Where AI earns its place, and where it has none

Where it has no place at all

Not 'use with caution': no place. These are the outputs where a plausible wrong answer is indistinguishable from a right one at review speed, and where being wrong is not recoverable.

• Any number that becomes a design value. Loads, capacities, ratings, sizing, tolerances, setpoints. Module 2 is entirely about why.
• Any code or standard requirement stated as fact. Clause numbers, limits, factors, applicability. Module 3 is entirely about why.
• Material property values, allowable stresses, derating factors: taken from the published source, every time, never from a chat window.
• Any judgement about whether a design is safe. That judgement is what your licence, your title, and your signature exist to represent.
• Anything going into a submission, certification or sign-off that you have not independently verified line by line.

The middle ground, and how to work in it

Most real work sits between the two lists, and the workable rule is that AI may help you think and may not tell you the answer. Ask it to enumerate failure modes for a connection detail, and you review the list against your own judgement and the code: useful, because a list is easy to audit and an omission is easy to spot. Ask it what the connection capacity is, and you have crossed into the second list, because a number is not auditable by inspection.

Anchoring is the quiet risk in the middle ground

Worth naming early because it survives good intentions. Once you have seen a number, your own subsequent calculation is measured against it, if your hand calculation lands near the suggested figure you relax, and if it lands far away your first instinct is to hunt for your own error rather than to doubt the suggestion. This is ordinary human cognition and knowing about it does not switch it off. The only reliable defence is not asking for the number in the first place.