💡Quick Answer
To conduct an AI-assisted systematic literature review, follow four core phases aligned with PRISMA 2020 guidelines: (1) Define your research question using the PICOS framework, (2) Use AI tools for automated database searching and abstract screening, (3) Leverage NLP for structured data extraction from full-text PDFs, and (4) Synthesize findings with AI while maintaining human-in-the-loop quality control at every stage.
Time Investment: Traditional vs. AI-Assisted SLR
SLR Phase | Traditional Method | AI-Assisted Method |
|---|
Search strategy & Boolean string development | 2–4 weeks | 2–4 hours |
Screening 10,000 abstracts (title/abstract) | 3 months | 3 days |
Full-text review & data extraction (500 PDFs) | 3–5 months | 1–2 weeks |
Synthesis & research gap analysis | 2–3 months | 1–2 weeks |
Reference formatting (APA/MLA/Vancouver) | 1–2 weeks | 5 minutes |
Total estimated time | 6–18 months | 3–6 weeks* |
* Includes human-in-the-loop quality control at every phase. AI accelerates the process; it does not replace the researcher's critical judgment.
Traditional vs. AI-Assisted SLR Methodology
Systematic literature reviews (SLRs) represent the highest level of evidence in academic research. They follow a rigorous, reproducible methodology — most commonly the PRISMA 2020 framework — to identify, screen, extract, and synthesize all relevant evidence on a specific research question.
The challenge is well-documented: a traditional SLR takes an average of 67.3 weeks from start to publication. Screening alone — reading thousands of titles and abstracts — can consume months of researcher time and is prone to fatigue-induced errors. This is precisely where AI provides the most value: automating the repetitive, high-volume tasks while the researcher maintains control over interpretation and quality.
💡 Key principle throughout this guide:AI does not replace the researcher. It automates the heavy lifting — database searching, abstract screening, data extraction — while you maintain intellectual ownership of the research question, eligibility criteria, synthesis, and conclusions. This is the "human-in-the-loop" approach that both PRISMA guidelines and journal editors increasingly expect.
PRISMA 2020 Flowchart: Where AI Fits Into Each Phase
The PRISMA 2020 flow diagram defines four stages of a systematic review. Here's how AI tools map onto each stage — and where human judgment remains essential:
1. Identification— Records identified from databases (PubMed, Scopus, Web of Science) and registers.AI Role:Search & Discover — AI-powered search query expansion, synonym suggestion, Boolean string generation, cross-database deduplication.
2. Screening— Title/abstract screening, removal of irrelevant records.AI Role:Bulk Screening — NLP-based relevance scoring, automated inclusion/exclusion classification, priority ranking of abstracts (AI-Assisted + Human Review).
3. Eligibility— Full-text assessment, data extraction, quality appraisal (Risk of Bias).AI Role:Extract & Appraise — PDF parsing, structured data extraction, RoB flagging (AI-Assisted + Human Verification).
4. Included— Final study set, synthesis, meta-analysis, reporting.AI Role:Synthesize & Write — Cross-study pattern analysis, research gap identification, structured literature review drafting (AI Draft + Human Authorship).
Phase 1: Define Your Research Question Using PICOS
Every systematic review begins with a precisely formulated research question. The PICOS framework ensures your question is structured, searchable, and reproducible:
Element | Definition | Example (Diabetes SLR) |
|---|
P | Population — Who are you studying? | Adults with Type 2 Diabetes (age ≥18) |
I | Intervention — What treatment or exposure? | AI-assisted glucose monitoring systems |
C | Comparison — What is the alternative? | Traditional self-monitoring of blood glucose |
O | Outcome — What do you measure? | HbA1c reduction, hypoglycemic events |
S | Study design — What types of studies? | RCTs and quasi-experimental studies |
Your PICOS definition directly feeds into the next critical step: building your search strategy.
Phase 1.5: Using AI to Generate Boolean Search Strings
The backbone of any systematic search is the Boolean query — the precise combination of AND, OR, and NOT operators that defines what databases return. Traditionally, crafting these strings is painstaking work requiring deep knowledge of MeSH terms, database syntax, and field-specific vocabulary. AI can dramatically accelerate this process.
💡 📋 Prompt Template — Boolean String Generation
I am conducting a systematic literature review on [your PICOS question].
Generate a comprehensive Boolean search string for PubMed that includes:
MeSH terms and free-text synonyms for each PICOS element
Appropriate use of AND, OR, NOT operators
Field tags ([tiab], [MeSH]) where applicable
Truncation wildcards (*) for term variations
My PICOS: P: [Population] I: [Intervention] C: [Comparison] O: [Outcome] S: [Study design — e.g., RCT, cohort]
💡 Critical:AI-generated Boolean strings are a starting point, not a final product. Always validate the output with a librarian or information specialist. Check that MeSH terms are current, syntax matches the target database, and no key synonyms are missing. Document any modifications you make — this is essential for PRISMA reproducibility.
Phase 2: Automated Literature Search & Screening
This is where AI delivers its most measurable time savings. Once you've executed your search across databases and collected thousands of records, the screening phase begins — and it's traditionally the most labor-intensive part of the entire SLR process.
AI screening tools use natural language processing to analyze each title and abstract against your eligibility criteria, assigning relevance scores and flagging records for inclusion, exclusion, or human review. The best tools maintain recall rates above 95% — meaning they catch virtually all relevant studies while dramatically reducing the number of irrelevant records you need to read.
📋 Prompt Template — Abstract Screening Criteria
Screen the following abstracts against these eligibility criteria:
INCLUDE if the study: Involves [population] as defined in our PICOS; Examines [intervention] compared to [comparison]; Reports [outcome measures]; Uses [study design: RCT, cohort, etc.]; Published between [year range]; Written in [language(s)].
EXCLUDE if the study: Is a review, editorial, letter, or conference abstract only; Involves [specific exclusion, e.g., pediatric populations]; Does not report quantitative outcomes.
For each abstract, classify as: INCLUDE / EXCLUDE / UNCERTAIN. Provide a one-sentence justification for each decision.
Once you've identified your eligible studies, structured data extraction begins. This means reading each full-text paper and pulling out specific data points into a standardized form. For a review with 50-200 included studies, this alone can consume months.
AI tools using NLP can parse PDF documents and extract structured data fields — transforming unstructured academic text into tabular data that's ready for analysis.
📋 Prompt Template — Structured Data Extraction
Extract the following data points from this PDF into a structured table:
First Author, Year, Country, Study Design, Sample Size (n), Population, Intervention, Comparator, Primary Outcome, Key Finding, P-value, Effect Size (CI), Limitations, Risk of Bias Notes. Flag any data points you are uncertain about with [VERIFY].
Accuracy warning:A JMIR study (2025) found that AI-extracted data was accurate in approximately 51% of cases, with 13% imprecise and 22% missing. Always verify AI-extracted data against the original PDF. AI handles the bulk formatting; you handle the verification. Never submit AI-extracted data without human cross-checking.
Phase 3.5: Risk of Bias Assessment with AI Assistance
Quality appraisal is a non-negotiable component of any rigorous SLR. The most widely used frameworks —Cochrane RoB 2for randomized trials andROBINS-Ifor non-randomized studies — require systematic evaluation of each included study across multiple bias domains.
AI can serve as a first-pass assistant here: scanning full-text papers for signals relevant to each bias domain and flagging areas of concern. But the final scoring must always be done by the researcher.
📋 Prompt Template — Risk of Bias Pre-Assessment
Analyze this study for Risk of Bias using the Cochrane RoB 2 framework. For each domain (randomization process, deviations from intended interventions, missing outcome data, measurement of outcome, selection of reported result), extract relevant text from the paper and provide an initial assessment. Suggest: Low Risk / Some Concerns / High Risk. Flag anything marked [INSUFFICIENT INFORMATION] for manual review.
Human-in-the-loop is non-negotiable here.AI can flag that a paper doesn't mention randomization concealment, but only a domain expert can judge whether this omission represents poor reporting or genuine methodological weakness. Use AI to speed up the pre-assessment; score the final Risk of Bias table yourself.
Phase 4: Synthesis, Gap Analysis & Writing
With your extracted data verified and quality appraised, the synthesis phase transforms individual study findings into a coherent narrative. This is where AI's cross-study pattern analysis capabilities become valuable — it can identify recurring themes, contradictory findings, and underexplored areas across dozens or hundreds of papers simultaneously.
However, synthesis is also where human expertise is most critical. AI can surface patterns; only you can interpret what those patterns mean for your field, construct arguments, and draw conclusions that advance knowledge.
📋 Prompt Template — Research Gap Identification
Based on the following extracted data from [n] studies on [topic]: [Paste your extraction table or summary] — Identify: 1) CONSISTENT FINDINGS: What do most studies agree on? 2) CONTRADICTORY FINDINGS: Where do results diverge? 3) RESEARCH GAPS: What questions remain unanswered? 4) METHODOLOGICAL PATTERNS: What study designs dominate? Present findings in a structured narrative suitable for a literature review chapter. AI identifies patterns; you provide the expert interpretation.
Documenting AI Use: PRISMA-trAIce Compliance
As AI use in systematic reviews becomes standard practice, transparent reporting is essential. ThePRISMA-trAIce checklist(published in JMIR, 2025) provides a 14-item framework specifically designed for reporting AI use in evidence synthesis — building on PRISMA 2020 but addressing the unique transparency requirements of AI-assisted research.
Key reporting items include:
AI Tool Identification:Name, version, and provider of every AI tool used
Phase-Specific Usage:Which PRISMA phases involved AI assistance
Prompt Documentation:The exact prompts or configurations used
Human-AI Interaction:How human oversight was maintained at each stage
Performance Metrics:Recall, precision, and accuracy rates where measurable
Limitations Disclosure:Known limitations of the AI tools used
💡 Practical tip:Start documenting your AI usage from Day 1 of your review. Keep a log of every prompt, every tool version, and every human override decision. This documentation is increasingly required by journals including Elsevier, Nature, and Cochrane — and it protects the reproducibility of your work.
AI Tool Comparison: Which Tool for Which Phase?
No single tool covers every phase of a systematic review equally well. Here's how the major platforms compare across PRISMA phases:
Tool | Search | Screening | Data Extraction | Synthesis | PRISMA Diagram | Citation Mgmt |
|---|
NevaScholar | ✓ | ✓ | ✓ | ✓ | Planned | ✓ |
Elicit | ✓ | Partial | ✓ | Partial | ✗ | ✗ |
Rayyan | ✗ | ✓ | Partial | ✗ | ✓ | ✗ |
Covidence | ✗ | ✓ | ✓ | ✗ | ✓ | Partial |
Litmaps | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
Scite.ai | ✓ | Partial | Partial | Partial | ✗ | Partial |
Table reflects publicly available features as of March 2026. Verify current capabilities on each platform's website.
Ethical Considerations & Known Limitations of AI in SLRs
Academic rigor demands honesty about what AI can and cannot do. Here are the documented limitations every researcher should understand:
Hallucination risk:AI can fabricate citations, author names, and study findings that don't exist. Every AI output must be verified against the source document.
Incomplete coverage:A JMIR study found that neither Elicit nor Connected Papers retrieved all records found by the PRISMA method. AI tools should supplement, not replace, traditional database searching.
Prompt dependency:The quality of AI output varies dramatically based on prompt design. Poorly constructed prompts produce unreliable screening and extraction results.
Paywall limitations:Most AI tools cannot access full-text articles behind publisher paywalls, limiting their extraction capabilities to open-access content.
Training data bias:AI models may overrepresent English-language publications and certain disciplines, potentially introducing systematic bias into your review.
The responsible position:AI significantly accelerates systematic reviews, but the PRISMA method continues to exhibit clear superiority in terms of reproducibility and accuracy. Use AI to reduce the burden of repetitive tasks. Rely on human expertise for quality assessment, synthesis interpretation, and final conclusions.
