Last updated: September 2026
Most guidance about AI in engineering is about what the tools can and cannot do. This is about something different: the situations where the right call is to not use AI at all, even when it could produce something plausible. That is a judgment call, not a capability gap. An AI can draft a passable-looking calculation for a load-bearing member, and you should still not let it, because the decision to rely on that output carries a professional and legal weight the tool cannot hold. Knowing where those lines are is part of the job.
This is the situational companion to our guide on the limitations of AI in engineering, which covers what the tools get wrong; here we cover when to hold back regardless. It sits under our roundup of the best AI tools for engineers.
Short answer: Hold back from AI whenever the stakes exceed what you can verify or are allowed to delegate. The hard red lines: anything that will be stamped or sealed, because a seal certifies personal responsibility and an AI holds no license; any safety- or life-critical calculation you cannot independently check; confidential, client, or export-controlled data on a consumer tier; any output you have no primary source or expertise to verify; novel problems at the edge of the training data; and code or standards work where the exact edition and jurisdiction decide the answer. Deadline pressure is the trap that turns a green light red, because it tempts you to skip the check. AI is fine for low-stakes drafting you will review; it is not fine for anything you would have to defend.
When “not to use AI” is a judgment call, not a capability gap
The tools keep getting more capable, which makes the boundary harder, not easier: the question is less often “can it?” and more often “should I let it?” The answer turns on two things you control, whether you can verify the output and whether you are permitted to delegate the responsibility. Where both hold, AI is a fine assistant. Where either fails, the professional move is to keep AI out of the decision, no matter how good the draft looks. The categories below are where that line usually falls.
The red lines: where an engineer should hold back
Anything that gets stamped, sealed, or signed
A professional seal is a statement of personal accountability. Under the NSPE Code of Ethics, engineers “shall not affix their signatures to any plans or documents” not prepared under their direction and control (NSPE Code of Ethics). State boards make the weight explicit: the Florida Board of Professional Engineers states that by signing and sealing, a PE “verifies the authenticity of the document” and “accepts responsibility for its accuracy and legitimacy,” and may only seal work they were in responsible charge of preparing (FBPE). An AI holds no license and can be in responsible charge of nothing, so sealable work is a red line, as we detail in whether you can use AI for stamped drawings.
Safety- and life-critical calculations
The first NSPE fundamental canon is to “hold paramount the safety, health and welfare of the public” (NSPE). Any calculation whose error could harm people, a structural member, an electrical load, a pressure boundary, is one you must be able to verify independently, which means AI can assist the work but never own the answer. Our guides on whether AI can do engineering calculations and when to stop using AI for engineering calculations go deeper.
Confidential, client, or export-controlled data on a consumer tier
If the input is a client’s confidential material, unreleased IP, or export-controlled data, a consumer AI tier is the wrong place for it, full stop. This is its own topic, covered in AI tool data security and engineering IP and whether it is safe to use AI with client drawings; the short version is that the tier and its terms, not the convenience, decide.
Any output you cannot verify
If you have no primary source to check the answer against, or you lack the domain expertise to catch an error, you cannot safely use the output, because confident-wrong output is structural, not occasional. As a 2023-model analogy, a study of chat-model citations found 55 percent of GPT-3.5 references and 18 percent of GPT-4 references were entirely fabricated (Walters and Wilder, Scientific Reports, 2023). Researchers argue this is baked in: models “hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty” (Kalai et al., 2025, a preprint). The mechanism is covered in AI hallucination in engineering.
Novel or edge-of-domain problems
Models are weakest on exactly the problems that are not well represented in their training data. A 2025 benchmark of engineering tasks found that model accuracy declines as task complexity rises, degrades under minor perturbations, and remains substantially below human performance on high-level engineering work (EngiBench, 2025). For a genuinely novel design or an unusual edge case, that is the regime where AI is least reliable and most confident.
Code and standards work where edition and jurisdiction decide
When the answer depends on the exact edition of a code and the jurisdiction that adopted it, AI is risky, because it may cite a superseded edition or a rule that does not apply where you are building. The seal rules above are themselves jurisdiction-specific. Confirm the governing edition yourself.
Under deadline pressure, when you are tempted to skip the check
The most dangerous moment is when speed tempts you to accept an answer unverified. In the 2025 Stack Overflow Developer Survey, 66 percent of developers said their biggest frustration is AI answers that are “almost right, but not quite,” 45.2 percent said debugging AI-generated output is more time-consuming, and only about 3 percent highly trust AI accuracy while about 46 percent distrust it (Stack Overflow, 2025). The almost-right answer is the one that slips through under deadline, so pressure is the signal to verify more, not less.
A green-light, red-light decision guide
| Situation | Signal | Why |
|---|---|---|
| Draft an email, summarize a spec, brainstorm approaches | Green | Low stakes, easily verified |
| First-pass boilerplate or documentation you will review | Green | Verifiable by a competent reviewer |
| Interpreting a code clause without confirming edition and jurisdiction | Caution | AI may cite a superseded or non-applicable edition |
| Any calculation feeding a life-safety outcome | Red | Duty of care; you must verify it independently |
| Work that will be stamped or sealed | Red | The seal is personal accountability; requires responsible charge |
| Confidential or IP data on a consumer tier | Red | Wrong place for sensitive material |
| Output you cannot check against a primary source | Red | Fabricated, confident-wrong output is structural |
Why holding back is the professional’s call
Every red line above comes back to one principle: the engineer, not the tool, is accountable for the work. An AI cannot hold a license, cannot be in responsible charge, and cannot be sued for negligence, so the duty of care stays with you whether or not you used AI to get there. That is why “when not to use AI” is not a limitation of the technology but a decision you make. Used inside the green zone, AI is a genuine accelerator; pushed past the red lines, it is a liability wearing the costume of a finished answer.
Related: for the other side of this question, see when engineers should use AI.
Frequently asked questions
Is it ever unsafe to use AI in engineering?
Yes, situationally. Any output that feeds a life-safety calculation you cannot independently verify is unsafe to rely on, because the engineer’s duty is to hold public safety paramount and confident-wrong output is a known failure mode. AI can assist the surrounding work, but a qualified engineer must own and verify anything that could harm people if it were wrong.
Can AI sign off on or stamp a design?
No. Sealing a document certifies personal responsibility for its accuracy and legitimacy and requires being in responsible charge of the work, and an AI holds no license and cannot be in responsible charge. You may use AI as a tool in preparing the work, but the licensed engineer who seals it is the one accountable, so the seal can never rest on unverifiable AI output.
How do I know when I cannot trust an AI answer?
When there is no primary source to check it against, or you lack the domain expertise to catch an error. Fabrication and hallucination research shows confident-wrong output is structural rather than rare, so treat any answer you cannot verify as unusable. If you cannot independently confirm it, that is the signal to hold back rather than rely on it.
Is it OK to paste client drawings or specs into a chatbot?
Not on a consumer tier, and not with confidential, IP, or export-controlled material. The tier and its data-handling terms decide whether the input is safe, not how convenient the tool is. For sensitive work, use a business or enterprise tier with a no-training guarantee, and confirm the terms before you upload anything.
Sources
- NSPE Code of Ethics (Fundamental Canon 1; signature and competence rules)
- Florida Board of Professional Engineers, signing and sealing
- EngiBench, arXiv 2509.17677 (LLM performance on engineering tasks)
- Stack Overflow 2025 Developer Survey (AI trust and frustration)
- Walters and Wilder, fabricated-citation rates, Scientific Reports, 2023
- Kalai et al., Why Language Models Hallucinate, 2025 (preprint)
Written by the CognitiveFuture editorial team. We build our guidance from professional codes of ethics and licensing-board rules, peer-reviewed research, and large practitioner surveys, each linked above; NSPE is the code owner and we cite a verbatim access copy. We do not independently benchmark tools, and none of this is legal advice; the duty of care and any sealing decision rest with 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.
