A new study is challenging the prevailing narrative that artificial intelligence (AI) tools universally turbocharge software development productivity.

According to a randomized controlled trial published by METR and widely reported by Reuters, experienced open-source software developers who were allowed to use AI coding assistants in early 2025—tools akin to GitHub Copilot or ChatGPT—actually took 19% longer to complete programming tasks compared to peers who worked without such aids. This surprising result, outlined in the Reuters coverage, contradicts the fervent industry belief that AI integration would deliver immediate, dramatic productivity gains for all developers.

The METR study highlights a deployment gap versus controlled benchmarking environments. Real-world coding often involves complex, context-rich repositories and ambiguous requirements, which AI tools still struggle to navigate efficiently. The Reuters report points out that while AI models can quickly generate code snippets or boilerplate functions, the process of integrating, debugging, and verifying those suggestions often introduces new friction for seasoned experts. The study’s findings imply that for the most experienced professionals—already adept at code synthesis and architecture design—AI-generated suggestions can serve as more of a distraction or an additional layer to vet, rather than a source of timesaving automation.

Contrasts and Nuance: Productivity Gains Aren’t Universal

Contrast this with the broader, often-optimistic industry consensus that AI is accelerating software engineering. Publications like Jellyfish’s 2025 State of Engineering Management report note widespread adoption of AI-powered development tools, with many engineering leaders citing improved team velocity and enhanced code quality. Jellyfish’s CEO, Andrew Lau, told the publication that “unlocking [AI’s] full value demands more than access. It requires intentional measurement, structured enablement, and cultural investment.” However, only 20% of organizations report using robust engineering metrics to rigorously quantify AI’s impact—the rest may be relying on anecdote or vendor promises.

Surveys cited by Brainhub reinforce that the impact of AI is not monolithic. While approximately 30% of developers worry about potential job displacement by AI by 2040, most today use AI for targeted tasks such as bug detection, code refactoring, or generating routine code. These technologies can automate low-level programming chores and free up developers to focus on higher-order design and problem-solving. Yet, as the METR study shows, the current generation of tools may be less beneficial—or even counterproductive—for the deeply experienced cohort of engineers who already work at maximum efficiency.

Measurement, Management, and the Learning Imperative

As AI adoption accelerates, pressure mounts for engineering organizations to close the gap between investment and measurable outcomes. Jellyfish underscores the need for Software Engineering Intelligence (SEI) platforms to track how AI changes productivity and team well-being. This is particularly critical in a context where AI-fueled automation is expected to reshape development workflows and business models, but where the risk of burnout and workflow disruption still looms large.

The imperative for lifelong learning and adaptability is clear. As Brainhub emphasizes, developers must now constantly update their skills—not only learning new AI tools, but also refining critical thinking and problem-solving abilities that AI cannot easily replicate. The rapid evolution of AI means today’s roadblocks may become tomorrow’s efficiencies, but the path will likely remain uneven across experience levels and specializations.

The Road Ahead: Cautious Optimism and Open Questions

For industry insiders, the key takeaway from the latest research is that AI’s impact on software development is highly context-dependent. While junior or less experienced programmers may see substantial productivity benefits from AI assistance, senior engineers must carefully weigh when and how to incorporate such tools. The Reuters report and associated studies suggest that AI’s promise is real, but not evenly distributed—and that measuring, managing, and upskilling remain essential as organizations seek to realize its full potential. The next wave of AI advancements will almost certainly refine these tools further, but for now, the rush to automate everything should be tempered by a sober assessment of real-world results.