Two things compound in AI coding tools that don’t get talked about together.

The first is sycophancy. Half of the major coding assistants will agree with a wrong assumption rather than push back. It’s not a bug. It’s a training artifact. Models got rewarded for pleasant interactions, so they produce pleasant interactions.

The second is variable-ratio reinforcement. Sometimes the output is junk. Sometimes it’s a working function. Occasionally it’s a refactored module that would have taken you a day. Same architecture as a slot machine, accidentally.

The wins get cached as “I’m shipping faster.” The losses get rationalized. The reward signal stays positive even when the average outcome doesn’t.

That’s how the depth gap compounds silently between two engineers using the same tool. Wrote up the full argument, with the data, here:

https://joaofogoncalves.com/articles/2026/05/2026-05-04-ai-as-the-great-filter/