The findings are striking. In a randomised controlled trial, developers using AI coding assistants scored 17% lower on comprehension tests than those who coded manually. That's nearly two letter grades. The gap was largest in debugging, the skill you need most when AI-generated code doesn't work as expected.
Speed gains? Marginal and statistically insignificant.
This isn't about whether AI tools are useful. They clearly are. It's about what happens to us when we use them.
The Handwriting Parallel
Here's what caught my attention: this mirrors something we've known about for years.
Studies consistently show that physically writing notes with pencil and paper leads to better retention than typing them digitally. The slower, more effortful process forces your brain to engage differently. You can't transcribe verbatim when writing by hand, so you summarise, prioritise, think.
The same principle seems to apply to AI assistance. When AI writes the code, or takes the meeting notes, or drafts the email, our brains shift into a different mode. We become reviewers rather than creators. Consumers rather than producers.
I notice this in myself. When I receive obviously AI-generated meeting notes, they hold less value to me than notes I perceive as written by a human. There's something about knowing that a person had to filter, interpret, and decide what mattered that makes the output more trustworthy. Or maybe it's just that I know the person who wrote them actually processed the meeting, not just recorded it.
The Uncomfortable Question
So here's the trade-off we're all navigating, whether we acknowledge it or not:
AI assistants make us faster at tasks while potentially making us worse at the underlying skills those tasks develop.
For a senior developer who already has deep expertise, using Copilot might be pure productivity gain. The skills are already there. For a junior developer still building their mental models, the calculation is different. Every line of code the AI writes is a learning opportunity not taken.
The Anthropic research found that usage patterns matter. High performers used AI to clarify concepts through follow-up questions, treating it as a tutor. Low performers simply delegated, accepting whatever came back without interrogation.
The tool is the same. The outcome depends entirely on how you use it.
Does It Actually Matter?
Here's where I'm genuinely uncertain.
We've lost skills before. Nobody worries that we can't navigate by the stars anymore. Calligraphy went from essential business skill to niche hobby. Mental arithmetic gave way to calculators, and the world didn't collapse.
Maybe coding skill atrophy is the same pattern. Maybe in ten years, debugging AI-generated code will be as quaint as knowing how to tell the difference between Orion's Belt and the Plough. The AI will debug itself, or the code won't need debugging because it will be generated differently.
But there's a counterargument. Those lost skills, navigation, calligraphy, mental maths, were replaced by tools that work reliably without human oversight. AI is different. It's confident but not reliable. It produces plausible outputs that require skilled review to trust.
If you can't debug code because you never learned how, you can't catch the bugs the AI introduced. If you can't write because you've always had AI draft your emails, you can't tell when the AI's tone is subtly wrong for your audience.
The skills AI replaces might be exactly the skills you need to supervise AI effectively.
What We're Doing About It
We think about this a lot when building automations for clients. The goal isn't to remove humans from processes entirely. It's to remove the tedious parts so humans can focus on judgement, relationships, and the decisions that actually need a brain.
The best automations we build handle the mechanical work while preserving the cognitive engagement that keeps people sharp. Data gets collected automatically, but humans still interpret it. Reports get generated, but humans still decide what to do about them.
It's a deliberate choice. We could automate more aggressively. Sometimes clients ask us to. But there's value in keeping people engaged with the work, not just informed about the outcomes.
No Easy Answers
I don't have a tidy conclusion here. The research raises questions that don't have clean answers yet.
What I do know is this: the organisations that will navigate this best are the ones thinking about it deliberately. Not banning AI tools, that ship has sailed. But designing workflows that capture efficiency gains without accidentally atrophying the skills that matter.
For individuals, the lesson from the research is clear. If you're using AI as a delegator, accepting outputs without engagement, you're probably losing more than you're gaining. If you're using it as a collaborator, questioning, learning, staying cognitively involved, you might get the best of both worlds.
The tool doesn't determine the outcome. How you use it does.
Thinking about how AI fits into your team's workflows without replacing the skills that matter? Let's have a conversation about getting the balance right.