Slop is a judgment failure: nobody noticed a choice was being made

Slop, the low-quality AI output people complain about, is not mainly a machine problem. Gabriel Martinez, an engineering manager at G2i, argued on the AI Engineer show that it is a human one. Nobody noticed that a decision was being made. AI widens that gap by producing more unexamined work, faster.

What was said

Martinez said slop is not AI-generated code. He defined it as work that looks finished before the thinking is. In such work, unclear points get settled by guesses and trade-offs are left to chance. His central claim: "The danger is that the engineer never noticed that there was a choice to make." The engineer never met the ambiguity or the constraint, so the work arrived already smoothed over.

He described software as a record of technical, product and operational choices. If people cannot see the choices, he said, they cannot own them.

He gave an ordinary example. He hands someone ten bullets of scope and gets back a 20-page document. It reads as polished, but he then spends hours checking whether real thinking sits inside. His summary: "Generation is cheap, but evaluation is expensive." He also pointed to pull requests (proposed code changes) of thousands of lines, which he said no human can fully understand. He admitted that anxiety about team speed once led him to merge code with less review than he should have given it.

Why it matters

For engineering leaders, that points at the review step rather than the writing step. Martinez said rewards should go to value, sound system design and clear tests, not to how fast code gets merged. If a team praises volume, it pays people to skip the noticing.

For buyers of AI-assisted work, the question is plain. Who owns this change, and who saw the choices inside it?

The other side

Martinez said he is not anti-AI. He also said not every line of code must be read by a person. What must be understood, in his view, are the main flows, the boundaries and the data model. That line is not sharp, and the excerpts do not say how a team should draw it.

His evidence is his own experience and examples. He gave no data on how common slop is. He noted that engineers who skip careful review often seem to produce more. The excerpts describe that incentive problem, but they do not show it solved.