Direct Answer: Disclosure and Verification Decide Honesty
Using AI during a PhD literature review is not inherently dishonest. It becomes academically unacceptable when the researcher presents AI-generated text, citations, summaries, or conclusions as their own unaided work, conceals material use that the supervisor or institution requires them to disclose, or relies on fabricated and misrepresented evidence. The ethical question is therefore not simply whether a tool was used, but what it did, how it was used, and whether a competent researcher can verify and defend the resulting work.
Also worth reading: How Should AI Structural Engineering Agents Be Secured Before They Can Design Buildings? · How Should Structural AI Audit Trails Be Built for Traceable Engineering Decisions? · Is AI-Assisted Structural Engineering Verification Reliable, and How Should Engineers Use It?
A useful distinction is between assistance and substitution. AI can help generate search terms, organize known references, compare terminology, identify broad review articles, and flag passages that require closer reading. Substitution occurs when AI is asked to produce the review itself, and the researcher accepts its claims without checking the underlying publications, design methods, results, limitations, or applicability to structural engineering. Universities and publishers are still developing consistent AI rules, so there is no universal rule that automatically makes every use acceptable or prohibited.
As of October 1, 2026, the defensible position is straightforward: use AI only within your institution’s policy, disclose it when required, preserve an audit trail, and personally validate every source and technical claim. AI can be part of a legitimate research process, but it cannot perform the scholarly accountability attached to earning a PhD.
What “Dishonest” Means in a Literature Review
A literature review is more than a decorated bibliography. It requires selection, appraisal, synthesis, and an argument about what the available evidence means for a particular research problem. Those activities are scholarly acts because they depend on disciplinary judgment. For example, two machine-learning approaches for estimating structural response may report lower prediction error, but one may have been trained on data leakage from the same building, while the other may operate under measured field conditions. A responsible reviewer must notice the difference rather than repeat a generic claim that one method is “better.”
Dishonesty can take several forms. A researcher may submit AI-written prose without disclosure, invent references through hallucination, quote a paper that does not contain the claimed passage, or state that no systematic search was conducted when phrases were generated without a documented method. Less visible misconduct includes using AI to screen abstracts but then describing the process as fully manual, or relying on a commercial system whose training data included confidential manuscripts. The severity depends on institutional rules and intent, but unsupported claims and false representations of method are serious problems in any case.
Intent still matters, although it does not erase a violation. A PhD candidate who misunderstands a journal’s AI policy may unknowingly submit undisclosed AI text, but correcting the record promptly is better than preserving the misleading account. Conversely, calling routine assistance “manual” to avoid disclosure is deceptive. The safest standard is behavioral rather than intent-based: an examiner should be able to reconstruct how the review was prepared and see that the candidate understands every included study.
How AI Can Be Used Legitimately in Structural Engineering Research
AI is reasonably useful for the mechanical parts of discovery. A structural engineering literature search may involve combinations such as “AI,” “machine learning,” “deep learning,” “structural response,” “field monitoring,” “digital twin,” “damage detection,” and “multi-hazard resilience.” AI can propose synonyms that a researcher might miss and help separate terminology used in civil, structural, geotechnical, and computational engineering. Such assistance is comparable in purpose to using a thesaurus or citation-database recommendation engine, provided the researcher checks the output.
It can also help summarize a paper that the researcher has already located and read, create comparison tables from verified fields, or rephrase text for clarity. In principle, a researcher can ask AI to compare the datasets, input features, validation methods, and error metrics in five confirmed papers. The output still has to be checked against each paper, especially where units and terminology differ. Structural results may involve displacement in millimeters, acceleration in meters per second squared, drift ratios, stress in megapascals, or probability of exceedance; an apparently small wording change can alter the engineering meaning.
AI should not be treated as a systematic-review engine unless its behavior has been tested and documented. Current systems can omit relevant studies, prioritize recent or frequently mentioned work, mishandle publication dates, and conflate preprints with peer-reviewed articles. Broad discovery support is therefore reasonable; claims of exhaustive or reproducible retrieval without a conventional database protocol are not. Published work on AI-assisted structural realignment and reviews of AI-driven field reconstruction shows why source quality matters: the underlying methods may involve actual intervention, grouting, reinforcement, sensor data, or reconstructed responses, none of which can be assessed from an AI summary alone.
A Practical Workflow That Survives Scrutiny
Begin with a written question and scope, such as how machine-learning models reconstruct structural response under incomplete field measurements or how AI affects inspection decisions for aging structures. Record the databases, date range, search strings, inclusion criteria, exclusion criteria, and search date. As of October 1, 2026, a search should include the final date, because relevant engineering literature changes monthly and a review claiming current coverage needs a defensible cutoff.
Next, use AI to propose search variants, but execute and save the searches yourself. Confirm that important databases in the relevant field have been checked, such as Web of Science, Scopus, Google Scholar, IEEE Xplore, ASCE Library, ScienceDirect, SpringerLink, and institutional repositories. Database indexing and coverage differ, so no single source should be treated as exhaustive. Keep the exact query, filters, result count, and export file; a useful benchmark is to review the first 200 results and then document why screening continued or stopped.
After retrieval, deduplicate records and screen titles and abstracts before reading the full text. AI may help flag probable duplicates, but author names, article versions, errata, conference papers, and preprints require manual resolution. For each included study, record the structure type, material, loading condition, data source, sample size, model family, validation design, uncertainty treatment, and principal limitation. A review of 40 studies should contain 40 verifiable evidence records, not 40 polished paragraphs created without source inspection.
Finally, write the synthesis from primary sources and retain a verification log. A practical rule is to open and inspect the abstract, methods, results, tables, and conclusion of every cited paper; ideally, confirm the exact page or table supporting each material claim. Re-run factual checks immediately before submission because AI systems can change between versions. The candidate should then be able to answer, without the chat interface, why each study was included and what its evidence contributes.
AI Assistance Compared with Conventional Research Tools
Traditional and AI-assisted methods can coexist, but they do not offer equal speed, reproducibility, or accountability. The table below illustrates a realistic division of work rather than a ranking in which AI replaces scholarly review.
| Feature | AI-assisted workflow | Conventional scholarly workflow |
|---|---|---|
| Discovery speed | Produces synonyms and candidate papers in minutes | Search refinement is slower but fully recorded |
| Reference verification | Can propose citations, but fabricated citations remain possible | Database records and PDFs provide traceable evidence |
| Screening | Useful for initial suggestions on a known, supplied set | Manual screening is slower and easier to audit |
| Synthesis | Can draft comparisons after data are supplied | Researcher performs and defends the synthesis |
| Reproducibility | Model, prompt, version, and output may change | Search dates, queries, and decisions can be preserved consistently |
| Best use | Search ideation and clerical assistance | Source appraisal, technical judgment, and final argument |
| Main risk | Plausible errors presented confidently | Human bias, incomplete searching, and time pressure |
| Disclosure | Required according to institutional or publisher policy | Usually no AI disclosure is needed unless a tool was materially used |
There is no single established market price for responsible AI-assisted literature review work. Individual subscriptions may range from free consumer tiers to roughly $20–$200 per month for higher-tier services, while institutional access can cost thousands of dollars annually. These figures are prices, not proof of quality. A $200 monthly tool does not guarantee accurate citations, and a free model can still support legitimate tasks if the researcher verifies everything; conversely, no paid model removes the need to read sources.
Common Mistakes and Failure Signals
The most obvious mistake is trusting citations because they look realistic. Generative systems can invent a plausible title, author, journal, year, and DOI. Verification should use the DOI resolver, publisher page, database record, or a trusted library catalogue; a citation should not be accepted merely because it appears in several AI-generated answers. The research context also illustrates why “cite checking” is independently important: an AI boom can increase both useful discovery and the volume of unchecked claims.
Another error is confusing an article’s title with its finding. “AI-Driven Field Reconstruction of Structural Responses: A Systematic Review” indicates a review about structural-response reconstruction, not evidence that every AI reconstruction method has been field validated. The wording does not establish model accuracy, sensor independence, transferability, or compliance with building codes. Similarly, news about Arup and YJK launching AI Designer concerns a commercial design tool and does not by itself establish independent performance evidence or professional endorsement of unreviewed model output.
Researchers also mishandle version history and confidentiality. They may paste copyrighted full texts, sensitive structural drawings, proprietary test data, or unpublished thesis chapters into a service whose retention terms they have not examined. They may use an institutionally prohibited account or fail to obtain supervisor approval. Before entering material, the candidate should check data classification, licensing, consent, retention, training use, and geographic or contractual restrictions.
The final common error is polishing AI prose until it sounds personal. Fluency can conceal weak reasoning, and style editing can make unverified material harder to detect. A candidate should be able to explain every included model, dataset, equation, and limitation in ordinary language. If the generated review contains claims the candidate could not defend during a viva or examination, it is not ready for submission.
When to Use AI, and When to Stop
AI is appropriate when the task is exploratory, reversible, and easy to verify. Examples include generating 20 search synonyms, identifying alternate abbreviations, suggesting a taxonomy after supplying verified articles, or formatting a table from researcher-entered facts. It is also reasonable for a disabled or overloaded researcher to use accessibility assistance, subject to policy. The key threshold is whether errors can be detected before they enter the scholarly record.
AI is inappropriate when it supplies the final argument, generates references that cannot be independently resolved, or replaces appraisal of safety-relevant evidence. It should not be used to fabricate a negative result, conceal a search failure, or make inaccessible personal judgments about authorship. Formal systematic reviews, scoping reviews with registered protocols, and safety-critical engineering assessments require especially strong provenance and repeatable procedures.
A practical 70% rule can guide process expectations: spend at least 70% of the literature-review effort on source acquisition, reading, verification, critical comparison, and writing from evidence. This is not a university compliance threshold, but it can expose a reversed workflow in which most effort goes into prompting and rewriting. The candidate should also be prepared to regenerate the review manually if the tool becomes unavailable. If the underlying work cannot survive loss of access to the model, the model has probably become too central.
Before submission, obtain supervisor confirmation of local policy, read the target journal’s or publisher’s author instructions, and preserve prompts and outputs only where lawful and necessary. A disclosure might state which tool class was used, its purpose, affected stages, and that all citations and claims were checked against primary sources. Exact wording should follow institutional guidance rather than a generic internet template.
The Standard for Responsible AI Structural Engineering
The durable standard is verifiability. In AI structural engineering more broadly, models may assist design, condition assessment, reconstruction, robotics, cost prediction, and multi-hazard analysis, but engineers remain responsible for assumptions, uncertainty, code compliance, and public safety. Literature review is the scholarly counterpart of that responsibility: AI may help locate and organize evidence, while the researcher must establish whether the evidence is valid and relevant.
For a PhD candidate, a defensible conclusion is “No, using AI is not automatically dishonest; yes, undisclosed use, fabricated citations, or unverified authorship claims can be dishonest.” The candidate should be able to show a search log, inclusion decisions, source files, extraction records, prompt history where required, disclosure statements, and a coherent rationale for the review’s conclusions. Those artifacts convert an opaque tool interaction into auditable scholarship.
This answer reflects the policy position as of October 1, 2026, not a guarantee of future institutional rules. Policies may become stricter, and services may change their data practices. Nevertheless, the core obligations are unlikely to change: preserve research integrity, respect confidentiality, disclose material assistance, avoid false claims, and take personal responsibility for every part of the submitted work.