Is Using AI for a PhD Literature Review in Structural Engineering Honest?

Direct Answer to the Ethics Question

Using AI tooling during a PhD literature review is not inherently dishonest, but presenting AI-generated synthesis as your own unaided scholarship is not academically acceptable. The central issue is not whether a tool can help you search, classify, summarize, compare, or code; it is whether you disclose material use and retain the intellectual responsibility for every claim, citation, interpretation, and conclusion. A literature review remains an original scholarly act even when software assists with retrieval or organization. If the university permits generative AI, the defensible approach is to treat it as an unverified assistant rather than as an author or final authority.

Also worth reading: How Is AI Structural Engineering Used in Practice, and What Should Professionals Verify in 2026? · How Should Structural AI Audit Trails Work in Engineering Systems in 2026? · How Do Structural Engineering Firms Handle AI Capacity Planning for Massive Data Centers and Heavy Workloads?

The ethical line depends on the function assigned to the AI. Asking a model to suggest search terms, map broad subject areas, or identify potentially relevant terminology is substantially different from asking it to write a review, fabricate references, conceal gaps, or make a final argument. Tool-assisted work becomes questionable when the researcher cannot explain how a source supports a statement, has not read the cited paper, or cannot account for studies the system omitted. In structural engineering, where design decisions can affect public safety, source fidelity matters more than polished prose. A 2026 literature review should therefore be reproducible, traceable, and explicit about limitations.

There is no universal percentage at which AI use becomes dishonest; the threshold is functional rather than numerical. As a practical rule, any use that affects the wording, selection, analysis, or presentation of scholarly work should be acknowledged under your institution’s policy. Purely administrative tasks may require less disclosure, but a paragraph written by AI and retained without substantial revision should be disclosed. If you cannot separate your reasoning from the tool’s output, you have not yet met the standard of accountable authorship. The safest formulation is simple: AI may help you process evidence, but you must verify evidence and own the review.

How to Use AI Without Misrepresenting Your Work

Begin with a documented search protocol rather than prompting a model to “write my literature review.” Define the research question, database scope, date range, disciplines, intervention terms, outcome terms, and exclusion criteria before opening an AI tool. A defensible initial database might cover structural engineering, structural health monitoring, AI-assisted design, reinforced concrete, steel structures, seismic assessment, and multi-hazard resilience, with a cutoff date of 1 October 2026. Record the databases searched, the exact query strings, the search date, and the number of records retrieved. This turns a casual search into a method that another researcher can audit.

Use AI for bounded support: generating synonyms, proposing controlled vocabulary, grouping papers by method, comparing structured metadata, or producing questions for manual inspection. Do not treat generated summaries as substitutes for reading the original article, especially when a paper’s conclusion depends on assumptions about geometry, loading, material properties, uncertainty, code version, or experimental limitations. For example, a paper involving machine-learning prediction of construction cost may be useful for a review of project-cost methods, but its findings should not be transferred directly to structural design. The domains may share data and modeling techniques while having different safety requirements and validation standards.

The review should distinguish evidence types. Peer-reviewed experimental studies, numerical studies, systematic reviews, case studies, commercial software reports, conference papers, and preprints do not carry equal weight. AI can help classify those categories, but it may misclassify a document or incorrectly describe its methodology. Manually confirm the title, authors, year, journal, DOI, dataset, and central result. A 2019 paper cited through an undated blog or generated reference is not the same as a verified 2019 publication, and a confident model response is not evidence that the publication exists.

A strong disclosure sentence names the tool, its purpose, and the human verification process. A weaker disclosure says only “AI was used,” while an improper submission hides the use entirely. Universities and journals may have different rules, so the student’s supervisor, research-integrity office, or journal author instructions should control the final wording. A transparent methodology can state that AI-assisted tools were used for keyword generation or metadata organization, that records were checked against source repositories, and that interpretation and conclusions were completed by the researcher. This wording is useful because it communicates the boundary between assistance and authorship.

A Practical Workflow for an AI Structural Engineering Review

The first stage is scope definition. A PhD literature review in AI structural engineering should not attempt to cover every application of machine learning in civil engineering. It should define whether the main question concerns design generation, structural assessment, damage detection, code checking, structural response reconstruction, robotics, cost prediction, or resilience. The more specific the scope, the easier it is to set inclusion thresholds and recognize irrelevant results. For a review focused on AI-assisted structural design, papers about cost prediction or robotics may belong only in adjacent-work sections. For a review of structural response measurement, sensor-based monitoring and field reconstruction may be central, while generative design is peripheral.

The second stage is searching. Use library databases, publisher platforms, Google Scholar where appropriate, official standards bodies, and targeted searches of organizations such as Arup, YJK, and relevant engineering societies. AI-generated keywords should be expanded into controlled terms and abbreviations, not accepted without review. For instance, searches may combine terms such as “machine learning,” “deep learning,” “structural response,” “digital twin,” “structural health monitoring,” “reinforced concrete,” “seismic,” “fire,” “foundation,” and “uncertainty.” Keep a search log and deduplicate by DOI, title, author, and year rather than simply deleting repeated titles from different databases.

The third stage is screening. Screening can be assisted by AI, but every included study should receive a human decision based on explicit criteria. A practical exclusion log might record the title, year, reason for exclusion, and which criterion applied. If only 40 records out of 500 are retained, the review should show how that selection was made rather than presenting the model’s selection as objective. The number of records is not a quality measure; a small, relevant evidence base can be more useful than a large collection of loosely connected studies.

The fourth stage is extraction and analysis. Use a spreadsheet or reference manager to record the engineering problem, structural type, data source, model family, validation method, performance metric, uncertainty treatment, code or standard context, limitations, and relevance to the research question. For example, report accuracy only with its baseline, test-set definition, and whether the model is predicting a known quantity. AI may help summarize these fields, but the researcher should inspect the original tables and methods. Then compare studies by evidence strength, application maturity, and failure modes rather than by a single headline metric.

What Counts as Strong Evidence in Structural Engineering?

Structural engineering research places unusually high demands on traceability because an apparently accurate prediction can conceal an invalid assumption. A neural network trained on displacement data may perform well statistically while ignoring stiffness, boundary conditions, material nonlinearity, load combinations, and measurement uncertainty. A study using machine learning to estimate structural response should therefore be read for its physics assumptions, not just its reported error. Numerical validation should be distinguished from experimental validation, and interpolation within a training range should not be confused with extrapolation to a new building.

A review should also separate design, assessment, and operation. AI-assisted design tools may propose member sizes, reinforcement layouts, or structural systems, whereas assessment tools infer damage or capacity from images, vibration data, or point clouds. Field reconstruction tools estimate structural responses from incomplete measurements, which is different from predicting a future load. This distinction is reflected in research such as the work titled “Artificial intelligence assisted structural realignment of high-rise buildings through lifting, grouting and reinforcement,” where the engineering process includes physical intervention, grouting, and reinforcement rather than only a software prediction.

The date and maturity of a source matter. Commercial announcements can establish that a product or collaboration exists, but they cannot independently establish design safety, regulatory approval, or performance in a particular project. Arup and YJK’s launch of AI Designer is evidence of an industry initiative aimed at structural engineering, and Arup’s Hong Kong announcement provides a specific geographic and organizational example. It is not, by itself, peer-reviewed proof that the tool improves design outcomes. Likewise, a systematic review of AI-driven field reconstruction can identify research patterns, but its conclusions remain subject to the quality and geographic coverage of the included studies.

For a PhD review, set minimum documentation standards. Prefer peer-reviewed primary studies and authoritative standards, record DOI or stable publisher links, and note whether a result is theoretical, simulated, laboratory-based, field-tested, or commercial. A source claiming a 95% accuracy should be recorded with the dataset size, class balance, baseline, and metric definition. Without those details, the number is not meaningfully comparable. The researcher should also identify whether the study reports uncertainty, failure cases, out-of-distribution testing, or human review.

Comparison of Reasonable and Problematic AI Uses

The following comparison evaluates common uses of AI in a structural-engineering literature review by its appropriate role, disclosure expectation, and main risk. It does not assume that every tool has the same capability or that all institutional policies are identical.

This table shows that the most important distinction is not whether AI is used, but whether the researcher preserves source control and accountability. A tool can be useful for discovery and organization while remaining unsuitable for evidentiary judgment.

Common Mistakes and How to Prevent Them

One common mistake is confusing fluency with factuality. Language models can produce a smooth technical paragraph containing an incorrect citation, an invented publication, or a false claim about a building system. Another mistake is allowing a tool to resolve disagreements among authors. The researcher should inspect the original methods, compare sample populations and materials, and explain why apparently conflicting findings differ. A conflict may arise from concrete grade, steel type, geometry, seismic intensity, sensor placement, training data, or evaluation metric, not from genuine scientific disagreement.

A second mistake is using a narrow dataset while claiming a broad field. If a study examines 100 images from one laboratory, the review should not describe it as a general solution for structural health monitoring. If an AI system is trained on one software’s output, its portability to other codes, finite-element packages, or physical buildings is unproven. Researchers frequently omit the negative cases, code assumptions, and human interventions that determine whether a result can be applied in practice. These omissions are especially serious when a review is later used to justify automated design decisions.

The third mistake is poor source hygiene. Researchers may cite a search snippet, a press release, or a model-generated bibliography entry as if it were a peer-reviewed article. Verify the title, author list, publication venue, year, DOI, and final version through an authoritative repository. A preprint should be labeled as a preprint, and a conference abstract should not be presented as a full journal paper. If the record cannot be located, exclude it or describe it explicitly as an unverified lead rather than allowing it to influence the review’s conclusions.

The fourth mistake is neglecting nontechnical evidence. Structural engineering research can be affected by building regulations, fire codes, seismic standards, material specifications, labor practice, software licensing, cybersecurity, and public acceptance. An AI workflow that improves a design metric but cannot explain decisions to a regulator or peer reviewer may still be unsuitable. A responsible review should therefore discuss explainability, traceability, data governance, model maintenance, and liability alongside prediction accuracy.

When to Act, Pause, or Escalate

Act with AI when the task is exploratory, reversible, and easy to verify. Keyword expansion, citation-manager cleanup, formatting, and preliminary grouping are suitable uses if the researcher checks the output. Use more caution when the tool produces evidence tables, summarizes quantitative findings, or drafts passages that will enter a thesis chapter. Pause if a model cites a source you cannot locate, if it combines results from different structural systems, or if it cannot identify the limitations of a dataset. Escalate to the supervisor or research-integrity office when the tool appears to have generated fabricated references, when the intended use is undisclosed, or when the work will support a safety-related engineering decision.

The date context for this answer is 1 October 2026. That date should be stated in a search protocol as the review cutoff, but it should not be treated as proof that research published after an earlier search has been assessed. A database update should be documented, and newly published commercial announcements should be separated from peer-reviewed evidence. If a project begins in October, record the exact search date, because databases change daily and search results can vary by subscription, language, and index coverage.

Timing also affects authorship policy. Institutions may update guidance after a publication, training event, or misconduct case. A policy that permitted a limited AI use in 2025 may be interpreted differently in 2026, particularly if the tool can now generate images, code, datasets, or long-form prose. Check the current policy rather than relying on an informal statement from another student. Journals also distinguish acceptable language assistance from unacceptable use that conceals missing research work.

Cost and pricing are secondary to integrity, but they should be planned. Many language models have free or low-cost entry tiers, while institutional subscriptions, paid databases, reference-management tools, and commercial engineering software can add recurring expense. A student can reduce unnecessary cost by using library access, open repositories, DOI resolution, and a simple reference manager. Paid tools may offer better context windows, search integration, or document handling, but price does not guarantee citation accuracy or scholarly validity. Budget for independent verification rather than treating the subscription fee as an audit.

A Defensible Submission Standard

Before submission, test the review against five questions. Can another researcher reproduce the search? Can every citation be located in a trusted source? Can you explain each included study’s method and limitations in your own words? Does the synthesis distinguish peer-reviewed evidence from announcements and preprints? Does the submission disclose material AI assistance in the manner required by your institution or target journal? If any answer is no, the review is not ready for submission.

The best default practice is to retain a human-readable audit trail containing search logs, inclusion decisions, extraction spreadsheets, notes from reading the original papers, and a record of tool use. Keep prompts and outputs only when policy or confidentiality permits, especially if unpublished research, proprietary drawings, personal data, or commercial information are involved. Do not upload confidential structural plans or sensitive project data to a public AI service without checking its terms and your obligations. An academically honest review is also a data-security review.

The final judgment is therefore conditional. AI can make literature discovery faster, help compare terminology, and reduce clerical effort, but it cannot replace the researcher’s responsibility to read, assess, contextualize, and criticize evidence. In AI structural engineering, the most credible work will connect computational claims to physical engineering: loads, materials, boundary conditions, uncertainty, codes, constructability, inspection, and failure consequences. Use the tool to improve the process, not to erase the process.

If you are deciding whether a particular use is acceptable, describe the action in one sentence and ask whether you would be comfortable making the same description to your supervisor, ethics or integrity office, journal editor, and the eventual reader. If yes, the use may be defensible with appropriate controls. If no, disclose it, revise it, or avoid it. The standard is not whether AI usage sounds modern; it is whether the resulting literature review remains truthful, traceable, and yours in the scholarly sense.