The Limitations of AI in Engineering (and How to Work Around Each One)

Last updated: August 2026

AI is a genuinely useful assistant for engineering work, and it has clear limits that show up in predictable places. Knowing them is what separates engineers who get real value from AI from those who get burned by a confident wrong answer. This guide walks through the seven limitations of AI in engineering that matter most, each grounded in published evidence, and the practical workaround for each. For the tools themselves, start with our pillar guide to the best AI tools for engineers.

The limitations of AI in engineering, at a glance

The limits are not random. They follow from what a language model actually is: a system that predicts likely text, trained on the internet, with no physical model of the world and no accountability. That single fact explains why it is fluent and unreliable at the same time. The seven sections below take each limitation in turn, with the evidence and the fix.

1. It calculates by predicting, not computing

A language model generates text that looks like the right answer rather than running the arithmetic, which is why it makes unit and sign errors that read as confident. Giving a model a real calculator changes everything: allowing GPT-4 to write and execute code lifted hard-math accuracy from 53.9 percent to 84.3 percent (arXiv 2308.07921). The failure is structural, down to how numbers are split into tokens (arXiv 2505.14178), and it shows up in engineering specifics: unit and measurement conversions are a distinct weak point (NUMCoT, ACL 2024), and a tuned model scored 82 percent on statics yet still misidentified tension and compression in truss members (Hope et al., arXiv 2502.00562).

Workaround: route every calculation to a deterministic engine, a code interpreter, a spreadsheet, or a symbolic solver, and require explicit dimensional analysis. Our guide on whether AI can do engineering calculations covers the full workflow.

2. Its reasoning is fragile

What looks like reasoning is often pattern matching, and it breaks under small changes. On GSM-Symbolic, adding a single clause that seems relevant but changes nothing dropped model accuracy by up to 65 percent (Mirzadeh et al., arXiv 2410.05229). On EngiBench, a tiered engineering benchmark, accuracy declined with task complexity, degraded under minor perturbation, and stayed substantially below human experts (Zhou et al., arXiv 2509.17677). The lesson is that a correct answer on a textbook problem does not promise a correct answer on your problem.

3. It hallucinates facts, standards, and citations

Because a model produces likely text rather than retrieving verified facts, it will invent plausible-looking standards clauses and references. A Stanford study of general-purpose models found hallucination on 58 to 88 percent of legal-reference queries, with ChatGPT-4 at 58 percent, PaLM 2 at 76 percent, and Llama 2 at 88 percent across more than 800,000 queries (Dahl et al., Journal of Legal Analysis, 2024, testing 2023-era general-purpose models). A systematic review of language models in engineering reached the same practical conclusion: reliability is the core barrier, and outputs need rigorous validation before deployment (Frontiers in Education, 2025).

Workaround: require source grounding against a controlled standards library and verify every quoted clause, code section, or reference against the primary document. Never trust a cited standard number at face value.

4. It reads technical drawings unreliably

Multimodal models can describe a drawing, but they are not a reliable source of truth for it. The best vision models score only around 60 to 70 percent on mechanical-drawing comprehension and stumble on tolerances, datums, and multi-view integration (MechVQA benchmark, 2025). The problem is real enough that researchers had to fine-tune a dedicated vision model on hundreds of expert-annotated drawings to beat general models at extracting GD and T information (Springer, 2025).

Workaround: keep a human interpreting any drawing whose tolerances or geometry carry risk, and if you upload one, read our guide on what happens when you upload CAD files to ChatGPT before trusting the readout.

5. It has no physical judgment and cannot be accountable

A model has no physical model of the world and cannot carry professional responsibility. Only a licensed engineer in responsible charge, meaning direct control and personal supervision, can review and seal engineering work (NSPE Board of Ethical Review). And even where surface accuracy is high, understanding is not: one study found a leading assistant was right about 80 percent of the time but only 60 percent better than chance, with the lead author noting the models do not understand what they are talking about (Washington State University, 2026). “The software generated it” is not a defense.

6. It can expose proprietary data

Uploading a proprietary design to the wrong plan can put your IP into a training set. On consumer tiers, inputs may be used to improve models unless you opt out; business and enterprise tiers are excluded by default. Adoption data shows why governance is front of mind: a 2025 industry survey put AI use at just 27 percent of architecture, engineering, and construction firms, with risk, cost, and integration named as the top barriers (Bluebeam, 2025, vendor-sourced).

Workaround: keep proprietary designs and client data off consumer tiers that may train on inputs, use enterprise or zero-retention agreements, and write down a data-handling policy your team actually follows.

7. It drifts and answers inconsistently

Ask the same question twice and you may get two answers. The Washington State University study found only about 73 percent consistency when an identical question was asked ten times, which means roughly one answer in four disagreed with the others. Models also change between versions, so an approach that worked last quarter can quietly shift.

Workaround: pin model versions where you can, re-run critical prompts and compare the answers, and treat any single response as provisional until you have reproduced it.

How to work around the limits

Every fix above reduces to the same discipline. Keep it on a card by your screen:

None of this is a reason to avoid AI. Used within these limits, it is a strong assistant for drafting, explaining, and first-pass analysis. See where it earns its place in our AI engineering tools comparison table and the free options in our guide to the free AI tools for engineers.

Frequently asked questions

What are the main limitations of AI in engineering?

Seven recurring ones: unreliable calculation, fragile reasoning that degrades under small changes, hallucinated facts and standards, unreliable reading of technical drawings, no genuine physical or safety judgment, data-privacy and IP exposure, and version drift with inconsistent answers.

Can AI replace engineers?

No. Benchmarks put current models well below human experts on high-level engineering, and only a licensed engineer can be in responsible charge and seal work. AI is an assistant, not an accountable professional.

Why does AI hallucinate in technical work?

Because a language model generates statistically likely text rather than retrieving verified facts. Without source grounding it invents plausible-looking standards clauses and citations, which is why general-purpose models hallucinated on the majority of legal-reference queries in one study.

Is AI reliable for engineering calculations?

Not on its own. It makes unit and sign errors even when overall scores look high. Routing the math to a code or compute engine raises accuracy sharply, so the fix is external computation plus verification, not trust.

Can AI be held liable for engineering errors?

No. Accountability rests with the licensed engineer in responsible charge, who must review and approve or reject the work before sealing it. “The software generated it” is not a defense.

Can AI read engineering drawings and GD and T?

Only unreliably. The best multimodal models score around 60 to 70 percent on mechanical-drawing comprehension and stumble on tolerances, datums, and multi-view integration, so a human and, where needed, a purpose-built tool with QC are required.

Does AI give the same answer every time?

No. One study found only about 73 percent consistency when the identical question was asked ten times, so a single answer should be treated as provisional until reproduced.

The bottom line

The limitations of AI in engineering are real, specific, and manageable. Each one has a workaround, and all of them share a spine: the model assists, a deterministic tool computes, and a licensed engineer owns the result. Work inside those lines and AI speeds up the parts of engineering that are drafting and lookup, without ever standing in for the judgment that engineering requires. For where the assistants differ, our ChatGPT vs Claude for engineering comparison and the sub-cluster hub for the best AI tools for civil engineers are the next reads.

---

Sources

• Code execution lifts hard-math accuracy 53.9 to 84.3 percent: arXiv 2308.07921 (2023)
• Tokenization limits on arithmetic: arXiv 2505.14178 (2025)
• Numeral and unit conversion errors: NUMCoT, ACL 2024
• Statics score with persistent tension and compression errors: Hope et al., arXiv 2502.00562 (2025)
• Reasoning drops up to 65 percent under a single added clause: Mirzadeh et al., arXiv 2410.05229 (2024)
• EngiBench, below human experts and fragile to perturbation: Zhou et al., arXiv 2509.17677 (2025)
• Legal-reference hallucination 58 to 88 percent: Dahl et al., Journal of Legal Analysis (2024)
• Systematic review, reliability as the core barrier: Frontiers in Education (2025)
• Mechanical-drawing comprehension benchmark: MechVQA (2025)
• Fine-tuned vision model for drawing extraction: Springer (2025)
• Responsible charge and sealing drawings: NSPE
• Accuracy, consistency, and understanding: Washington State University (2026)
• AEC AI adoption and barriers: Bluebeam (2025)

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 and peer-reviewed research for the claims we make and update our recommendations as models and prices change. We do not test products ourselves; our assessments synthesize vendor documentation, primary research, and practitioner reporting.

Tool pricing and features change frequently. Always check the official website for the latest information before signing up.