Generative and Agentic AI Shift Concern from Tech Debt to Cognitive Debt
margaretstorey.comI feel like no one talks about how people who are "supposed to be reviewing the LLM outputs, guiding the agents, etc.", actually acquire the knowledge to do a decent job. When I see discussion on LLMs making SWEs more productive, I assume they mean compared to someone who knows significantly less than they do.
Here is a real scenario encountered in corporate America:
A new CS grad in their first job after college is given a task in a domain they're unfamiliar with to solve a problem they've never seen. They ask an agent to implement it in code in a language they've never used, and have it give a breakdown of it's process, the tradeoffs encountered, things to consider in the future, etc.
They have no time to actually learn anything, because they should be moving faster, being more productive, AI has already solved all of our problems.
So they submit a 700+ line PR to which whoever reviews it just pushes it through because they don't have time as they need to be moving just as fast and do not have the cognitive capacity to sit through and comprehend what's actually happening.
I recently read an article acknowledging the issue of this kind of debt, but then suggesting, “you need more review, not less AI”. I laughed at the naïveté of the author. More review is simply not going to happen, and as you’ve pointed out, who is going to do the reviewing?