The economic case for AI coding tools has rested on throughput: generate more code, ship more features, move faster. A Golem analysis of software development takes aim at that assumption, arguing that what the industry is actually getting is more code and less progress.
The argument is not that generated code is worthless. It is that volume was never the right measure. For decades code was expensive to produce, so its quantity served as a rough proxy for effort and, loosely, for progress. When generation collapses that cost, the constraint does not disappear — it moves downstream, to the parts of engineering that were already scarce: reading code, deciding whether it is correct, understanding why it is shaped the way it is, and living with it for years.
That shift has consequences teams already recognise. Review becomes the throughput limit: a reviewer facing a thousand lines of plausible-looking generated code is not faster than one facing a hundred handwritten lines, and is often slower. Comprehension debt accumulates — code nobody wrote, nobody fully read, and nobody can confidently change. Tests can stay green while the design quietly rots.
There is a genuine counterargument, and Golem's framing leaves room for it: tooling, tests, static analysis and model quality are all improving, and debt accumulated today may be cheaper to repay tomorrow. But that is a bet on future capability, not evidence of present progress — and the return on it has to be measured, not assumed.
The uncomfortable reading of the piece is that the industry has optimised the cheapest stage of the pipeline and called it acceleration. Measuring progress by code volume under AI assistance is roughly as informative as measuring a hospital by the number of prescriptions written.
Golem's full argument, with its own evidence and examples, is in the original article.
