Can ChatGPT do structural calculations?

Ask a modern chatbot to design a beam and it will oblige — inputs restated, method named, equations typeset, a confident answer at the bottom. It looks like a calculation package. The question that matters professionally is narrower: can you trust the numbers, and could you or a reviewer reproduce them? On the published evidence, the answer to both is no.

What the trials show

In STRUCTURE magazine's December 2025 feature "AI vs. Hand Calculations," Dowell and Shajihan gave GPT-5 five statically indeterminate structural problems, running each one three times. Their finding: the model "often gives the wrong answers, even on repeat attempts for the same problem that it got correct on another attempt," and "sometimes the results were off by a lot." Same problem, same wording, different day — different answer. The failure isn't just accuracy; it's non-determinism. A large language model is a probabilistic text generator, and probabilistic is exactly what a structural calculation must never be.

Why this happens

None of this is a bug that the next model version cleanly fixes, because it follows from what the tool is:

• LLMs predict tokens; they don't solve equations. When the arithmetic comes out right, it's because the right continuation was also the statistically likely one. Multi-step problems — moment distribution, compatibility equations, anything indeterminate — multiply the chances that one step drifts.
• Errors arrive fluently. A wrong stiffness coefficient is presented in the same confident, well-formatted prose as a right one. There is no wobble in the voice to warn you.
• Citations can be plausible rather than real. A chatbot can attach a section number that looks like the right neighborhood of the code and isn't. Every clause reference needs checking against the book — at which point you're doing the lookup anyway.
• The same prompt does not guarantee the same output. Which means a reviewer cannot re-run your "calc" and get your numbers — the basic contract of a checkable calculation is broken from the start.

The liability angle is not hypothetical

Naser's March 2026 STRUCTURE analysis of machine learning's legal posture draws the line in terms every practitioner should memorize: engineering work stays under the professional-service standard when "any competent structural engineer could reproduce the entire design process from first principles" using the documented assumptions, code references and rationale. Output that can't be reproduced from its documentation drifts toward product liability territory — a much worse place to be standing. An AI transcript whose numbers change on re-run fails the reproducibility test by construction, no matter how good the transcript looks.

Where language models genuinely help

Dismissing the technology entirely would be as lazy as trusting it blindly. Surveys of practicing engineers (Pathak's 24-engineer survey in STRUCTURE, April 2026, among others) land on a consistent posture: AI as a bounded productivity aid, with traceability and verification non-negotiable. In practice that looks like:

• Explaining and locating — "what is §12.8.7 about, in plain terms?" — as a pointer to the text you then read yourself.
• Drafting prose — basis-of-design narratives, transmittal language, report boilerplate — where the engineer's review is the whole workflow anyway.
• Scripting help — writing the Python that post-processes your analysis output, which you then test against a known case.
• Document drudgery — extracting schedule data, summarizing a geotech report, flagging inconsistencies between drawings — with each flagged item verified at the source.

What's common to every legitimate use: the model proposes, a deterministic check disposes — either the engineer's own verification or real computation.

The tool shape that actually works

This is the thesis StructuralClaw is built on, so take the vested interest as read — but the logic stands on its own. A calculation tool an engineer can defend has three properties a chatbot structurally lacks:

• Determinism. Same input, same output, every time. StructuralClaw's engines are compiled, tested code — not sampled text.
• Citations on values, not vibes. Every value in the calc trace names its code section and equation — ACI 318, AISC 360, NDS, ASCE 7 — so checking a number means opening the book to the cited clause, not reverse-engineering a paragraph.
• Reproducibility by hand. The printed trace shows the substituted equation, so a reviewer can follow it with a pencil. That is exactly the documented-assumptions standard the liability analysis asks for.

And the discipline cuts both ways: deterministic software deserves skepticism too. Benchmark any tool — ours included — against hand calcs and published verification examples before you rely on it. Software that shows its work makes that audit an afternoon instead of an act of faith; software that hides it is asking for trust no engineer should extend, whether the black box is a neural network or a vendor's binary.

Bottom line

Can ChatGPT do structural calculations? It can perform the appearance of one. For the numbers themselves, the published trials say the reliability isn't there, and the professional standard — reproducibility from documented assumptions — is something a probabilistic model cannot offer even in principle. Use language models for language. Use deterministic, code-cited computation for loads and capacities. And keep the engineer's judgment, and the engineer's pencil, in the loop for both.

• Dowell & Shajihan, "AI vs. Hand Calculations," STRUCTURE, December 2025.
• Naser, "Thought Experiments for Understanding the Legality of Machine Learning in Structural Engineering," STRUCTURE, March 2026.
• Pathak, "Cloud and Browser-Based Structural Analysis: Unlocking Practical Automation and AI-Ready Workflows," STRUCTURE, April 2026.
• Stasulis, "Harnessing the Power of Software Without Losing Sight of Fundamentals," STRUCTURE, October 2024.