Last updated: August 2026
Being handed a colleague’s calculation package to check is a job with your name on the outcome, and the fastest way to do it badly is to read back through their arithmetic and nod along. A real check is an independent verification, not a proofread, and that distinction is the whole discipline. This guide covers how to check someone else’s engineering calculations properly, where AI can genuinely help a checker and where it is dangerous, and why the accountability stays with you. It pairs with our method for how to verify an AI engineering answer and the pillar on the best AI tools for engineers.
In one line
Do not re-read the other engineer’s working; reproduce the result independently, ideally by a different method, then check the assumptions and inputs, the units and dimensions, the applicable standards, and the order of magnitude. Independence from the original author is what makes it a check. AI can help you understand a method or spot an inconsistency, but it cannot be the independent recomputation and it cannot carry the responsibility. That stays with the licensed engineer in charge.
What a real check actually is
A check is not confirming that the author’s steps follow from each other; it is confirming that the answer is right, arrived at from the inputs by a route you did not copy. The defining feature is independence: a good check recomputes the result separately, so that a shared assumption or a repeated slip cannot survive in both versions. If your check simply retraces the original calculation, it will inherit the same blind spots. That is the difference between reviewing and verifying, and it is the same principle behind our guide on whether AI can do engineering calculations at all.
The five layers of an independent check
Work through these in order, and treat any one that fails as a stop, not a note for later.
- Recompute independently. Re-derive the result yourself, ideally with a different method, model, or software than the author used, rather than re-reading their numbers. Agreement between two independent routes is the strongest evidence a result is right.
- Check the assumptions and inputs. Confirm the loads, material properties, boundary conditions, and source data are correct and justified. Most real errors live in a wrong input, not a wrong sum.
- Check units and dimensions. Run dimensional analysis on every expression. It cannot prove an answer is right, but it can prove one is wrong, which is exactly what you want from a fast filter (University of British Columbia).
- Check the standards. Confirm the applicable code, the correct edition, and the specific clauses the design relies on. A calculation can be internally perfect and still fail the governing standard.
- Sanity-check the magnitude. Ask whether the result is physically plausible. An order-of-magnitude estimate catches the errors that pass every other test.
How independent should the check be?
Industry practice recognizes that not all checks are equal, and the level should match the consequence of an error. A useful way to think about it is a ladder of independence.
| Level | Who checks | Use it for |
|---|---|---|
| Self-check | The author reviews their own work | Low-consequence, routine work |
| Independent check | A different engineer in the same team | Most design calculations |
| Independent verification | A separate organization, own model and software | High-consequence or safety-critical work |
Where AI helps a checker, and where it must not
AI is a useful assistant to a checker on the language and comprehension parts of the job, and a liability on the arithmetic. The reason is structural: research shows language models struggle to detect errors, because their internal validation relies on surface-level numerical alignment and happens before the computation is complete, so they miss even simple mistakes (Bertolazzi et al., arXiv 2502.11771). It also gets units wrong for the reasons in our guide to why ChatGPT gets units wrong. So it can never be the independent recomputation.
| AI can assist with | Never delegate to AI |
|---|---|
| Explaining an unfamiliar method | The independent recomputation |
| Re-deriving a formula symbolically | Confirming a number is correct |
| Flagging a narrative-versus-numbers mismatch | Detecting the arithmetic errors |
| Suggesting a standard clause to look up | The final sign-off |
Used as a prompt for your own judgment, it speeds the check up; treated as the checker, it fails silently, which is why the broader limitations of AI in engineering matter here, and why you still use ChatGPT for engineering only as an aid.
Who is accountable for the check
Checking is not a clerical step; it carries professional responsibility. Under the model law that most licensing boards follow, an engineer must be in responsible charge, defined as direct control and personal supervision of the work (NCEES Model Law). Merely reviewing drawings after they were prepared, without genuine involvement, does not meet that bar (NSPE). No tool changes that. Whatever software or AI you used to help, the licensed engineer who checks and signs remains personally accountable for the result.
Frequently asked questions
How do I check someone else’s engineering calculations?
Reproduce the result independently rather than re-reading their working, ideally by a different method, then verify the assumptions and inputs, check units and dimensions, confirm the applicable standard and edition, and sanity-check the magnitude. Independence from the original author is what makes it a genuine check.
What is the difference between checking and reviewing a calculation?
Reviewing confirms the author’s steps follow from one another; checking confirms the answer is correct by an independent route you did not copy. A review can pass a calculation that shares a hidden wrong assumption, which an independent recomputation would catch.
Can I use AI to check engineering calculations?
Only as an aid, never as the checker. AI can explain a method, re-derive a formula symbolically, or flag an inconsistency, but research shows it struggles to detect its own arithmetic errors, so it cannot be the independent recomputation or confirm that a number is right. Do that yourself or with a deterministic tool.
How independent does a calculation check need to be?
Match the independence to the consequence. Routine work may need only a self-check, most design calculations need a different engineer, and high-consequence or safety-critical work needs a separate organization using its own model and software so no shared error survives.
Who is responsible if a checked calculation is wrong?
The engineer in responsible charge, meaning the one with direct control and personal supervision, remains accountable, and the checker who signs off shares that responsibility. Using AI or any software does not transfer the accountability away from the licensed engineer.
Sources
- Language models struggle to validate their own arithmetic: arXiv 2502.11771
- Dimensional analysis as a check: University of British Columbia
- Responsible charge, direct control and personal supervision: NCEES Model Law
- Reviewing after preparation does not satisfy responsible charge: NSPE
About the author: this guide was written and edited by the CognitiveFuture editorial team, which researches how AI tools fit real professional workflows. We cite primary sources for the claims we make and update our recommendations as tools and standards change. We do not test products ourselves; our assessments synthesize official documentation, primary research, and practitioner reporting.
Professional practice and standards vary by jurisdiction. Always confirm the requirements that apply to your work and route load-bearing decisions to a licensed engineer.
Richard Johnson
Richard Johnson is an AI specialist at one of the world's largest technology companies, where he has spent the past three years helping organizations adopt AI. CognitiveFuture extends that work publicly: gathering the available evidence on each tool, from vendor documentation to independent reviews and user feedback, and cutting a crowded market down to the right choice for the job in front of you.