I think you might be fooling yourself with AI
Claims that AI-assisted coding boosts developer productivity are challenged by mixed evidence, including an often-cited 2025 study showing developers felt faster with AI yet were actually slower, and a newer follow-up with inconclusive results. Commenters contrast “fast” AI-driven prototyping and “slow” AI as a rubber‑duck tutor with traditional hand coding, reporting everything from negligible gains to dramatic speedups depending on task type, codebase complexity, and developer skill. Alongside productivity, many raise concerns about the long‑term economics and environmental cost of large-scale AI, while others argue that even if today’s business models are unsustainable, the underlying technology is clearly useful and unlikely to disappear.
AI and Developer Productivity
- Strong split: some report 2–10x gains, others see neutral or negative impact once verification and debugging are included.
- Gains are largest for: simple CRUD, glue code, helper libraries, unfamiliar domains, prototypes, and “bureaucratic” tasks (emails, summaries).
- Several note no benefit or even slowdown for complex, low-level, or library code where correctness and architecture matter deeply.
Workflows: Fast vs Slow AI Use
- “Fast AI”: let agents generate large chunks or whole features quickly → great for demos/POCs, risky for production and understanding.
- “Slow AI”: incremental use, lots of questioning and review → feels less stressful than hand-coding yet not much faster, but more sustainable.
- Many mix both: AI for initial scaffolding, humans for refactor, testing, and “humanwashing” the code.
Skill, Understanding, and Code Quality
- Concerns that heavy AI use erodes skills and makes devs unable or unwilling to work without it; some evidence from a study where many refused no-AI conditions.
- Others emphasize codebases becoming harder to understand when “vibecoded,” even if feature lists grow faster.
- Some treat AI as a powerful “rubber duck,” forcing clearer requirements and mental models.
Economic Sustainability and Costs
- Debate over whether massive AI capex and hidden debt make current offerings unsustainable vs. tokens being profitable on the margin.
- Skeptics predict price hikes, consolidation, or failures; supporters argue tech will persist like the internet post–dot-com, possibly at lower unit costs.
- Disagreement over whether current productivity gains justify near‑trillion‑dollar investments.
Environmental and Ethical Concerns
- Worries about energy use (largely fossil-fuel powered), climate impact, and community harm; some argue this is being hand‑waved away.
- Others counter that the core issue is dirty energy, not AI per se, and expect energy to get cheaper/cleaner, but critics invoke Jevons paradox (more efficiency → more total use).
Evidence, Studies, and Epistemics
- The 2025 METR study (developers felt faster but were ~19% slower with AI) is heavily cited by skeptics.
- Pro‑AI posters stress a 2026 follow‑up showing a speedup with newer tools, though the authors themselves flag serious methodological limits.
- Broader argument over relying on single studies vs. accumulated personal experience, and over selectively citing research that matches priors.
Adoption Patterns and Use Cases
- Clear enthusiasm in small teams, indie projects, and data science workflows where individuals control architecture and context.
- Corporate‑level ROI remains “unclear”; several note that traditional org bottlenecks (testing, requirements, politics) can eat coding gains.
- For non‑engineers, AI can make previously impossible or too‑expensive software tasks feasible, which many see as genuinely transformative.
Social Dynamics and Emotions
- Participants accuse both “pro‑AI” and “anti‑AI” camps of tribalism, insecurity, and motivated reasoning.
- Some critics admit anger and anxiety about AI’s direction; some boosters display high confidence that skeptics are simply “wrong.”
- Several argue that dismissing obvious everyday utility undermines more serious debates about externalities and long‑term risks.