Can Generative AI Accurately Extract Sample Size and Participant Characteristics From a Research Paper?

Can AI Identify How Many Participants Were Actually Included in a Study?

You ask generative AI to extract the sample size from a research paper. It reports 350 participants. Yet the methods section says 350 individuals were recruited, the participant flow diagram shows 312 completed the study, and the primary analysis includes only 298.

Which number should appear in your literature review matrix?

Extracting participant information may appear straightforward, but research papers frequently report different numbers at different stages. The challenge is not simply finding a number. It is determining what that number represents and whether the associated participant characteristics accurately describe the population relevant to the study's findings.

AI Can Extract Participant Information, but It May Confuse Different Sample Counts

Why Extracting Sample Size Is More Complicated Than Finding a Number

Research papers may describe participants at several stages, from initial recruitment to the final analysis. These numbers serve different purposes, and none should automatically be substituted for another.

The appropriate count depends on what the researcher needs to understand. For instance, recruitment numbers describe the initial study population, while analytical sample sizes indicate how many observations contributed to particular findings.

Which Sample Size Should AI Extract?

Before requesting a number, distinguish the different participant counts that may appear in a paper.

These categories are not necessarily applicable to every study, and terminology varies across disciplines and designs.

For example, a qualitative interview study may report the number approached, consented, interviewed, and included in the analysis. A secondary-data study may report records rather than recruited human participants.

AI should extract the relevant categories rather than force all studies into a conventional recruitment-to-analysis sequence.

Why Can a Paper Report Several Different Sample Sizes?

Participant numbers can change because of eligibility screening, nonparticipation, withdrawal, loss to follow-up, incomplete measurements, exclusions, or analysis-specific data requirements.

Consider a hypothetical survey of 500 teachers. Researchers may receive 430 responses, exclude 25 records for failing prespecified quality checks, and analyze 405 responses for descriptive statistics.

A regression model requiring complete information on additional variables might include only 376 participants.

All these numbers may be correct, but they answer different questions.

Reporting the regression sample as the total number of survey respondents would be inaccurate. Likewise, reporting the initial invitations as the number analyzed would overstate the evidence available for the statistical findings.

What Is the Difference Between Sample Size and Analytical Sample Size?

The term sample size may refer broadly to the number of participants or observations included in a study. The analytical sample size refers more specifically to those contributing to a particular analysis.

These counts can differ when data are missing or different analyses have different eligibility requirements.

For example, a study might analyze 420 responses for descriptive statistics but use 387 complete cases for multiple regression.

Some analytical approaches, including certain multiple-imputation and likelihood-based methods, can use information from participants with incomplete observations. In such cases, the number contributing information to an analysis cannot always be understood as a simple complete-case count.

AI should therefore report the analysis-specific count and the missing-data approach when relevant, rather than assuming that the smallest number mentioned in the results represents the final sample.

Can AI Extract Participant Characteristics Accurately?

Participant characteristics describe who or what was studied. Depending on the research question, they may include age, sex or gender as measured and reported, educational level, professional experience, socioeconomic characteristics, geographic setting, institutional affiliation, or other relevant attributes.

Accurate extraction requires preserving the definitions, categories, and units used by the authors.

For example, a study reporting a mean age of 21.4 years should not be summarized as involving participants aged 21 to 22 years. A mean does not describe the full age range.

Similarly, a sample consisting of 65% female participants should not automatically be described as representative of women in the target population.

Participant characteristics are descriptive information. They do not, by themselves, establish population representativeness or generalizability.

Why Do Denominators Matter When Extracting Percentages?

A percentage is meaningful only when its denominator is understood.

Suppose a paper reports that 180 participants were enrolled, but demographic information was available for only 170. If 102 of those 170 participants were female, the reported percentage is 60%.

AI might incorrectly divide 102 by 180 and report 56.7%, or repeat 60% while implying that all 180 participants provided the demographic information.

Both interpretations would misrepresent the underlying data.

Researchers should preserve the relevant numerator and denominator whenever percentages are extracted, especially when demographic variables contain missing values.

Can AI Confuse Participant Characteristics Across Study Groups?

Yes. Studies involving multiple groups often present characteristics separately for intervention and comparison groups, cohorts, sites, or subpopulations.

For example, an educational intervention might include 80 students in the intervention group and 75 in the comparison group. Their mean ages, baseline achievement scores, and prior technology experience may differ.

An AI system may incorrectly combine group-specific values or attribute characteristics from one group to the entire sample.

Researchers should check whether the extracted information describes the overall sample, a particular group, or a subgroup used in an additional analysis.

What About Clustered Studies and Different Units of Analysis?

Some studies include several levels of observation. A school-based trial might involve 12 schools, 48 teachers, and 960 students.

These are not competing sample sizes. They describe different units.

If schools were randomized but student outcomes were analyzed, the number of randomized clusters and the number of students both matter.

AI may report only the largest count and overlook the structure of the study. This can conceal information relevant to statistical independence, precision, and the interpretation of findings.

Researchers should identify the unit of recruitment, assignment, measurement, and analysis when these differ.

Can AI Extract Sample Size From Tables and Flow Diagrams?

Some systems can interpret tables and figures, but accuracy depends on the document-processing capabilities and the quality of the source.

Participant flow diagrams are particularly important because they may report exclusions, withdrawals, losses to follow-up, and analysis populations not fully described in the narrative text.

Complex PDF layouts can introduce extraction problems. Numbers may be associated with the wrong column, group, time point, or outcome.

For instance, a table might report separate sample sizes for baseline, immediate posttest, and delayed follow-up. AI could mistakenly identify the baseline count as the number completing follow-up.

Where participant counts appear in tables or diagrams, verify them visually against the original document.

What Do Reporting Guidelines Require?

Reporting guidelines provide useful references for determining which participant information should be available.

CONSORT 2025 addresses participant flow in randomized trials, including numbers assigned, receiving interventions, lost or excluded, and included in analyses. STROBE addresses reporting participant numbers, descriptive characteristics, and missing data in observational studies.

For qualitative research involving interviews and focus groups, COREQ includes reporting considerations concerning participant selection, nonparticipation, and sample description.

These guidelines are design-specific. They should not be treated as interchangeable requirements for all research papers.

They also help clarify an important distinction: incomplete reporting does not necessarily prove that a study was conducted incorrectly, but it may prevent readers from evaluating the sample adequately.

Can AI Determine Whether the Sample Size Was Adequate?

Extracting the sample size is not the same as evaluating whether it was sufficient.

A study involving 30 participants may be appropriate for one qualitative investigation but insufficient for a particular quantitative analysis. A study involving thousands of observations may still have substantial bias or inadequate representation of its target population.

Sample adequacy depends on the research question, design, analytical objectives, expected precision, relevant assumptions, and disciplinary conventions.

In quantitative research, power analysis or precision-based planning may inform sample size decisions. In qualitative research, considerations such as information power, methodological purpose, and depth of inquiry may be more relevant.

AI should not infer that a study is methodologically strong or weak from participant numbers alone. Evaluating adequacy belongs to the broader task of critically appraising the study.

What If Participant Information Is Missing or Inconsistent?

Research articles sometimes report incomplete or conflicting participant counts.

The abstract may describe 250 participants, while the methods report 248 and the results include 239. These differences may reflect exclusions or analysis-specific populations, but the explanation should be established from the paper rather than invented.

AI should identify the discrepancy and report where each number appears.

When the available text does not explain the difference, the appropriate extraction is not a guessed reconciliation. It is a transparent statement that the reported counts differ and the reason cannot be determined from the available information.

Extracting Participant Information When the Numbers Change Across the Study

The example shows why accurate extraction requires more than selecting the most prominent participant count. Different numbers describe different stages and analytical purposes.

If the study did not report the demographic composition of the 382 complete cases, AI should not assume that it was identical to the composition of the 400-person descriptive sample.

Common Misconceptions About AI Sample Size Extraction

How to Extract Sample Sizes and Participant Characteristics Reliably With AI

The most effective approach is to treat participant information as structured data with clearly defined meanings, rather than asking AI for a single number.

This is particularly useful when preparing literature matrices, extracting evidence for systematic reviews, or comparing populations across studies.

A Reusable Prompt for Participant Data Extraction

Use a Structured Extraction Table

For research synthesis, it may be useful to record participant information in separate fields.

This structure reduces ambiguity when several papers use different conventions for reporting participant numbers.

However, do not assume that every study provides all these fields. Record information as not reported when appropriate, and distinguish missing reporting from an actual absence of participants or observations.

Verify Extracted Values Against the Original Document

Check the methods section, participant flow diagram, descriptive tables, and analysis-specific notes. If AI reports a percentage, confirm its numerator and denominator. If it reports an average, confirm the associated unit and whether the value represents a mean or median.

When extracting information for a systematic review or other consequential evidence synthesis, consider independent verification by another reviewer according to the review protocol.

The Cochrane Handbook discusses data collection procedures and methods for reducing extraction errors. AI assistance does not remove the need for quality control appropriate to the review.

Accurate participant extraction also supports understanding the research design and interpreting the study's reported findings, although neither task can be completed from participant numbers alone.

Before Using AI-Extracted Sample Size and Participant Information

Frequently Asked Questions About AI Sample Size and Participant Extraction

Accurate Sample Extraction Requires Understanding What Each Number Represents

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.