Study Shows AI Can Slow Down Experienced Software Developers

Image credit: Unsplash / Study Shows AI Can Slow Down Experienced Software Developers

METR study finds AI coding assistants increased seasoned developers' task time by 19% in familiar codebases, challenging productivity hype.

Study Shows AI Can Slow Down Experienced Software Developers

When the promise of speed becomes a drag

Introduction

Have you ever imagined **losing time** by using an AI coding assistant? For many engineers that sounds impossible—but that’s precisely what researchers at the non‑profit METR found when they watched veteran programmers working on massive open‑source projects. Instead of accelerating, AI stretched delivery times by **19 %** citeturn4view0. The result throws cold water on the widespread belief that any AI automatically equals productivity.

What the METR study says

1. **Participants:** 20 developers with at least five years of experience on large projects.

2. **Tool tested:** Cursor, one of the GPT‑powered assistants most popular on GitHub.

3. **Method:** Each dev handled real tasks inside repositories they already maintained; half the time with Cursor, half without.

4. **Key outcome:** Average completion time **rose 19 %** when AI was enabled. Perception, however, still indicated a 20 % speed‑up—a measurable placebo effect citeturn4view0.

Why is the result so surprising?

Previous research—like Microsoft’s benchmark showing 56 % gains—looked at *green‑field* code or toy tasks. In real life, senior developers wrestle with legacy code full of quirks, and that’s exactly where AI stumbles by offering patches that are “almost” right but still need human polishing. Put differently, what speeds up a junior may slow down a veteran.

How coding assistants work today

Large language models (LLMs) generate snippets directly in your editor. They learn statistical patterns and therefore deliver **“the most likely code,”** not necessarily **“the perfect code.”** The older and more idiosyncratic the base, the more *refactoring* the AI-generated patch will need. The generate‑review‑fix loop is the friction METR quantified.

Business impact for engineering teams

- **Invisible cost:** If a squad of ten senior devs spends 19 % more hours on the same deliverable, the yearly budget can balloon by hundreds of thousands of dollars.

- **Perceived quality of life:** Developers reported lower *stress* thanks to AI assistance—even with delays. That morale boost could reduce talent churn.

- **Hybrid strategy:** Use AI for rapid prototypes and test scaffolding, but keep humans in the loop for critical, large‑scale edits.

Practical tips for engineering managers

- **Map scenarios:** Apply assistants only where the AI learning curve is short.

- **Measure real data:** Instrument your CI/CD pipeline to compare lead time with and without AI.

- **Manage expectations:** Explain the difference between *“easier”* and *“faster.”* Comfort is not speed.

- **Update code policies:** Consider human pair‑programming in areas where AI suggestions tend to misfire.

Expert voices

Software pioneer **Margaret Hamilton**, who led NASA’s Apollo software, once said that “every new abstraction adds convenience at the cost of new complexity.” **Kent Beck**, father of Extreme Programming, echoes: “Premature optimization of productivity without measuring first is the root of all engineering failure.” Both urge a sober view: AI is a **tool**, not a **shortcut**.

Alternatives and complements to Cursor

- **GitHub Copilot Enterprise:** fine‑tuned models, private repo context, stronger privacy controls—may reduce hallucination by understanding your internal code better.

- **Tabnine:** language‑specific models; reports fewer hallucinations but less flexibility.

- **Codeium:** free for small teams, with an ultra‑fast *snippet search* that works like semantic search on your own code.

Run an A/B pilot with two different assistants to discover which one fits your stack.

AI and Developer Experience (DevEx)

Productivity is not just **lines per hour**. Factors like team morale, user satisfaction, on‑boarding time and deploy frequency all shape success. If AI makes work more enjoyable—even if slightly slower—the overall balance might be positive. The most valuable metric is still **time to customer value**.

Extended conclusion

Bottom line? **AI isn’t magic.** It needs context, curation and metrics to truly add value. Before embracing any coding assistant, test it in your environment, with your people, and measure without bias. Otherwise you risk swapping your car’s engine mid‑race—only to finish behind the pack.

> “Powerful tools demand proportional discipline.” — **Conway’s Law reimagined for the AI era**

Stay tuned: we will soon benchmark Google’s brand‑new **Gemini Code Assist** against Cursor. Subscribe to our newsletter and don’t miss out!

Broader industry context

Investment in AI coding assistants has skyrocketed over the past two years, with venture capital pouring billions into startups that promise to upend traditional software workflows. Products like **Replit Ghostwriter**, **Amazon Q Developer**, and **DeepMind AlphaCode 2** tout lightning‑fast scaffolding of boilerplate and test generation. Yet hard data on *sustained productivity* remains scarce. The METR publication brings badly needed nuance to the table, showing that one size does **not** fit all.

According to market analyst **DevMetrics**, at least 68 % of mid‑size tech firms adopted some form of AI pair programmer in 2024. However, fewer than 30 % have formal KPIs tied to the rollout. “Enterprises chase the hype cycle,” warns DevMetrics partner **Larissa Chu**, “but without measurements, they risk sinking costs into tools that don’t fit their maturity model.”

Policy and ethics angle

The slowdown uncovered by METR also intersects with debates about **AI governance**. If experienced developers spend extra hours correcting AI suggestions, who is liable for bugs that slip through? Some legal scholars argue that unchecked use of generative code could clash with evolving regulations on software safety and accountability—especially in sectors like finance and healthcare.

What comes next?

METR plans to replicate the experiment with **junior engineers** and **private enterprise codebases**. If they observe speed‑ups among novices, companies might adopt a **tiered approach**: AI mentors for beginners, human gatekeepers for mission‑critical merges. Either way, the conversation is shifting from blind optimism to **evidence‑based adoption**—and that’s healthy for the craft of software engineering.

Feeling curious about where LLMs are headed? Read our deep dive into **GPT‑5’s multimodal powers** () to see how AI is expanding beyond text into video and audio.