Senior Devs Ship More AI Code, Feel Faster

Fastly’s survey shows senior developers both use and trust AI coding tools more than juniors, with more of their shipped code AI-generated and stronger claims of speed gains. Yet many developers frequently rework AI output, and an external RCT suggests AI can even slow experienced programmers, indicating a perception–reality gap. Despite mixed efficiency, AI improves enjoyment, and sustainability awareness—especially around AI’s energy footprint—is high.

Key Points

Senior developers ship much more AI-generated code and report larger speed gains than juniors (32% vs. 13% ship >50% AI code; 26% vs. 13% say AI makes them a lot faster).
A significant share of developers frequently rework AI output (28%), often offsetting perceived time savings; only 14% rarely need changes.
External evidence complicates the perceived gains: an RCT found experienced developers took 19% longer when using AI tools.
AI tools increase job satisfaction for most developers (nearly 80%), even when efficiency gains are mixed.
Awareness of energy use and AI’s carbon footprint is high, especially among more experienced developers; sustainability considerations are increasingly common.

Sentiment

Mixed and polarized: pragmatic agreement that AI helps with boilerplate and prototyping under tight supervision, combined with skepticism toward the survey’s framing and strong criticism of AI’s inconsistency, quality issues, and overhype.

In Agreement

Senior developers can get more from AI because they know where to delegate: use it for boilerplate, scaffolding, small functions, tests, refactors, API integrations, and documentation lookup.
Perceived speedups are real in narrow, well-scoped tasks; AI reduces cognitive load and helps teams ship dull or repetitive parts faster.
Meaningful time is still spent reviewing and fixing AI output, often offsetting some of the gains—aligning with the survey’s rework caveat.
Productivity gains are uneven and highly task-, tool-, and language-dependent; success comes from decomposing work and supervising like a junior dev.
External evidence showing AI can slow experts (e.g., RCTs) matches lived experiences where later debugging negates early ‘smooth’ autocompletion.

Opposed

The survey is marketing blogspam without methodological rigor; ‘AI-generated’ is undefined (does tab-complete count?), and the sample may be biased or too small.
AI code quality is often poor or inconsistent; reviewing AI output can be harder than writing from scratch, leading to technical debt and ‘organizationally poisonous’ outcomes.
For complex features, legacy codebases, or ecosystems like C/C++/Zig/Rust, AI frequently hallucinates, forgets context, and slows teams due to heavy supervision.
Over-reliance risks eroding developer skills and understanding; reducing cognitive load can harm long-term competence (some cite cognitive decline concerns).
Environmental and monetary costs (e.g., high daily spend) and managerial pressure to ‘use AI’ are problematic; the benefits are overhyped relative to real-world results.