How Should Structural Engineers Use AI for Literature Reviews Without Being Dishonest?

The Direct Answer: AI Assistance Is Not Automatically Dishonest

Using AI during a PhD literature review is not inherently dishonest. The ethical question is not whether a model generated text, summarized papers, ranked references, or proposed search terms; it is whether the researcher represented the work accurately, disclosed material use when required, and personally verified the evidence supporting the final argument. A tool such as ChatGPT, Claude, Gemini, or a specialized research platform can help a structural engineer search more efficiently, but it cannot replace scholarly judgment, source reading, or accountability.

Also worth reading: How Can a Responsible AI Literature Review Improve Trust in Structural Engineering Practice? · Is AI-assisted structural engineering research honest, and how should engineers use it responsibly? · How Do Engineers Validate Physics-Informed Neural Networks for Structural Analysis?

The central distinction is between assistance and substitution. Assistance means using AI to suggest terminology, identify broad subject areas, detect missing search concepts, organize an existing collection, or compare notes that the engineer has independently checked. Substitution means accepting generated claims, citations, equations, design assumptions, or conclusions without examining the original sources. The second practice can be scientifically invalid even if no specific university rule explicitly prohibits it. A literature review must represent the state of knowledge, not merely the output of a language model.

As of 28 September 2026, there is no single global rule governing AI use in engineering research. Universities, journals, funders, and professional institutions may issue different requirements about disclosure, authorship, confidentiality, and acceptable use. The safest policy is therefore procedural: follow the relevant institution’s written rules, disclose meaningful AI assistance in the methods or acknowledgements, preserve prompts and outputs when appropriate, and maintain an audit trail showing how each cited source was verified.

What AI Can and Cannot Do in a Structural Engineering Review

AI is useful for turning a large, poorly organized engineering problem into a manageable research process. A structural engineer may ask a model to identify terminology used in seismic assessment, nonlinear finite-element analysis, structural health monitoring, construction cost prediction, or high-rise realignment. These generated terms can improve Boolean searches in databases such as Web of Science, Scopus, Engineering Village, or Google Scholar. AI can also propose synonyms for concepts such as “residual drift,” “progressive collapse,” “lateral force-resisting system,” and “digital twin.”

The model can also help organize material after retrieval. If an engineer supplies abstracts or legally obtained papers, AI may group them by research question, classify methods, or create a provisional comparison table. This can reduce the time spent formatting notes. However, classification does not prove methodological quality. A study may use sophisticated machine-learning terminology while relying on a small, geographically narrow dataset, weak validation, or an inappropriate error metric. Those issues require inspection of the methods and results.

AI should not be treated as a reliable bibliographic database. Language models may invent references, attach a real author to a nonexistent paper, misstate a publication year, or summarize one study as though it reported another. This failure is especially dangerous in structural engineering because a wrong parameter, failure mode, code clause, or experimental result can propagate into research conclusions or design practice. A citation should enter the review only after the engineer has located the actual publication and checked the relevant pages.

## Why the Ethics Are Different from Ordinary Writing Assistance

A literature review is not merely prose about a topic. It is a bounded argument about what has been studied, where evidence agrees, where findings conflict, and what remains unknown. Structural engineering adds a further obligation: conclusions often depend on physical models, material properties, boundary conditions, loading patterns, code versions, and safety factors. A fluent sentence can conceal a technically serious error even when its grammar is excellent.

The situation becomes problematic when AI is used to conceal the fact that the researcher did not read the cited work. That is not helped by changing “the literature shows” into “AI suggests.” The intellectual contribution remains unsupported. It is also problematic to upload confidential manuscripts, client data, unreleased test results, or proprietary designs to a public AI service without checking the provider’s data-retention and training terms. A research project may contain information that is unpublished because of contractual obligations, privacy concerns, security requirements, or restrictions associated with critical infrastructure.

Disclosure should describe what the tool actually did. “AI was used to improve writing” is insufficient if the model selected references, extracted numerical results, or produced a substantial first draft. A more precise statement explains the purpose and the human control: for example, that an AI tool generated search synonyms and an initial coding framework, while the researcher screened all records, read the included papers, verified every citation, and revised the final synthesis. The wording should be adapted to institutional policy rather than copied blindly from another university.

A Defensible Practical Workflow

A defensible workflow begins with a written research question and scope. For example, a review might examine machine-learning methods for predicting seismic demand in reinforced-concrete buildings, or compare AI-assisted approaches for structural realignment of high-rise buildings. The researcher should define the population, intervention, outcome, date range, databases, languages, and exclusion criteria before asking an AI system for assistance. This limits vague recommendations and makes later decisions auditable.

The next step is to use AI for brainstorming, not authority. Ask several models for controlled vocabulary, Boolean query variants, and possible review frameworks, but compare the outputs against the Engineering Vocabulary, relevant standards, and terminology used by the retrieved papers. Search each database directly and save the exact query, filters, date, and number of results. A useful operational threshold is to review the first 200 to 500 records carefully when the result set is manageable; if it is larger, apply documented screening stages and record how many records were removed at each stage.

Screening should remain human-led. Titles and abstracts can narrow the set, but inclusion decisions should follow a written rubric. Full-text review should record the research question, structural typology, data source, model family, validation design, performance metrics, limitations, and relevance to the review question. Researchers should distinguish predictive accuracy from physical plausibility and external validity. A reported 95% accuracy is meaningless without knowing the class balance, test-set construction, leakage controls, uncertainty treatment, and whether the model was tested on buildings outside its training distribution.

The final synthesis should be written from verified notes, not from the model’s memory. Every numerical claim should have a page or table reference, and every citation should be checked against the publisher record, DOI, repository, or authoritative index. A practical quality-control rule is to require two independent checks for high-consequence claims: confirm the source metadata in the bibliographic database and confirm the relevant statement in the full text. The researcher should also preserve rejected studies and reasons for exclusion, because apparent consensus may disappear when a carefully selected set of negative or null results is examined.

Comparison with Traditional, Automated, and Specialist Tools

Traditional review methods are slower but transparent. A researcher using Scopus or Web of Science can inspect query history, export records, apply inclusion criteria, and reproduce the search. Manual review also exposes uncertainty: the researcher may miss a concept, but that limitation is visible and can be corrected through additional searches. The best approach is often a hybrid method, in which traditional database controls govern the evidence and AI improves only the repetitive portions of screening or note organization.

Automated literature tools offer stronger reproducibility than a general-purpose chatbot when they are designed to retrieve and preserve records. They may provide deduplication, citation export, screening queues, and change logs. Yet automation does not remove bias. A screening classifier trained on labels from one discipline may perform poorly on civil or structural engineering because abstracts use abbreviations, numerical results, country-specific standards, and highly specialized terminology. The tool should therefore be evaluated on a manually labeled sample before it is trusted.

Specialist AI systems for structural engineering, such as the Arup and YJK AI Designer announced in Hong Kong and Vietnam, are aimed at design support rather than academic literature review. Their existence shows that engineering AI is moving from general text generation toward domain-specific workflows. Nevertheless, a structural design assistant should not be treated as a peer-reviewed literature database or as independent evidence for a thesis. Its assumptions, training materials, model boundaries, and intended users may differ from those of scholarly research. Commercial AI products also change quickly, and subscription pricing is not a measure of output reliability.

## Common Mistakes and How to Avoid Them

The most common mistake is citation fabrication. A model may produce a plausible title and author list that do not correspond to a real publication. The second is citation drift: a real paper is cited, but the model attributes a conclusion that the paper does not make. The third is metadata error, such as an incorrect year, journal, volume, page range, or DOI. None should survive verification. If a reference cannot be found in a publisher site, trusted index, institutional repository, or legitimate author copy, it should not remain in the bibliography.

Another mistake is confusing an abstract with a full-text result. Abstracts may omit important limitations, while conference summaries may overstate validation. Researchers also make errors by treating review articles as equivalent to primary studies, ignoring publication bias, or counting multiple papers from one dataset as independent evidence. AI can make these mistakes more efficiently, and it can lend a false appearance of objectivity because the output is rapid and neatly formatted.

Prompting does not solve the evidence problem. Asking a model to “be accurate,” “use only peer-reviewed sources,” or “do not hallucinate” reduces risk but cannot guarantee compliance. A stronger control is to restrict the model to a supplied evidence packet and ask it to identify passages that are not supported by the packet. Even then, every final statement must be checked by the researcher. The model is a processing assistant, not an author with responsibility for the literature.

When to Use AI, When Not to Use It, and What It May Cost

AI is most helpful at the beginning of a project, during the transition from an unfamiliar topic to a structured vocabulary. It can also be useful when the researcher must compare many clearly defined concepts, reconcile inconsistent note labels, or create an initial table of study characteristics. For a small review of perhaps 20 to 40 papers, manual reading may be faster than designing and validating an automated screening process. For a large review containing hundreds or thousands of records, deduplication and assisted prioritization can save substantial time, provided the rules are documented.

AI should not be used to access paywalled or unlawfully obtained material, bypass database limits, or reproduce protected text beyond permitted quotation. It should not make final inclusion decisions where the researcher lacks the technical background to evaluate a paper. Nor should it be used to upload identifiable data, critical-infrastructure information, or confidential engineering plans to an unapproved service. When the evidence concerns public safety, structural capacity, seismic performance, or code compliance, the responsible licensed engineer or research supervisor must remain in the verification loop.

Costs depend on the tool and usage pattern. General AI tools commonly provide free or limited-access tiers, with paid individual or team subscriptions adding higher usage limits, advanced models, privacy controls, or collaboration features. University and institutional plans may be available through a library or research office. Database access can cost more than the AI tool itself, and commercial engineering platforms may require negotiated licensing. The relevant cost calculation should include researcher time, training, supervision, data storage, privacy review, and the cost of correcting an incorrect citation or model assumption. A $20 monthly software subscription is not economical if it generates evidence that must be manually reconstructed.

The Recommended Standard for Structural Engineering Research

The appropriate standard is not “AI use is forbidden” or “AI use is always acceptable.” It is that the researcher must be able to explain every source, every extracted value, and every conclusion in the literature review without relying on the model as a hidden authority. The engineer should know why a study was included, what its method actually was, which limitations affect transferability, and whether the cited result was observed or merely inferred. AI may assist with retrieval, organization, and clarity, but the scholarly responsibility remains human.

For a publication or thesis, the best practice is to document tool use in a way that is proportional to its role. A journal may require disclosure in a separate AI statement, while a university may require inclusion in the methodology or acknowledgements. The author should also follow the target’s policy on AI-generated text and citation responsibility. A useful test before submission is to pretend the supervisor will ask for the primary source behind each important sentence: the researcher should be able to provide it immediately.

This standard is consistent with the wider direction of engineering AI. Arup and YJK have promoted AI Designer for structural engineering, while research cited in the supplied context addresses applications of machine learning in construction cost prediction, AI-assisted structural realignment, and field reconstruction of structural responses. These developments indicate active experimentation, but they do not establish that any AI system can replace expert review or guarantee code-compliant design. The same principle applies to academic research: capability should not be mistaken for validation.

In practical terms, use AI to widen vocabulary, accelerate repetitive organization, and challenge assumptions. Do not use it as a substitute for reading, a citation resolver without manual confirmation, or a final technical reviewer. Keep a reproducible log, disclose meaningful assistance, protect confidential information, and require human sign-off before conclusions enter a dissertation, peer-reviewed paper, or engineering decision. Under that approach, using AI can be transparent and academically legitimate; without those controls, it can produce convincing work that is fundamentally untrustworthy.

A Clear Decision Rule for Researchers and Engineers

A simple decision rule is: if a mistake from the AI output could change the research conclusion, design assumption, public-safety judgment, or interpretation of the evidence, the output requires direct expert verification. This rule applies to a cited numerical result, a proposed model parameter, a statement about a failure mode, a bibliographic reference, and a summary of a complex paper. It is not enough that the model sounds confident or that its output contains citations; confidence is not evidence.

Researchers should also preserve enough information to reproduce the workflow. That includes the date of each search, database and query, inclusion and exclusion criteria, duplicate-removal method, tool name and version where known, purpose of use, disclosure statement, and the identity of the person who approved the final source set. A future auditor may need to distinguish between records found by the database, records suggested by the model, and records included after expert review. Those categories should not be blurred.

The strongest literature reviews therefore combine machine efficiency with engineering discipline. AI shortens the distance between a question and a candidate set of sources; human expertise determines whether those sources deserve trust. For structural engineering, that human expertise is particularly important because models operate within a physical world governed by uncertainty, material variability, construction tolerances, loading histories, and public consequences. The tool can assist the process, but it cannot own the responsibility.