Cognitive debt is emerging as a risk in AI-driven development

A growing conversation among developers is centered on a problem some are calling cognitive debt: the gap between how a system changes and how well a team still understands it. In a recent post, the author said generative AI and agentic AI may be making that gap worse by helping teams move faster than they can preserve shared understanding.

The author said the discussion picked up after a previous post on the topic drew responses from several communities. Rather than answer each comment separately, the new post tries to pull together what practitioners are saying and how their experiences connect.

Several developers, including Simon Willison, have described feeling this problem directly, according to the post. On a Hacker News discussion of a Martin Fowler article, participants said they were getting lost in their own projects and finding it harder to confidently add new features. They could ship faster, but they were losing the deeper sensemaking that ties decisions to intent and intent to code.

The post argues that this is not just a matter of code quality. It is about whether people on a team can keep a coherent mental model of what a system does, why it was built that way, and how it can be changed over time. The author says velocity can outpace understanding.

The costs show up in people as well as software. The post lists several effects: less confidence when making changes, more review work, harder debugging, slower onboarding, and more stress and fatigue. The software may still function, but the reasoning behind it becomes harder to recover.

The author also points to other writers who have described similar experiences. Siddhant Khare has written about AI fatigue, Steve Yegge has written about burnout from AI-accelerated development, and Annie Vella has described the emotional strain of uncertainty when systems become harder to reason about. Taken together, those accounts suggest this is not only an engineering issue, but also a human one.

Martin Fowler has argued that cognitive debt, like technical debt, must eventually be repaid, according to the post. The author agrees, but says the repayment is not just a matter of refactoring code. It also means restoring what the post calls the “distributed theory of the system” - the intent, constraints, and architectural reasoning that are spread across people, documentation, tests, conversations, tooling, and increasingly AI agents.

The post says that repayment becomes harder when teams are under pressure to move quickly, whether in startups trying to learn fast or larger organizations pushing AI adoption. Under those conditions, it can be easy to postpone the work needed to keep understanding in sync with change.

Some commenters argued that this is simply a failure of engineering discipline, pointing to specifications, rigorous reviews, testing, and architecture documentation. The author says those practices still matter, but adds that AI changes the incentives. Because AI lowers the cost of producing structure, the structure can evolve faster than shared understanding can stabilize.

The post says teams need to actively shape their practices to keep pace. It also notes that specifications and documents are not enough if they are not living artifacts that people actually use.

Readers shared mitigation strategies in response, according to the post. These included stricter review practices, writing tests that capture intent, updating design documents continuously, and treating prototypes as disposable. Some also described using AI to reduce the cost of that work and to help with dependency tracking, explanation, and other forms of cognitive support.

The author’s main question is how high-performing teams will adapt as AI becomes more common in both startups and large companies. The post asks how teams will use generative and agentic AI not just to produce code faster, but to preserve collective understanding as systems evolve.