01 · The Question
Can AI Turn Your Systematic Review Records Into a PRISMA Flow Diagram?
A systematic review may begin with thousands of database records and end with a relatively small number of included studies. Researchers need to explain what happened between those stages: how many records were identified, how many duplicates were removed, how many reports were retrieved, and why others were excluded.
The PRISMA flow diagram presents this selection process visually. Generative AI can organize numbers, prepare flowchart structures, and help identify inconsistencies. That makes it tempting to ask a chatbot to generate the entire diagram.
But a visually convincing diagram is not necessarily an accurate account of the review. Can AI generate a valid PRISMA flow diagram, and what information must researchers provide before doing so?
03 · What You Need to Know
How AI Can Assist With PRISMA Flow Diagram Preparation
What Is a PRISMA Flow Diagram?
The PRISMA flow diagram is a standardized visual representation of how records and reports move through the identification and selection process of a systematic review.
PRISMA stands for Preferred Reporting Items for Systematic Reviews and Meta-Analyses. PRISMA 2020 provides a reporting framework that helps authors describe how a review was conducted and what it found.
The flow diagram typically documents records identified, records removed before screening, records screened, reports sought for retrieval, reports assessed for eligibility, exclusions, and studies included.
PRISMA is a reporting guideline, not a method for conducting the review itself. Completing its diagram does not establish that the underlying searches or eligibility decisions were methodologically adequate.
Why Are Records, Reports, and Studies Different?
One of the most consequential PRISMA errors is treating every retrieved database record as a separate study.
A record is a bibliographic entry, such as a database search result. A report is a document presenting information about a study, such as a journal article, conference paper, or technical report. A study is the underlying investigation.
Record
A bibliographic entry identified through searching or another discovery method.
Report
A document providing information about an investigation. Several reports may describe one study.
Study
The underlying investigation, which may be represented by one or multiple reports.
For example, a randomized trial may have a conference abstract, a primary journal article, and a follow-up publication. Those are three reports, but they may describe one study.
AI-generated diagrams must preserve this distinction, particularly when multiple reports share an underlying study population.
Which PRISMA 2020 Flow Diagram Template Should You Use?
The official PRISMA 2020 website provides four principal flow-diagram templates, reflecting whether the review is new or updated and whether it uses databases and registers only or also identifies records through other methods.
The appropriate template depends on how the review was conducted, not on which layout looks most convenient.
| Review situation |
Template consideration |
| New review using databases and registers |
Use the corresponding new-review template. |
| New review using additional identification methods |
Use the template accommodating databases, registers, and other methods. |
| Updated review using databases and registers |
Use the update template that distinguishes previous and newly identified evidence. |
| Updated review using additional identification methods |
Use the update template accommodating other identification sources. |
Researchers should obtain the appropriate template directly from the official PRISMA 2020 flow diagram resources.
What Information Does AI Need to Generate the Diagram?
AI requires the actual selection counts and a clear description of how those counts were obtained. Ideally, the information comes from database exports, deduplication logs, screening software, retrieval records, and full-text exclusion documentation.
Depending on the template, the necessary information may include:
- Records identified from each database, register, or other source.
- Duplicate records removed before screening.
- Records removed through other documented prescreening procedures.
- Records screened and excluded.
- Reports sought for retrieval and reports not retrieved.
- Reports assessed for eligibility.
- Reports excluded at full-text assessment, with specific reasons.
- Studies and reports ultimately included in the review.
The model should not infer missing values merely to make the diagram complete. If a count is unavailable, researchers must investigate the underlying records rather than accept a plausible estimate.
Can AI Calculate the Numbers Automatically?
AI may help check relationships among supplied counts, but the calculations must reflect the actual review process.
For a straightforward database-search workflow, the number of records screened is usually the number identified minus records removed before screening. Reports sought for retrieval are then determined by the screening decisions.
This calculation establishes the number of records screened in the simplified example. It does not establish how many reports were retrieved, how many studies were eligible, or whether the original search was comprehensive.
More complex workflows may require separate accounting for databases, registers, citation searching, organizational websites, or previous review versions.
Why Must Full-Text Exclusion Reasons Be Verified?
At the full-text stage, researchers should document why reports failed the predefined eligibility criteria. Examples might include an ineligible population, study design, intervention, or outcome.
AI can help organize verified exclusion reasons into categories. However, it should not create reasons that were never recorded or assign a reason simply because it appears plausible from the article's title.
When several criteria are unmet, the team should follow a consistent procedure for assigning the primary exclusion reason. Counts should represent excluded reports according to the selected diagram's structure, not multiple reasons counted repeatedly for the same report.
AI-assisted full-text screening therefore needs traceable decisions if its outputs are later used to populate the PRISMA diagram.
Can AI Create the Visual Diagram Itself?
Generative AI may produce a draft flowchart, structured diagram data, or code for a diagramming application. Some AI-enabled tools can also populate a visual template.
However, the output format matters. A generated image may look attractive but be difficult to edit or correct. A structured, editable diagram is generally preferable when counts or labels may change during manuscript preparation.
The official PRISMA website provides templates and resources specifically designed for the reporting standard. Researchers should prioritize these over a decorative flowchart that changes required terminology or omits necessary boxes.
AI can assist with preparation, but the final diagram should retain the structure and meaning of the appropriate PRISMA template.
What If AI Was Used During Screening?
AI-assisted screening does not eliminate the need to report actual record counts. If AI classified titles, abstracts, or full-text reports, the diagram should still reflect the records processed and the final selection decisions.
Researchers should separately describe material AI involvement in the methods, including its role, validation, and human verification procedures.
PRISMA 2020 does not provide a universal replacement flow diagram simply because AI was used. The diagram documents the selection process, while the methods explain how that process was conducted.
Where automated procedures removed records before screening, researchers should examine the applicable PRISMA reporting categories and accurately distinguish those removals from exclusions made during screening.
What Should Researchers Do When the Counts Do Not Reconcile?
An arithmetic inconsistency is a signal to investigate the source records. It is not an invitation to adjust one number until the diagram balances.
Common causes include duplicate records counted twice, missing retrieval outcomes, inconsistent screening exports, reports confused with studies, and full-text exclusion reasons assigned more than once.
AI may help identify where an inconsistency occurs, but researchers must determine which source record is incorrect and document the correction.
Watch Out
Never ask AI to estimate PRISMA counts from the final number of included studies. Many different selection histories can produce the same final study count. A mathematically consistent diagram may still be fabricated if its numbers do not come from the actual review records.
04 · A Practical Example
Preparing a PRISMA Flow Diagram From Verified Screening Records
Hypothetical Example
A Systematic Review of AI-Supported Feedback in Higher Education
A research team searches several bibliographic databases for studies examining AI-supported formative feedback among university students. The team maintains search exports, deduplication logs, screening decisions, and full-text exclusion records.
After completing selection, the researchers use generative AI to organize their verified counts for a PRISMA 2020 diagram.
| Selection stage |
Verified count |
| Records identified from databases |
2,400 |
| Duplicate records removed |
500 |
| Records screened |
1,900 |
| Records excluded during title and abstract screening |
1,650 |
| Reports sought for retrieval |
250 |
| Reports not retrieved |
10 |
| Reports assessed for eligibility |
240 |
| Reports excluded after full-text assessment |
200 |
| Reports of included studies |
40 |
| Studies included in the review |
36 |
How Should the Counts Be Checked?
Identification and deduplication
2,400 records identified − 500 duplicates removed = 1,900 records screened.
Title and abstract screening
1,900 records screened − 1,650 excluded = 250 reports sought for retrieval.
Retrieval and eligibility
250 reports sought − 10 not retrieved = 240 reports assessed for eligibility.
Full-text selection
240 reports assessed − 200 excluded = 40 reports of included studies.
Study identification
The 40 included reports describe 36 underlying studies, because some studies have more than one report.
AI can check these relationships and organize the information into the corresponding template. However, it cannot establish whether the counts are authentic. That requires comparison with the original review records.
How Should Full-Text Exclusions Be Presented?
Suppose the 200 excluded reports have one documented primary exclusion reason each:
| Primary exclusion reason |
Reports excluded |
| Ineligible population |
55 |
| Ineligible intervention |
65 |
| Ineligible study design |
60 |
| Ineligible outcome requirement |
20 |
|
Total
|
200
|
The categories reconcile with the 200 excluded reports. This does not mean each report failed only one criterion. It means the team consistently assigned one primary reason for reporting.
The final diagram should be prepared using the official template applicable to the review and checked against these verified records.
06 · What This Means for You
Generate the Diagram From an Audit Trail, Not From a Narrative Guess
Before using AI, assemble the actual counts from the review's search and selection records. Confirm which PRISMA 2020 template applies and whether all necessary stages have been documented.
AI can then assist with organizing the data, checking arithmetic relationships, and preparing a draft. The final diagram must be compared with the source records and official reporting guidance.
A practical diagram-preparation framework
If all selection counts are documented
Use AI to organize and check them, then populate the appropriate official PRISMA template.
If some counts are missing
Investigate search exports, screening logs, and retrieval records before finalizing the diagram.
If the counts do not reconcile
Identify the source of the discrepancy rather than adjusting values merely to balance the flowchart.
If multiple reports describe one study
Preserve the distinction between included reports and included studies.
What Should Be Preserved for Reproducibility?
Retain the original search exports, deduplication records, screening logs, retrieval outcomes, full-text exclusion reasons, and the final list of included studies and reports.
If AI assists with diagram preparation, preserve the input counts, relevant instructions, generated draft, and corrections. This allows the final diagram to be checked without relying on the model's account of what happened.
Material AI involvement can be explained through the review's AI-use reporting, particularly when the system performed more than editorial formatting.
