Industrial Series - Don't use LLMs
I hold a PhD in Computer Science and have been published in a variety of international peer-reviewed journals.
AI is going to be a problem. I don't know what will cause the first "big issue"; it might be from a courtroom where a defendant is sent to jail based off erroneous AI-generated data, it could be a death in a medical setting.. but, something is going to happen.
Let's take the existing adversarial AI research (there's been plenty) and make it useful.
I'm here to bring you up to speed.
As far as industrial engineering goes, I'm not saying don't ever use LLMs: I'm saying don't use them yet.
LLMs are good at text; they're bad with numbers. They're not particularly well suited to combinations of text and numbers as seen in logic problems.
(retrieved from Grok, 20250822) Notice the assumption of its knowledge of the problem. Notice the confidence.
Why does this matter for a chemical plant? Because industrial systems are full of similar logic problems: "If pressure in tank A exceeds X while valve B is closed and pump C is running, how do we prevent an explosion?". The response is the difference between normal operations and emergency shutdowns, or worse.
LLMs in particular are predisposed towards regurgitating training data rather than solve for unique circumstance. Appropriately handling unique circumstances is a serious safety issue. On AI safety - it's a perspective. It's a phrase that can mean different things to different industries. Asking a Google or Microsoft employee about AI safety, they'll likely talk about how the LLM can't say anything nasty (e.g., it can't be racist, inflammatory, etc.).
In a chlorine unit, safety means "let's not release pure chlorine into the atmosphere and kill everyone 20 miles downwind".
These aren't competing definitions—they're completely different universes of risk.
Right now, industrial operators I've interviewed share a simple philosophy: "never let an AI be in a position to affect the control board". I sure hope it stays that way. But, as commercial entities are beholden to boards and shareholders, this will inevitably change towards more "AI enabled automation". So the question isn't whether AI will enter critical industrial systems—it's whether we'll implement appropriate safeguards before it does.
So this series will look at use of AI in industrial settings. We'll look at directly introduced risk (poisoned models, cyber risks) and indirect risk (e.g., an engineer or operator asking assistance from an LLM). More importantly, we'll argue for increased oversight and proactive governance on usage of LLMs in critical industrial sectors to mitigate potential impact of LLM and ML usage. Nobody likes regulation - but unlike a chatbot that gives bad restaurant recommendations, industrial AI failures can have catastrophic impact.
