03 · What You Need to Know
What Researchers Should Report About AI-Assisted Systematic Reviews
Why Is AI Reporting a Methodological Issue Rather Than Just a Disclosure Issue?
AI can contribute to a systematic review in fundamentally different ways. It may correct grammar without affecting the evidence, or it may classify thousands of records and influence which studies enter the final synthesis.
These activities have different implications for reproducibility and methodological credibility. A reader evaluating a review needs to know whether AI merely assisted with presentation or participated in the identification, selection, appraisal, or synthesis of evidence.
Reporting should therefore distinguish between the use of AI as a writing assistant and its use as part of the research method.
AI disclosure
Identifies that AI-assisted technology was used, generally following the journal's disclosure or acknowledgment requirements.
Methodological reporting
Explains what AI did during the review, how the procedure worked, how outputs were verified, and how its involvement may have affected the evidence.
Both may be necessary. An acknowledgment does not explain a screening algorithm, while a detailed screening procedure does not necessarily satisfy a publisher's separate AI disclosure policy.
What Does PRISMA 2020 Actually Require About Automation?
PRISMA 2020 is a reporting guideline for systematic reviews. It is not an AI-specific approval framework, but several checklist items explicitly require details of automation tools when applicable.
Three particularly relevant items concern study selection, data collection, and risk-of-bias assessment. The explanation and elaboration document provides additional guidance on describing automated procedures and their validation.
| PRISMA item |
Reporting requirement |
AI-related information |
| Item 8: Selection process |
Describe how eligibility decisions were made, including reviewer involvement and automation. |
AI screening role, human reviewers, independence, decision rules, and validation. |
| Item 9: Data collection process |
Describe how data were collected from reports, including reviewer involvement and automation. |
AI extraction tasks, source verification, error checking, and independent extraction. |
| Item 11: Study risk-of-bias assessment |
Describe the assessment methods, tools, reviewer involvement, and applicable automation. |
Whether AI located evidence, suggested signaling-question responses, or proposed ratings. |
Other PRISMA items remain relevant when AI contributes to searching, synthesis, or reporting. For example, AI-assisted synthesis methods should be described accurately under the applicable synthesis-method items.
Researchers should consult the PRISMA 2020 explanation and elaboration rather than assume that adding an AI disclosure automatically satisfies the checklist.
What Should Be Reported About the AI Tool Itself?
At minimum, identify the system or service used and its relevant role. For methodological applications, additional information may be necessary to understand or reproduce the workflow.
Useful details include the provider, product, model identifier or version where available, date or period of use, relevant settings, input format, and whether the system accessed external documents or retrieval services.
For example, reporting that ChatGPT was used may be insufficient when the specific model, document-processing procedure, and instructions could affect screening results.
Researchers should not invent model version information that the service did not disclose. When exact technical details are unavailable, report what can be established and explain the limitation.
Should Researchers Report Their AI Prompts?
Prompts can materially influence how a generative model interprets eligibility criteria, extracts variables, or produces methodological recommendations.
Where AI performs consequential review tasks, preserving the relevant instructions, templates, and decision rules is advisable. These may be included in supplementary materials or an accessible repository when permitted by confidentiality, copyright, and journal policies.
Not every conversational exchange needs to appear in the main manuscript. The priority is to preserve the information necessary to understand how consequential outputs were generated.
Researchers should also distinguish a fixed prompt applied consistently across records from prompts modified during screening. Material changes can affect comparability and should be documented.
How Should AI-Assisted Screening Be Reported?
When AI participates in study selection, researchers should explain which records it assessed, how the eligibility criteria were supplied, what classifications it produced, and how those classifications affected subsequent decisions.
For example, an AI system might recommend inclusion or exclusion during title and abstract screening. Researchers should specify whether its recommendations were advisory, independently checked, or used to remove records automatically.
If automated exclusions occurred, the reporting should make their extent and validation procedures clear. Readers need to understand the possibility that eligible studies were missed.
When AI assists with full-text screening, the report should identify who made final eligibility decisions and how exclusion reasons were verified.
What Validation Information Should Be Included?
Statements such as "AI outputs were checked" or "the model was highly accurate" provide little methodological information without describing the verification process.
Researchers should specify the reference standard, number and characteristics of records evaluated, performance measures, types of errors examined, and procedures for resolving discrepancies.
For screening, sensitivity and false exclusions are especially relevant. Overall accuracy alone may be misleading when most retrieved records are ineligible.
The report should distinguish between model-level performance and the performance of the complete human-AI workflow. For example, an AI system might incorrectly exclude several eligible records that are subsequently recovered through human verification.
These are different results and should not be conflated. Detailed screening error measurements help make the distinction explicit.
How Should Manual Audits Be Described?
If researchers manually audit a sample of AI decisions, they should explain how the sample was selected, its size, the population from which it was drawn, and the human reference-assessment procedure.
A random sample, a stratified sample, and a targeted selection of difficult records support different conclusions. Results should be interpreted accordingly.
For example, examining 100 AI exclusions and finding no errors does not establish that the remaining exclusions are correct. Researchers should acknowledge sampling uncertainty and avoid claiming perfect reliability.
Where an audit of AI decisions identifies errors, report the corrective action taken and whether additional records were reassessed.
How Should AI-Assisted Data Extraction Be Reported?
AI may generate provisional extraction entries for study characteristics, outcomes, or numerical results. The methods should identify which variables were extracted, how the source documents were processed, and how the entries were verified.
Researchers should also explain whether multiple human reviewers independently collected or checked the data, particularly when the review follows standards requiring duplicate outcome-data extraction.
For AI-assisted data extraction, reporting should distinguish values copied from source reports, calculated values, and information that was unavailable or ambiguous.
Where extraction errors were evaluated, describe the reference dataset and results by variable type when relevant. A single accuracy percentage may conceal errors in numerical fields that directly affect meta-analysis.
How Should AI-Assisted Risk-of-Bias and Certainty Assessments Be Reported?
If AI identifies methodological passages or suggests risk-of-bias judgments, researchers should report the assessment framework used, the AI's specific role, and how human assessors reached the final judgments.
For example, using AI to locate passages relevant to RoB 2 is different from allowing it to assign final domain ratings. The report should make that distinction clear.
The same principle applies to certainty-of-evidence assessment. AI may organize information relevant to GRADE domains, but authors should document how final outcome-level certainty ratings were determined and justified.
Readers should be able to distinguish verified methodological assessments from AI-generated suggestions.
Where Should AI Use Be Disclosed in the Manuscript?
The appropriate location depends on how AI was used and the target journal's policies. The International Committee of Medical Journal Editors (ICMJE) recommends that authors disclose AI-assisted technology used in producing submitted work and explain its purpose.
Its recommendations distinguish writing assistance, which may be described in acknowledgments, from AI used in data collection, analysis, or figure generation, which should be described in the Methods section. The ICMJE also recommends disclosure in the cover letter.
| AI application |
Appropriate reporting location |
| Language editing or writing assistance |
Acknowledgment or designated AI disclosure, according to journal policy. |
| Search preparation or study screening |
Methods, with any additional required disclosure. |
| Data extraction or analysis |
Methods, including verification and relevant analytical procedures. |
| Risk-of-bias or certainty assessment support |
Methods, identifying the AI's role and final human assessment procedures. |
| Figure or diagram preparation |
Methods or designated disclosure, depending on whether AI affected the research process or only presentation. |
These are practical reporting distinctions rather than a universal journal template. Authors should verify the current instructions of the target publisher and journal.
Can AI Be Listed as an Author?
No. Under ICMJE recommendations, AI-assisted technologies should not be listed as authors or coauthors because they cannot assume responsibility for the accuracy, integrity, and originality of scholarly work.
Human authors remain responsible for checking AI-generated content, providing appropriate citations, ensuring originality, and addressing errors.
Disclosure does not transfer responsibility to the technology provider. Nor does it establish that the AI-assisted methodology was appropriate.
Does Disclosure Make an Unreliable AI Workflow Acceptable?
No. Transparent reporting enables readers to evaluate the methodology, but it does not correct invalid screening decisions, fabricated references, or inaccurate extraction.
A review that discloses autonomous AI screening but provides no defensible validation may still have serious methodological limitations.
Likewise, researchers cannot satisfy a requirement for independent human assessment merely by explaining that a model replaced one reviewer.
The appropriate methodological standards must be met independently of the disclosure statement.
What If the AI Model Changes After the Review?
Hosted generative AI systems may be updated, retired, or modified. Even when the same instructions are used, subsequent outputs may differ from those generated during the original review.
Researchers should therefore preserve available model information, relevant prompts, input records, output files, and verification logs. These materials may improve transparency even when exact computational reproduction is impossible.
It is important to distinguish reproducibility of the documented research procedure from the ability to obtain identical outputs from a changing commercial service.
Watch Out
A statement such as "ChatGPT was used, and all outputs were verified by the authors" is inadequate if AI made consequential screening or extraction recommendations and the manuscript does not explain how verification occurred. Report the actual workflow, including its limitations, rather than relying on a reassuring but unverifiable declaration.
04 · A Practical Example
Writing an AI Methods Statement for an Educational Technology Review
Hypothetical Example
A Systematic Review of AI-Generated Formative Feedback
A research team conducts a systematic review examining the effects of AI-generated formative feedback on university students' learning outcomes.
The team uses a generative AI model to produce provisional title and abstract screening recommendations and prepare candidate study-characteristic extraction entries. Human reviewers retain responsibility for final eligibility decisions and verify the extracted information against original reports.
What Would an Inadequate Disclosure Look Like?
"ChatGPT was used to assist with the systematic review. All outputs were checked by the researchers."
Although the statement acknowledges AI involvement, it does not explain which records were processed, what decisions the model made, how errors were evaluated, or how human verification was conducted.
What Would a More Informative Methods Statement Look Like?
The following hypothetical statement illustrates the level of methodological detail that could be appropriate. The model identifier, sample sizes, validation results, and procedures must be replaced with the research team's actual information before publication.
Illustrative Methods Statement
AI-Assisted Screening and Data Extraction
"A generative AI system ([provider, model, and version]) was used to generate provisional eligibility recommendations for [number] titles and abstracts according to the predefined review criteria. The model was instructed to classify each record as retain, exclude, or uncertain and to identify the criterion supporting its recommendation. Before implementation, performance was evaluated against [number] independently assessed records, yielding a sensitivity of [value] and specificity of [value]. AI recommendations did not independently determine final study eligibility. Human reviewers assessed potentially eligible full-text reports and resolved disagreements according to the review protocol."
"The same system was used to prepare provisional extraction entries for [specified study-characteristic variables]. Entries were checked against the original study reports, and discrepancies were corrected before synthesis. Numerical outcome data were extracted and verified according to the review's independent data-collection procedures. Relevant prompts, validation methods, and verification records were preserved in [specified supplementary material or repository]."
Why Is the Second Statement More Useful?
It distinguishes the AI's screening role from final eligibility decisions, identifies the need for validation evidence, specifies the types of extracted information, and describes human responsibility.
It also separates provisional AI outputs from the verified dataset used in synthesis. That distinction allows readers to understand whether the model could have introduced errors into the evidence base.
The statement should be adapted to the actual workflow. Researchers must not claim independent human screening, numerical validation, or complete verification unless those procedures genuinely occurred.
Step 1: Identify each material AI application
Record whether AI contributed to searching, screening, extraction, appraisal, synthesis, or reporting.
Step 2: Document the technical procedure
Identify the system, relevant model details, inputs, instructions, and outputs.
Step 3: Describe validation and verification
Explain the reference decisions, performance measures, human checking, and correction procedures actually used.
Step 4: Identify responsibility for final decisions
Specify which decisions were made independently by human reviewers and which were informed by AI.
Step 5: Complete the required disclosures
Place methodological details in the appropriate manuscript sections and comply with the journal's additional AI disclosure requirements.
The objective is not to produce the longest possible AI statement. It is to provide enough information for readers to understand and critically evaluate the procedures that influenced the review.
06 · What This Means for You
Report AI at the Level of Detail Its Methodological Role Requires
Begin by separating editorial assistance from activities that could influence the evidence base. A minor language-editing task may require a brief declaration under the journal's policy, while AI-assisted study selection requires a much more detailed methodological description.
For every consequential application, ask whether a reader could determine what the system did, how its decisions were checked, and whether the review's findings might have changed because of its involvement.
A practical AI reporting framework
If AI was used only for language or editorial assistance
Follow the journal's disclosure requirements and verify that the final manuscript remains accurate and appropriately attributed.
If AI influenced screening or eligibility decisions
Describe the records processed, classification rules, validation, human involvement, and treatment of exclusions.
If AI extracted study information
Identify the variables, source documents, verification procedures, and independent human extraction arrangements.
If AI assisted with appraisal or synthesis
Explain the specific analytical or interpretive assistance and how final judgments were established.
If model details or outputs cannot be reproduced exactly
Report the available technical information and acknowledge the limitations without inventing missing details.
What Records Should You Preserve During the Review?
Maintain an internal record of the AI tools used, their tasks, relevant model information, instructions, inputs, outputs, validation results, and human corrections.
Preserve the original bibliographic records and study reports separately from AI-generated classifications or extraction tables. Where outputs are modified, document which entries were corrected and why.
These records support accurate reporting and may be useful if reviewers request clarification about the AI workflow.
How Much Detail Belongs in the Main Manuscript?
The main Methods section should describe the procedures necessary to evaluate the review's methodological credibility. Lengthy prompts, detailed validation tables, and extensive technical logs may be placed in supplementary materials where appropriate.
The main text should still identify the AI's actual role and the safeguards used. Referring readers to a supplement should not conceal essential information about automated exclusions or unverified extraction.
Before submission, compare the completed report with PRISMA 2020 and the target journal's current instructions. The two sources address related but distinct reporting obligations.
