A headline number claiming AI writes 80% of code at top firms sounds like a countdown to obsolescence. It is actually measuring keystrokes generated, not decisions made, and those are separate jobs entirely.
Why it matters: Engineers reading that statistic as a threat to their careers are answering the wrong question, and the real shift, from typing code to judging it, demands a skill most teams never trained anyone to build.
The Details:
- The percentage assumes a fixed amount of total code, but cheaper generation means teams attempt work they'd never have justified before: throwaway scripts, edge-case tests, risky refactors. Some share of that 80% is new output filling gaps, not old work getting reassigned to a model.
- High generation numbers show up almost exclusively in codebases with strong testing, clean conventions, and changes scoped small enough to verify quickly. The impressive figure is really a scorecard for prior engineering discipline, not proof that judgment became optional.
- Review now carries the weight that writing used to carry. A change that took an afternoon to type might take ten minutes to generate and forty to check properly, and the failure mode showing up across teams is reviewing AI output at typing speed instead of reading speed.
- Junior engineers traditionally built intuition by writing simple, well-understood code first. If that easy work disappears into automation, teams need a deliberate substitute: pairing juniors on reviews, logging caught failure modes, and still assigning hard problems with the tool switched off.
Bottom Line: The 80% figure doesn't shrink the engineering job. It relocates the hard part from generation to specification and review, and that relocation is a harder problem to solve than the number lets on.
