03 · What You Need to Know
How to Create a Defensible Record of AI-Assisted Literature Searching
Why Does AI Assistance Need Documentation?
Literature searching involves methodological decisions that influence which studies become available for screening and synthesis.
These decisions include selecting information sources, choosing terminology, constructing queries, applying limits, and identifying additional publications through supplementary methods.
Generative AI can influence any of these activities. Sometimes its contribution is minor, such as explaining a Boolean operator. In other cases, it materially changes the search by introducing new terminology, recommending a database, or selecting publications for further examination.
Documentation helps distinguish these situations.
It also allows researchers to investigate potential errors. If an AI-suggested term excluded relevant studies or an AI-generated reference proved nonexistent, a preserved record can help explain how the problem arose and whether it affected the final search.
The Cochrane Handbook emphasizes maintaining sufficient documentation to report literature searches correctly. PRISMA-S provides detailed guidance for reporting search methods and information sources in systematic reviews. These principles provide a methodological basis for documenting AI assistance, although they should not be misrepresented as a universal AI-specific prompt-reporting requirement.
What Is the Difference Between Documenting AI Use and Disclosing AI Use?
Documentation and disclosure serve related but different purposes.
Internal documentation
The research team's working record of AI tools, prompts, outputs, search decisions, verification, and retrieval activities. It supports auditability, quality control, and later reconstruction.
Published disclosure
The information reported to readers, reviewers, editors, or institutions about how AI contributed to the research, according to applicable policies and methodological requirements.
A research team may preserve detailed prompt histories internally while reporting only the consequential AI-assisted activities in the manuscript and supplementary materials.
Conversely, a statement such as "ChatGPT was used during the literature review" may satisfy neither methodological transparency nor a particular publisher's disclosure requirements.
The appropriate disclosure depends on what the tool actually did and the policies governing the work.
Which AI Activities Are Worth Recording?
The strongest reason to document an activity is that it materially influenced how the literature was identified, retrieved, or selected.
| AI-assisted activity |
Useful documentation |
Why it matters |
| Generating search terminology |
Prompt, candidate terms, accepted or rejected suggestions |
Explains how the search vocabulary developed |
| Constructing Boolean expressions |
AI-generated draft, modifications, final executed query |
Distinguishes proposed syntax from the actual search |
| Translating database strategies |
Original query, translated draft, verified database-specific version |
Shows whether conceptual or syntactic changes occurred |
| Suggesting controlled vocabulary |
Candidate headings and verification against official thesauri |
Prevents unverified AI terminology from being treated as authorized indexing |
| Recommending publications |
Candidate records, source information, verification outcome |
Distinguishes genuine publications from unsupported recommendations |
| Retrieving scholarly records |
Connected source, query or retrieval method, date, result count and exported records where available |
Supports evaluation of source coverage and retrieval behavior |
| Assisting screening |
Eligibility criteria, model role, validation procedure, consequential decisions |
Shows how automated judgments affected record selection |
Not every conversational interaction requires the same level of preservation. Asking AI to explain the meaning of AND may have little methodological consequence. Accepting an AI-generated exclusion term that changes which studies are retrieved is considerably more consequential.
Should You Save Every Prompt and AI Response?
Not necessarily. There is no universal methodological rule requiring researchers to publish every prompt used during literature searching.
However, preserving consequential prompts and outputs can be valuable, particularly when they influenced search terminology, database selection, query construction, or the identification of records.
A useful record should preserve the exact prompt where its wording affected the outcome, together with the response that informed the decision.
If several prompts were used to refine a strategy, retain the important iterations rather than presenting only the final successful exchange.
This helps distinguish a strategy that was developed through systematic evaluation from one selected because it happened to produce appealing results.
For systematic reviews, the final executable search strategies and actual retrieval records remain more important than an extensive collection of prompts that cannot establish what was searched.
What Tool Information Should Be Recorded?
Record the AI product or service and the model identifier or version where available and relevant.
Also record the date of use and whether the system had retrieval capabilities, connected databases, uploaded documents, or other information sources available during the task.
If the tool used a specialized scholarly retrieval service, identify that service when possible.
Some commercial systems do not disclose their complete model configuration, retrieval algorithms, or underlying indexes. Researchers should report what can be established rather than inventing unavailable technical details.
For example, identifying the product and date may be possible even when the exact underlying model snapshot is not disclosed.
Uncertainty about the system's configuration should be treated as a limitation, particularly when it affects reproducibility.
How Should You Record AI-Generated Search Terms and Boolean Strings?
Preserve the distinction between suggestions and executed searches.
Suppose AI proposes several alternative terms for faculty trust in generative AI. The researcher rejects some terms because they represent broader constructs and retains others after examining relevant publications.
The record should identify the consequential suggestions and explain the important inclusion or exclusion decisions.
When AI drafts a Boolean expression, preserve the draft where useful, but also retain the exact query executed in each database.
These expressions may differ because the researcher corrected errors, added terminology, changed field restrictions, or adapted syntax.
The process of constructing Boolean search strings with AI is therefore not adequately documented by saving only the AI's initial response.
How Should You Document Search Strategy Translation?
When AI helps adapt a query from one database to another, record the original expression, the proposed translation, and the final verified version.
Document important differences in field coverage, controlled vocabulary, phrase searching, proximity operators, or other retrieval functions.
For example, a PubMed title-and-abstract search may be translated into a Scopus expression that also searches keyword fields. This changes the search scope even if the Boolean concepts remain similar.
The final executed strategy should reflect the destination database's actual syntax.
Researchers should not present a translated draft as though it were the query used to retrieve records.
These distinctions are central to adapting literature searches between databases.
What If AI Recommends Research Papers?
Record whether the recommendations were generated from model knowledge or retrieved from identifiable scholarly sources, to the extent that this can be established.
For consequential recommendations, preserve enough information to identify the candidate publications and determine how they were handled.
A paper may be verified and included in subsequent searching, rejected because it is irrelevant, or excluded because the citation does not correspond to a genuine publication.
These outcomes should not be confused.
In particular, asking AI for research papers is not automatically equivalent to executing a literature search.
Where AI recommendations lead to additional database searches or citation searching, document those activities separately.
What If AI Performs the Search Through a Connected Database?
Retrieval-enabled AI introduces additional documentation requirements because the system may perform search operations on the researcher's behalf.
Ideally, researchers should identify the underlying source, query or retrieval method, date, result count, and records returned.
However, some AI systems expose only a ranked selection of publications and do not reveal the full query or candidate set.
In such cases, researchers should describe the limitation rather than claiming that the search was fully reproducible.
For reviews requiring comprehensive evidence identification, an opaque retrieval process may be more appropriate as a supplementary discovery method than as the sole search strategy.
How Should You Record AI-Assisted Screening?
Screening differs from search development because it directly influences which retrieved records proceed toward inclusion.
If AI classifies titles or abstracts as relevant or irrelevant, researchers should document the criteria supplied to the system, the role of human reviewers, and the procedures used to check its decisions.
Where automated exclusion is permitted by the review methodology, validation becomes particularly consequential because false exclusions can remove eligible evidence.
Researchers should avoid claiming that AI performed a validated screening task merely because its classifications appeared plausible.
Search reporting guidance and screening reporting guidance address related but distinct methodological activities. Documentation should make clear which stage AI assisted.
What Does PRISMA-S Actually Require?
PRISMA-S is a reporting guideline for literature searches in systematic reviews. It contains sixteen reporting items addressing information sources, search strategies, limits, supplementary methods, and record management.
Among its recommendations is the provision of full search strategies for each database and information source, copied as actually executed.
It also addresses other search methods, including approaches whose exact results may not be reproducible because of proprietary or opaque algorithms. In such circumstances, transparent reporting of the method and relevant starting references remains important.
PRISMA-S was not developed as a comprehensive AI governance standard. Its provisions should not be interpreted as requiring a particular AI disclosure formula or universal prompt archive.
Nevertheless, its principles provide a useful foundation for documenting AI-assisted retrieval. The PRISMA-S explanation and elaboration discusses these reporting expectations.
How Should You Write an AI Assistance Statement?
The statement should describe the actual contribution rather than rely on a generic declaration.
For example, if AI was used only to suggest terminology, the statement should not imply that the system retrieved or screened the studies.
If the tool assisted with screening, the statement should identify that activity and the validation procedures used.
Illustrative Reporting Statement
AI Used for Search Strategy Development
"A generative AI assistant was used to propose alternative search terminology and preliminary Boolean expressions. The research team evaluated the suggestions against relevant publications and database documentation, revised the strategies, and independently executed the final searches in the selected bibliographic databases. The final database-specific search expressions and search dates are provided in the supplementary materials."
This is a hypothetical example, not a universally required disclosure statement.
The actual wording should identify the tool and relevant dates where required, reflect the procedures performed, and comply with applicable institutional or publisher policies.
What Should Be Kept Internally but Not Necessarily Published?
Research teams may preserve more detailed information than appears in the final manuscript.
Internal records can include consequential prompt histories, rejected terminology, draft queries, verification notes, retrieval exports, and decisions made during strategy development.
These materials can support peer review, later updates, or investigations of unexpected search results.
However, researchers should not upload confidential manuscripts, unpublished research materials, personal data, or restricted documents to external AI services without appropriate authorization and safeguards.
Documentation should also respect applicable data protection requirements, institutional policies, and contractual restrictions.
Where a prompt contains sensitive information, preserve records using an approved secure process rather than assuming that every prompt should be publicly disclosed.
