ChatGPT Gave Me a Wrong Parts List. It Cost Me Real Money.

I asked a chatbot for a parts list, costs, and step-by-step build instructions for my hardware project. The answer was beautiful: organised, confident, itemised. I ordered from it. The prices were wrong, key instructions were wrong, and the components at the heart of the build were wrong for the job. The money was gone before the flaws surfaced, because the flaws only surface when physics grades your homework. This post is the autopsy of that answer — and the verification method that would have saved it.

Why language models invent part numbers

A large language model generates the most statistically plausible next token. Part numbers are exactly the kind of text where plausible and true diverge: "DRV8825" and a subtly wrong sibling look equally reasonable to a model that has read millions of forum posts but has never opened a distributor catalogue. The model is not lying — it has no concept of a catalogue at all. It is completing a pattern, and hardware punishes patterns that are 95% right.

The four failure modes of an AI parts list

• Invented or mismatched part numbers: components that don't exist, are discontinued, or are a different variant than the description implies.
• Hallucinated prices: training data is months to years old, and prices for motors, sensors, and chips move constantly. A BOM total from a chatbot is a guess presented with two decimal places.
• Spec mismatch: the part exists, the price is even right — but nobody calculated whether its torque, current rating, or voltage matches your project, because you never gave the model numbers to check against.
• System incompatibility: each part is individually fine, but the driver can't supply the motor's current, the sensor speaks a protocol your board doesn't, or the battery can't survive peak draw. Compatibility lives between the lines of a BOM, and that's exactly where a text predictor doesn't look.

What LLMs are genuinely good at in hardware

This is not an anti-AI post — I'm building an AI company. Chatbots are excellent at explaining concepts, comparing architectures, walking through the torque formula, and turning your vague idea into a list of questions you should answer. They are a brilliant tutor and a dangerous purchasing agent. The failure is not using AI; it is letting a text predictor do a catalogue-and-physics job without grounding.

The 5-step verification for any parts list — AI or human

• Existence: paste every part number into a real distributor (Digi-Key, Mouser) — the part must appear, in stock, matching the description.
• Price: replace every price in the list with today's live distributor price. Expect the total to move; budget on the new number.
• Datasheet: open the datasheet for every part over $20 and find the one line that matters for your project — torque curve, continuous current, operating voltage.
• Math: write your requirement as a number (load, current, runtime) and check the datasheet number clears it with a 30% margin. One page of arithmetic.
• System pass: walk the power chain end to end — supply → regulation → drivers → loads — and confirm every link supports the peak demands of the link after it.

If a parts list survives those five steps, it is safe to order — no matter who or what wrote it. If any step fails, the list just saved you the exact money it would have burned. This verification is precisely the work we now do by hand for early founders: a bill of materials with real part numbers, live prices, linked datasheets, and the math written out for you to check.