This post argues that AI coding tools carry a hidden cost — cognitive debt — that erodes developer skill even as output speeds up.
1. The Speed Illusion: A 40-Point Perception Gap
- METR/Anthropic RCT (2025): 16 experienced open-source developers, 246 real tasks. The AI-assisted group finished 19% slower, yet predicted they'd be 24% faster and still believed they were 20% faster afterward — a ~40 percentage point gap between perception and reality.
- Tilburg University (GitHub Copilot): Core developers' code reviews rose 6.5% while their own output fell 19%. Junior developers produced more, at the cost of seniors' cognitive load reviewing uneven AI code.
- Anthropic (Shen & Tamkin, 2026): 52 professional developers learning a new async library showed 17% lower comprehension with AI assistance. Debugging was the most impaired skill.
- LLM group showed ~55% reduced neural connectivity (alpha band ~47%), near sleep-state levels.
- 83% couldn't quote what they had "written"; 78% still couldn't after AI was removed — suggesting persistent, not temporary, effects.
- Crucially, participants who worked independently first, then used AI showed *increased* connectivity. Order of use matters more than total use: AI should extend your thinking, not replace it.
- Low-reliance users: read requirements carefully, verify answers, make fine-grained edits when pasting, navigate contextually.
- High-reliance users: skip comprehension, copy-paste wholesale, repeatedly revisit LLM answers, and ultimately accept wrong information even after hesitating.
- Addy Osmani (O'Reilly), citing Margaret-Anne Storey: cognitive debt is accumulated cost from rising cognitive load as codebases evolve. AI creates a 5–7x speed-comprehension gap (AI generates 140–200 lines/min; humans comprehend 20–40).
- Triple Debt Model (Storey et al., arXiv:2603.22106): 1. Technical Debt — structural code problems 2. Cognitive Debt — erosion of shared team understanding 3. Intent Debt — unrecorded rationale, goals, and constraints
- arXiv:2602.20206: Among 78 novice programmers, unrestricted AI users had a 77% failure rate vs. 39% for scaffolded AI guidance. The problem is usage, not the tool.
- Kosmyna, N. et al. (2025). "Your Brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task." *arXiv:2506.08872*. MIT Media Lab.
- Becker, M. et al. (2025). "The Effects of AI Assistance on Software Development: A Randomized Controlled Trial." *METR / Anthropic*.
- Shen, Z. & Tamkin, A. (2026). "Developer Comprehension Under AI Assistance." *Anthropic*.
- CHI 2026. "Behavioral Indicators of Overreliance During Interaction with Conversational Language Models." *ACM CHI Conference*.
- Storey, M.-A. et al. (2026). "Triple Debt Model: Technical, Cognitive, and Intent Debt in AI-Accelerated Software Development." *arXiv:2603.22106*.
- He, J. et al. (2025). "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." *Tilburg University*.
- Osmani, A. (2026). "Comprehension Debt: The Hidden Cost of AI-Generated Code." *O'Reilly Radar*.
- "Mitigating Epistemic Debt in Generative AI-Scaffolded Novice Programming." *arXiv:2602.20206*.
- Addy Osmani on Cognitive Debt: https://www.aicerts.ai/news/cognitive-debt-the-hidden-cost-of-ai-coding/
2. The Brain Outsourced: MIT EEG Findings
MIT's "Your Brain on ChatGPT" (Kosmyna et al., arXiv:2506.08872) monitored 54 participants over 4 months:
3. Initial Anchoring: Why the First Answer Is Most Dangerous
A CHI 2026 study ("Behavioral Indicators of Overreliance...") tracked 77 participants working with LLMs seeded with deliberate errors. Key behavioral markers:
The LLM's fluent, authoritative, single-answer style anchors users to the first answer and switches off verification — a cognitive autopilot.
4. Engineering Definition of Cognitive Debt
5. Three-Tier AI Usage Model
| Tier | Mode | Outcome | |---|---|---| | 1. Copy-paste | Paste output unverified | Cognitive debt, skill decay (danger) | | 2. Concept exploration | Think first, use AI for exploration | Neural connectivity improves (safe) | | 3. Think-first | Design independently, AI for details | Best comprehension and debugging (optimal) |
How you use AI determines your skill level five years from now.
6. From AF447 to Cursor: The Automation Paradox
The aviation industry saw this first: degraded manual flying skills after autopilot reliance contributed to the 2009 Air France AF447 crash that killed 228 people. Programming is following the same path — not because tools like Cursor, Claude Code, and Copilot are bad, but because their very efficiency lets the brain quietly dismantle the neural circuits that constitute core engineering competence.
7. Five Principles Against Cognitive Surrender
1. Write before asking: work independently 15–20 minutes before consulting AI (MIT-confirmed ordering effect). 2. Explain to the AI: articulate your reasoning first to activate slow, deliberate thinking. 3. Review line by line: never "accept all." 4. Regular AI-free days: manual-flight training for the brain. 5. Keep "cognitive receipts": after using AI, note what it did and why you agreed or changed it.
Conclusion
Tools are neutral; usage is not. Speed is not capability, and efficiency is not growth. The most dangerous skill degradation in the AI era is not being unable to code — it's *believing you can*. What will separate great engineers in five years is whether they retain high-order thinking after AI absorbs the low-order tasks.