AI coding creates two kinds of debt. You’re only measuring one
Summary
Dr. Margaret-Anne Storey introduced "cognitive debt" in October 2025, defining it as the accumulated loss of understanding a system's function, purpose, and team knowledge distribution, distinct from technical debt. This debt arises when AI generates code faster than engineers can build a shared mental model of the system. Infobip's longitudinal study, spanning eight cohorts and 225 interns since 2022, revealed that AI-era cohorts (from Winter 2024) experienced significantly lower skill growth across technical, teamwork, problem-solving, and personal excellence dimensions compared to pre-AI groups. For instance, technical growth plummeted from +2.89 in Winter 2023 (0% heavy AI usage) to +0.69 in Summer 2025 (54% heavy AI usage). Crucially, this decline occurred despite rising Net Promoter Scores and high intern satisfaction, indicating that traditional metrics fail to surface cognitive debt. The issue stems from "passive delegation" to AI, where users accept output without understanding, hindering the "productive struggle" vital for learning and knowledge distribution, as demonstrated by both Anthropic's January 2026 study and Infobip's intern observations.
Key takeaway
For engineering leaders overseeing AI tool adoption, recognize that rising satisfaction scores can mask declining skill growth and accumulating cognitive debt. You must actively foster "cognitively engaged" AI use, not passive delegation, by implementing structured practices like "try first, check with AI second" and AI-free checkpoints. Prioritize measuring skill trajectory over satisfaction and conduct targeted retrospectives to ensure shared system understanding, preventing hidden knowledge gaps that impede future adaptability and control.
Key insights
AI coding can induce "cognitive debt," a hidden loss of system understanding and skill growth, often obscured by rising satisfaction metrics.
Principles
- Cognitive debt is distinct from technical debt.
- Software understanding is a shared mental model.
- Productive struggle is vital for skill development.
Method
Infobip implements "Try first, check with AI second," AI sequencing (understand then accelerate), and AI-free checkpoints. Mentors act as coaches, ensuring visibility of professional standards. Teams should use retrospectives to map architecture to user needs and identify who understands each subsystem.
In practice
- Implement "Try first, check with AI second" rule.
- Conduct AI-free checkpoints for skill calibration.
- Use retrospectives to map system understanding.
Topics
- Cognitive Debt
- AI Coding Assistants
- Skill Development
- Engineering Leadership
- Software Engineering Metrics
- Mental Models
Best for: CTO, VP of Engineering/Data, AI Engineer, Software Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by LeadDev.