On what basis did you sign? A calculation the AI got wrong
Published
What the tool produced
An engineer asks an AI tool to calculate a load, a flow rate or a stress. The tool writes a formula, substitutes values and returns a number. The number enters a report, and the report is signed.
Months later an auditor asks who approved it and on what basis. There is no documented answer.
The calculation looks excellent. It was also wrong.
Why it looked right
A direct calculation inside the model fails in three places: the arithmetic, the choice of formula and the unit conversion. Each of them looks valid in the output line.
A study of two national licensing exams for mechanical engineers in buildings classifies 172 errors of one tool. Of them, 104 were knowledge errors. 48 were reasoning errors, in calculation questions. 20 were context errors, in questions that require reading a drawing.
Another important finding: the answer consistency was 97 percent. A wrong answer stayed wrong on the next attempt. Asking again is not a check at all.
This is what people search for as AI hallucinations engineering calculations: a formula that looks right, with a wrong number inside.
How the engineer catches it
The first rule: the tool drafts and does not calculate. It may suggest a formula and identify variables. Run the arithmetic in a spreadsheet or in a deterministic calculation tool, not inside the model.
The second rule: every value that came from the tool is a draft until it is checked against a source. A source is a manufacturer manual, a data sheet, a standard in its official wording or a direct measurement.
The third rule: the signature is a commitment. A commitment is any action that moves a deliverable from draft to reality. A signature, a submission, a purchase order, a work instruction, a report to a customer and an entry into an organizational system are included.
Only a deliverable may be signed. A deliverable is an output where every item was checked or carries the missing mark, with a full log.
In the request to the tool, state constraints: units, the binding standard and length. A request without constraints is not a request by the method. The number that returns may enter a design decision, so it must rest on data that was supplied, and not on a guess.
Even a calculation that came out right may arrive with faulty reasoning. This is a G7 failure, and it passes a source check. It is caught by the estimate you wrote before the tool.
The check to run
The log is the answer to the auditor's question. For every calculation, record the following fields.
- What the tool said, word for word.
- What source it was checked against, and where in it: page, clause or table.
- What the result is: match, gap, not verified, or missing.
An item that was verified and matches the source passes, and you record the source and the location. A gap is recorded with the original output, the source, the correction and the gap type. In our case the type is direct calculation.
An item that cannot be verified is blocked. If the deliverable depends on it, the deliverable stays a draft. There is no partial verification.
Before all this comes an estimate. Write the order of magnitude of the answer in advance, before you open the tool. An estimate made after you saw the answer is anchored to it.
What the auditor wants to see
An auditor does not look for perfection. He looks for a record that can be counted. So the log is built from fixed fields. Five of them are closed lists: discipline, deliverable type, item type, result and gap type.
Closed lists make the log countable. You can ask how many calculations were checked in a quarter. You can ask how many gaps were found, and of what type.
This yields an answer to the question the auditor will ask anyway: who approved, against which source, and when.
A source is a document with an owner and a standing. A manufacturer manual qualifies. Output from another tool does not.
What the check does not do
The check does not replace engineering knowledge. A simple calculation that enters a signed inspection report is still a task at use level zero, meaning no tool, until someone tags it otherwise. The level is set by the consequence of an error, not by the difficulty.
The document behind the method has not been peer reviewed. It measures failures, not the method.
The first step
A small start is better than a big procedure that never gets going.
Choose one report that has not yet been signed. Take from it the three numbers that matter most for the decision. Check each of them against a source, and record the result in the log.
If one of them is not verified, the report stays a draft and is not signed. It takes a short time, and it is the way to a documented answer to the auditor. Such a habit also saves a late argument about responsibility.
Self-check
- Did the calculation run in a spreadsheet and not inside the tool?
- Were the conversion units checked against a manufacturer table?
- Does the log record what the tool said, word for word?
- Is there a source and a location for every value that entered the report?
- Was an estimate written before the tool was opened?
- Were gaps recorded with a type, for example direct calculation?
- Did the deliverable stay a draft when one item was not verified?
- Can you answer the auditor on what basis the number was signed?
Do you have a report waiting for a signature? Get in touch.
Source: Choi, H.; Lee, J.; and Kim, J. (2025). Performance Evaluation of GPT-4o on South Korean National Exams for Building Mechanical Equipment Maintenance. Scientific Reports, 15: 30436. doi 10.1038/s41598-025-16118-x