Can Generative AI Screen Full-Text Articles for a Systematic Review?

Can AI Read an Entire Research Article and Decide Whether It Qualifies?

After title and abstract screening, researchers must retrieve and examine the full texts of potentially relevant reports. This is where many difficult eligibility decisions occur. A promising abstract may turn out to describe an ineligible study design, a different population, or an intervention that does not meet the review's criteria.

Generative AI can process lengthy documents, locate methodological descriptions, and compare reported study characteristics with instructions. It may therefore appear capable of deciding which full-text articles should be included or excluded.

The difficulty is that full-text screening requires more than recognizing relevant words. Researchers must interpret eligibility criteria, resolve incomplete reporting, distinguish studies from their publications, and document defensible reasons for exclusion.

AI Can Support Full-Text Screening, but Final Decisions Need Independent Verification

How AI-Assisted Full-Text Screening Works

How Is Full-Text Screening Different From Title and Abstract Screening?

Title and abstract screening identifies reports that might satisfy the eligibility criteria. Full-text screening examines the available reports in sufficient detail to determine whether the underlying studies actually qualify.

This distinction changes both the information available and the responsibility involved. At the preliminary stage, an uncertain abstract may simply be retained. At the full-text stage, reviewers must investigate the uncertainty and reach a justified decision or record that eligibility remains unresolved.

The Cochrane Handbook identifies studies, rather than individual reports, as the primary units of interest. Multiple publications may describe one study, and information needed to establish eligibility may be distributed across those reports.

The earlier stage is addressed separately in AI-assisted title and abstract screening. Its validation results should not automatically be treated as evidence that the same system can reliably assess full-text reports.

What Information Can AI Examine in a Full-Text Article?

A model may be instructed to locate information about participants, study design, interventions, comparison groups, outcomes, setting, follow-up periods, and publication characteristics. These details may appear in different sections of the article, supplementary materials, or related reports.

For example, an abstract may describe a controlled educational intervention, while the methods section reveals that participants were assigned to groups according to existing class schedules rather than randomized allocation.

If the review includes only randomized trials, that distinction could determine eligibility. An AI model might correctly identify the allocation procedure, but it could also mistake the word "random" in another context for evidence of randomized assignment.

The researcher must verify whether the model's interpretation is supported by the actual methods.

What Does a Full-Text AI Screening Workflow Look Like?

A controlled workflow supplies the model with the relevant source document and explicit eligibility criteria. Rather than asking whether an article is simply relevant, the instructions require assessment of specific criteria and identification of supporting evidence.

This is a possible workflow, not a validated screening method by itself. Its reliability must be established before it is used to influence consequential decisions.

What Does Published Research Show About AI Full-Text Screening?

Khraisha and colleagues (2024) evaluated GPT-4 on systematic review tasks involving title and abstract screening, full-text assessment, and data extraction across different literature types and languages.

The study found that some full-text screening configurations using highly reliable prompts achieved performance approaching that of human reviewers. However, results varied by task and dataset. Performance assessments also became less favorable when accounting for chance agreement and imbalanced datasets.

The authors therefore cautioned against assuming that large language models could replace humans throughout systematic reviews. Their findings suggest potential under particular conditions, not general validation of autonomous full-text screening.

The distinction matters because a model performing well on a particular collection of reports may struggle with different disciplines, complex study designs, multilingual articles, or eligibility criteria requiring substantial interpretation.

Why Can Full-Text Documents Be Difficult for AI to Interpret?

A PDF is not necessarily a clean, machine-readable representation of a research article. Text extraction may disrupt columns, omit footnotes, misplace tables, or fail to capture scanned pages.

Even when a model accepts a long document, researchers should not assume that every part was processed accurately. Important details may appear in appendices, figures, supplementary files, or tables whose structure is lost during extraction.

Can AI Apply Every Inclusion and Exclusion Criterion Correctly?

Not reliably by assumption. Some criteria involve straightforward factual identification, while others require methodological interpretation.

For instance, identifying whether a study took place in a university may be relatively simple when the setting is explicitly stated. Determining whether a multicomponent intervention meets a narrowly defined intervention criterion may require much closer examination.

Researchers should distinguish between information extraction and eligibility judgment. A model may accurately reproduce a sentence from the methods section while drawing an incorrect conclusion from it.

Reliability therefore depends on whether the model can apply the actual eligibility criteria, including their exceptions and ambiguous cases.

How Should Missing Information Be Handled?

Missing information should not automatically become a reason for exclusion. A report may omit a methodological detail that is available in another publication, a trial registration, or information supplied by the investigators.

The Cochrane Handbook describes contacting investigators where appropriate and retaining studies awaiting classification when essential eligibility information cannot be obtained.

An AI model may help identify which details appear to be missing, but researchers should verify the omission and determine whether further investigation is warranted.

For example, "randomization not reported" is not necessarily equivalent to "study was not randomized." Treating those statements as interchangeable could lead to an unjustified exclusion.

What Are the Requirements for Independent Full-Text Assessment?

The Cochrane Handbook, Chapter 4, requires at least two people working independently to determine whether studies meet the eligibility criteria, with a predefined procedure for resolving disagreements. It specifically identifies independent assessment as essential for final eligibility decisions based on full-text reports.

This is an important distinction from initial title and abstract screening, where Cochrane guidance permits a single screener even though duplicate screening is desirable.

Researchers should not assume that an AI system automatically counts as a second independent person under this standard. The implications of replacing a human reviewer with AI require separate methodological consideration.

Other review organizations and disciplines may use different procedures, so the relevant standard should be identified before the workflow is established.

How Should Full-Text Exclusion Reasons Be Documented?

When a study is excluded after full-text assessment, the reason should be traceable to the review's predefined eligibility criteria.

PRISMA 2020 requires reporting of the study-selection process and identification of studies that might appear eligible but were excluded, together with explanations. Its explanation and elaboration guidance emphasizes documenting potentially contentious exclusions.

Researchers should avoid vague explanations such as "not relevant" when a more specific reason can be given. Examples include an ineligible study design, population, intervention, or comparison.

Where multiple criteria are unmet, a consistent approach to assigning the primary exclusion reason can help maintain coherent reporting. Cochrane guidance notes that one failed eligibility criterion is sufficient for exclusion and that criteria may be assessed in a defined order.

Using AI to Examine Full-Text Eligibility in an Education Review

How Might AI Interpret Different Articles?

What Would a Validation Exercise Reveal?

Suppose independent human reviewers establish that 30 of the 120 reports should be retained as eligible or awaiting clarification, while 90 are appropriately excluded. AI correctly retains 27 of the 30 but incorrectly excludes 3. It also retains 12 reports that the human reference process excludes.

The result means that AI missed 10% of reports that the reference assessment would have retained. It does not establish that the three missed reports would necessarily contribute to the final synthesis, but their premature exclusion could prevent appropriate eligibility determination.

Because full-text screening directly influences study inclusion, the team should investigate the missed reports and should not authorize autonomous exclusions on the basis of this result alone.

The example also illustrates why a binary validation exercise may simplify the real workflow. In practice, researchers may need to distinguish eligible studies, ineligible studies, and studies awaiting classification rather than forcing every unresolved case into a definitive category.

Misconceptions About AI Full-Text Screening

Use AI to Locate Eligibility Evidence, Not to Conceal Judgment

The most defensible starting point is to use generative AI as an information-location assistant. Ask it to identify where the article describes relevant study characteristics and to distinguish reported facts from provisional interpretations.

Researchers can then assess the evidence using the protocol's eligibility criteria and applicable independent-review procedures.

How Should Researchers Evaluate AI Full-Text Screening?

Validation should use representative full-text reports, including ambiguous cases and studies that fail different eligibility criteria. Researchers should compare AI classifications against independently established reference decisions and examine both missed eligible studies and incorrect exclusion reasons.

Overall accuracy alone is insufficient. The evaluation of screening errors should reflect the consequences of incorrect exclusions, while any claimed workload reduction should account for verification time.

When using an audit of AI-assisted decisions, researchers should recognize that a small convenience sample may fail to detect uncommon but consequential errors.

What Should Be Reported?

Document the AI system and relevant model details, the source documents supplied, instructions used, eligibility criteria assessed, validation procedures, human reviewer involvement, and final decision process.

Also preserve exclusion reasons and records of disagreements or corrections. These details allow readers to distinguish AI-generated recommendations from the decisions actually adopted in the review.

Transparent reporting of AI use should explain the system's methodological role rather than merely acknowledging that a chatbot was consulted.

Before Using AI for Full-Text Eligibility Assessment

Frequently Asked Questions About AI Full-Text Screening

AI Can Examine Full Texts, but Eligibility Still Requires Defensible Decisions

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.