Overview

Qualitative evidence synthesis brings together findings from qualitative studies, interviews, ethnographies, case studies, focus groups, to develop conceptual understanding that transcends any single study. Where meta-analysis pools numbers, qualitative synthesis interprets meanings: how people experience phenomena, what factors shape their behavior, why interventions work or fail in context. The methods include thematic synthesis, meta-ethnography, grounded theory synthesis, and framework synthesis, each with distinct epistemological commitments. AI tools can support several procedurally intensive stages of this work, coding, pattern identification, concept mapping, but the interpretive judgment at the heart of qualitative synthesis remains irreducibly human.

Title

Lesson 3.2: Qualitative Evidence Synthesis with AI

Purpose

This lesson teaches researchers how to synthesize qualitative evidence using rigorous approaches while leveraging AI for manageable scale. You'll learn to maintain interpretive integrity while using AI to assist with coding, pattern identification, and preliminary synthesis, ensuring AI enhances rather than replaces qualitative reasoning.

The Landscape of Qualitative Synthesis Methods

Before deploying AI assistance, researchers must select a synthesis method appropriate to their question and the epistemological stance of the included studies. Thematic synthesis, developed by Thomas and Harden (2008), is the most widely used method in health and social science: it moves from descriptive themes close to the data through to analytical themes that interpret underlying concepts. Meta-ethnography, developed by Noblit and Hare, focuses on translating the interpretive frameworks of studies into each other to build second-order and third-order constructs. Framework synthesis uses a pre-existing conceptual framework to organize and interpret findings. Grounded theory synthesis applies Glaser and Strauss's constant comparison method across studies to generate new theoretical propositions.

Each method requires different types of engagement with the data. Thematic synthesis requires line-by-line coding of findings sections; meta-ethnography requires identifying the key metaphors and concepts each study uses; framework synthesis requires mapping study findings onto framework dimensions. Understanding these distinctions is essential because they determine how you can and cannot use AI assistance without compromising methodological integrity.

For most systematic review contexts, thematic synthesis is the most tractable for AI-assisted workflows because its procedural steps, initial coding, developing descriptive themes, refining analytical themes, can be partially scaffolded by AI without displacing the researcher's interpretive role. Framework synthesis is similarly amenable because the mapping step is structured enough to be partially automated. Meta-ethnography and grounded theory synthesis are more epistemologically demanding and require heavier researcher involvement at every stage.

AI-Assisted Initial Coding

Initial coding in qualitative synthesis involves reading the findings sections of included studies and applying descriptive labels to segments of text that capture what participants reported or what the primary researchers found. This is procedurally intensive work: a synthesis of 30 qualitative studies, each with a 2,000-word findings section, means reading and coding approximately 60,000 words of dense qualitative text.

AI can accelerate initial coding substantially. The approach is to paste the findings text from one study into a prompt that instructs the AI to identify discrete findingsstatements, that is, the specific claims, themes, or descriptions the original authors report, and to apply brief descriptive labels to each. A suitable prompt is: 'Read this findings section and list each distinct finding as a numbered statement with a three-to-five word descriptive code. Focus on what participants reported or experienced, not the authors' interpretive commentary. Preserve the original language where possible.' The output provides a provisional coding list that the researcher then reviews, merges, splits, and re-labels based on their interpretive judgment.

The critical discipline is that AI coding at this stage is strictly first-pass. The researcher must read the original text alongside the AI-generated codes, identifying findings the AI missed (particularly nuanced or implicit ones), correcting codes that misrepresent the finding's meaning, and noting interpretive dimensions that require human judgment. A practical workflow is to complete AI-assisted first-pass coding of all studies, then conduct a second human-only pass that reviews and refines the code list before moving to theme development.

Consistency checking is another valuable AI function. After you have developed a provisional codebook from the first few studies, paste subsequent findings sections with your codebook and ask the AI to flag findings that don't map clearly to existing codes. These may indicate new themes or sub-themes that have emerged. This is not AI making interpretive decisions; it is AI identifying potential gaps in your coding framework for human review.

Pattern Identification and Theme Development

After initial coding is complete, the synthesis moves into theme development, identifying patterns across studies that reflect conceptually coherent groupings. In thematic synthesis, this involves grouping first-order codes into descriptive themes, then reflecting on what those descriptive themes collectively imply at a higher analytical level.

AI can assist with the pattern identification stage by helping you visualize and reorganize your code structure. Paste your full list of first-order codes (with their frequency and the studies they came from) into a prompt and ask the AI to suggest potential groupings, noting that these are provisional hypotheses for your review. The AI might, for example, identify that twelve codes relating to trust, disclosure, and relationship quality might cluster under a descriptive theme of 'relational conditions for engagement.' You then evaluate whether this grouping is conceptually coherent given the original study texts and whether it collapses distinctions that should be maintained.

AI can also help you identify areas of convergence and divergence across studies. Ask it to identify which studies contributed codes to each emerging theme and which themes were represented in only a few studies, suggesting tentative rather than robust patterns. This cross-study mapping is tedious to do manually with large numbers of codes and studies, but AI handles it efficiently.

The transition from descriptive to analytical themes is the most epistemologically demanding step and must remain with the researcher. Analytical themes go beyond what the studies say to interpret what the patterns mean, why certain experiences cluster together, what the synthesis implies for theory or practice that the individual studies couldn't reveal. This interpretive move draws on your domain knowledge, your understanding of the included studies' contexts, and your synthesis of the conceptual literature. AI can propose candidate analytical themes as prompts for your thinking, but accepting AI-generated analytical themes without critical evaluation would compromise the validity of the synthesis.

Quality Assessment in Qualitative Synthesis

Assessing methodological quality in qualitative research is more complex than applying a checklist, because the criteria for rigor differ across epistemological traditions. An interpretive phenomenological study should not be judged against criteria developed for grounded theory. Tools like CASP (Critical Appraisal Skills Programme), COREQ (Consolidated Criteria for Reporting Qualitative Research), and the JBI (Joanna Briggs Institute) qualitative appraisal tool each take different approaches to criteria for trustworthiness, reflexivity, and methodological fit.

AI can help with the preliminary stages of quality assessment by extracting relevant methodological information from papers: sampling approach, recruitment strategy, data collection method, analysis approach, reflexivity statement, member checking, and transferability claims. A prompt structured as: 'Extract the following methodological details from this methods section: [list your criteria]' provides a structured summary that the researcher then evaluates against quality criteria. This extraction step is mechanical and AI performs it reliably; the evaluation step requires human judgment.

When included studies vary in quality, the synthesist must decide how to handle this variation. Sensitivity analyses, re-running the synthesis excluding lower-quality studies, are increasingly expected. AI can help you draft the documentation for these sensitivity analyses and structure the comparison between full-synthesis and quality-restricted results. The decision about which threshold to apply, and whether quality differences explain thematic patterns, remains with the researcher.

Writing Up Qualitative Synthesis Findings

Writing qualitative synthesis is distinct from writing a literature review. A well-executed thematic synthesis or meta-ethnography presents findings that could not have been produced by reading any single included study, the synthesis itself is a new contribution. The writing should demonstrate that the analytical themes are grounded in the data, show which studies contributed to each theme, use verbatim quotations from included studies to anchor interpretations, and make explicit the interpretive steps that led from first-order codes to analytical themes.

AI can assist with structural aspects of the write-up: organizing the findings sections around themes, ensuring that each theme section includes a clear definition, supporting evidence from multiple studies, and an analytical statement about what the pattern means. A useful prompt is: 'I have a qualitative synthesis theme called [theme name] supported by codes from [n] studies. Draft a structured paragraph that introduces the theme, presents supporting evidence as a series of paraphrased study findings, and ends with an analytical statement.' The draft will need significant human editing to integrate the specific quotations and to ensure the interpretive language accurately reflects the synthesis rather than overstating consensus.

For the GRADE-CERQual (Confidence in the Evidence from Reviews of Qualitative Research) assessment, which many systematic review guidelines now require, AI can help you structure the four assessment dimensions: methodological limitations, coherence, adequacy of data, and relevance. Understanding and applying these criteria requires substantive knowledge; AI can assist with documentation and formatting once judgments have been made.

Maintaining Interpretive Integrity Throughout

The deepest risk in AI-assisted qualitative synthesis is epistemic: confusing AI-generated patterns with researcher-generated interpretations. When an AI suggests a thematic grouping, it is identifying textual co-occurrence patterns across a word vector space, a process fundamentally different from the interpretive, theory-informed judgment that a qualitative synthesist applies. The danger is that researchers, under time pressure, accept AI-suggested themes as conclusions rather than as prompts for deeper analysis.

Several practices protect interpretive integrity. First, maintain a reflexivity log throughout the synthesis, notes on your own interpretive decisions, where you agreed or disagreed with AI suggestions and why, and how your conceptual framework evolved. This log is evidence that the synthesis reflects human reasoning. Second, ensure that every analytical theme can be traced back to specific study findings through your code hierarchy; if a theme exists only because AI grouped codes and no researcher independently validated it against original texts, it lacks methodological grounding. Third, have a second researcher independently code a sample of studies and compare code lists, inter-rater processes remain essential in AI-assisted qualitative work because they test whether the patterns are recognizable to any trained reader or are artifacts of a particular AI-assisted process.

Finally, in the Methods section of your synthesis report, describe the role AI played at each stage with specificity: which steps were AI-assisted, what the AI produced, how the researcher reviewed and modified AI outputs, and which steps were conducted entirely by human researchers. This transparency allows readers to evaluate the synthesis appropriately and follows the emerging standards for AI disclosure in systematic reviews.