Things that AI is bad at, that engineers do without thinking
Three things AI is bad at, that engineers do without thinking
I have spent a lot of time over the past year working with AI on engineering problems. Some of it is genuinely useful, some of it is so-so and some of it is a mess, that ends up consuming way more time than it should have done and I could have done it better myself in a fraction of the time. The patterns of what AI is good at and what it is bad at have become much clearer.
Layout and Visual Reasoning. When I build a model in PSCAD or PowerFactory by hand, I spend a surprising lot of time laying out the wires and nodes carefully and trying to replicate the structure of the SLD I am working from. I do this so that I can debug the model later, follow the signal flow at a glance, and explain it to someone else. The AI approach is much more haphazard. Wires cross. Components sit at odd angles. Even when given explicit rules about layout, the result tends to look messy in a way that bothers me. Conceptually this makes sense. The AI sees data paths and data flows. It does not really think visually. Getting it to structure a graphical layout nicely is surprisingly hard and time-consuming, and an area where AI really does not perform anything close to a human.
Knowing when to stop. An experienced engineer recognises what good looks like. I can eyeball a fault calculation or a relay protecting study and know if its about right without having to check everything line by line. Take for example a fault level analysis, I know from experience that one decimal point in kA is good enough; looking at multiple decimal places is pointless as the input data wont be that accurate. The same with tuning a PID control loop, I can see when the response ‘looks goo’ AI does not have this instinct. Left to itself it will keep going, keep testing, keep refining, well past the point where the additional work has stopped adding value. It does not get tired and it does not get bored, which sounds like an advantage and sometimes is. More often it is a sign that the engineer needs to step in and say enough.
The boring sanity check. Did the model initialise cleanly? Does the load flow converge from a different starting point? Are the active and reactive power flow directions right? Have I accidentally turned something off, or opened a breaker that is going throw everything off? Are the steady-state values sensible?
When I do a study the small dull checks are the first things I look at, because they catch the silly errors that destroy the rest of the analysis. AI tends to skip them. It chases the dramatic edge cases and the headline result. It does not naturally pause to check the obvious and think that its output is right and starts looking for exotic reasons and justifications for the odd behaviour, rather than stopping and thinking ‘hold on a minute…’.
Summary
The pattern is consistent. AI is good at the things that look impressive, and sifting data, but it is not good at many of the simple things that experienced engineers do without conscious thought, the habits that come from years of catching their own mistakes.
This is not an argument against AI in engineering. It is an argument for being clear about which parts of engineering it can and cannot help with. The bookkeeping, the orchestration, the writing-up, the structured search through possibilities - these are real value. The aesthetic judgement, the knowing-when-to-stop, the reflex to sanity-check - these are still the engineer’s job.
A good agentic loop has the AI doing the work and the engineer doing the watching. Both roles are essential for future power systems engineering.