A July 2025 randomized controlled trial found that experienced developers using AI coding tools completed tasks 19% slower than a control group. The counterintuitive result carries real implications for how engineering leaders assess, hire, and develop genuine AI fluency.
The headline from a July 2025 randomized controlled trial should stop any engineering leader mid-scroll: experienced open-source developers who used AI coding tools took 19% longer to complete tasks than developers who did not. Not 19% faster — slower. In the METR study, developers worked on their own repositories — code they already knew deeply — and still lost ground to the control group.
The METR researchers themselves note this is a snapshot of early-2025 capabilities, and that the tools are evolving rapidly. But for engineering managers making headcount, training, and hiring decisions right now, the result is a critical signal: tool access is not the same as AI fluency, and the market is already pricing the difference sharply.
The Research Gap Between Tool Access and Genuine Fluency
The METR study's design matters. These weren't junior developers still learning an unfamiliar codebase — they were experienced contributors to open-source repositories they had built themselves. They still slowed down. The most plausible interpretation: AI tools introduce real friction in the form of context-switching, prompt iteration, output verification, and debugging AI-generated errors. Until an engineer has genuinely internalized how to work with these tools — not merely that they exist — the overhead can outweigh the benefit.
This distinction is urgent for engineering organizations today. The current tech layoff wave has been characterized explicitly as structural, not cyclical — companies are cutting traditional engineering roles and redeploying hiring budgets toward AI development and automation. Mid-level engineering, sales, and support roles are disproportionately affected; AI, data science, and cybersecurity roles are ramping up in their place. Teams are being rebuilt around AI fluency as a real operational competency, not a checkbox. Simply deploying Copilot or Cursor across a team is not a workforce transformation strategy.
What the Market Is Already Signaling
The labor market has registered this distinction clearly. Indeed Hiring Lab data shows U.S. machine learning engineer openings sitting roughly 59% above their February 2020 baseline — while general software engineering postings have fallen approximately 49% below that same baseline. That is a 108-percentage-point spread between engineers who build AI systems and engineers who simply have access to AI tools.
The compensation gap is equally sharp. Robert Half's 2026 Salary Guide puts the national range for software engineers at $109,250–$175,500. For AI/ML engineers, the floor rises to $134,000 — roughly $25,000 higher at the entry point — with the ceiling reaching $193,250.
| Role | Salary Range (Robert Half 2026 Guide) | Floor vs. SW Eng. |
|---|---|---|
| Software Engineer | $109,250 – $175,500 | — |
| AI / ML Engineer | $134,000 – $193,250 | +~$25,000 |
LinkedIn named AI engineer the #1 fastest-growing U.S. job for the second consecutive year — a sustained structural shift, not a spike. Apple, Google, and TikTok lead major tech employers in the volume of AI engineering openings, with Google advertising roughly 62% more engineering roles year-over-year. Beyond Big Tech, the fastest-growing demand is concentrated in fintech, observability, and security.
What Hiring Managers Should Do Differently
The practical upshot is clear, even if execution isn't easy: evaluate AI fluency through evidence of shipped work, not self-reported tool familiarity.
Ask candidates to walk through a project where AI tooling failed them — what broke, what they debugged, what they shipped anyway. Ask what they didn't trust the AI to generate. Engineers who have genuinely internalized these tools will have specific, nuanced answers about the failure modes and the workarounds. Engineers with surface-level exposure typically won't.
For teams that want to develop AI fluency internally rather than hire for it externally, the same challenge applies: fluency is trained, not installed. Ryzlink runs structured training programs designed specifically to turn capable engineers into AI-fluent, forward-deployed engineers — built on the premise that tool access without deliberate practice rarely moves the needle, a premise the METR data now supports directly.
The market will keep paying a steep premium for engineers who can actually demonstrate AI fluency at the level of shipped production work. The first step for any engineering leader is being precise about what that term means — and honest about what handing out tool licenses does not.
Comments
No comments yet — be the first to share your thoughts.