03 · What You Need to Know
How to Develop Defensible Inclusion and Exclusion Criteria With AI
What Are Inclusion and Exclusion Criteria?
Inclusion criteria specify the characteristics that individuals or other study units must satisfy to be eligible for participation.
Exclusion criteria identify conditions that make otherwise potentially eligible units inappropriate for the study under the defined protocol.
These criteria may concern population membership, relevant experience, intervention suitability, measurement requirements, participant safety, or other justified methodological conditions.
Inclusion criteria
Conditions that must be satisfied for a person or unit to qualify for the study.
Exclusion criteria
Additional conditions that prevent participation despite otherwise meeting relevant eligibility requirements.
For example, a study of currently enrolled undergraduate students may require participants to have undergraduate enrollment status during a specified academic period.
Whether prior AI experience should also be required depends on the research question. It may be necessary for a study of experienced users but inappropriate for estimating AI adoption across the student population.
How Are Eligibility Criteria Different From the Target Population?
The target population identifies the population about which the study intends to draw conclusions. Eligibility criteria operationalize relevant boundaries and other necessary participation requirements.
For example, a researcher may define the target population as all undergraduate students enrolled at a specified university during the current academic year.
The inclusion criteria may require verified enrollment at that institution and undergraduate status during the defined period.
Additional exclusions should not contradict the population definition without a clear methodological or ethical justification.
Researchers should therefore define the target population before asking AI to develop eligibility rules.
What Makes an Inclusion Criterion Appropriate?
An appropriate inclusion criterion should have a clear relationship to the research question, study population, intervention, or evidence required.
It should also be sufficiently precise that eligibility can be determined consistently.
For example, "must be a university student" may be ambiguous when the study concerns only undergraduate students enrolled during a particular semester.
A more operational statement would identify the institution, educational level, and relevant enrollment period.
However, precision should not be confused with adding unnecessary restrictions. A criterion can be very specific while remaining scientifically unjustified.
What Makes an Exclusion Criterion Appropriate?
An exclusion criterion should address a legitimate reason why participation would be inappropriate or incompatible with the study's objectives or procedures.
For example, an intervention study may exclude individuals for whom the intervention presents a specific, documented safety concern or cannot be administered under the approved protocol.
In educational research, exclusion may sometimes be justified by the absence of a necessary experience or by incompatibility with a clearly defined study condition.
However, researchers should examine whether a proposed exclusion is genuinely necessary or merely simplifies recruitment and analysis.
Excluding participants because their responses might be unusual, their circumstances are inconvenient, or their outcomes could complicate interpretation is not a defensible basis by itself.
Can AI Invent Arbitrary Eligibility Thresholds?
Yes. AI may generate numerical thresholds that sound methodologically precise but have no established justification.
For example, it may recommend including only students who have used generative AI for at least six months.
Such a threshold may be relevant if the study specifically investigates sustained experience with AI tools and the duration is theoretically or empirically justified.
However, it would be arbitrary if the researcher simply wants to examine current patterns of AI use.
Likewise, an age threshold, minimum grade point average, or required number of weekly AI interactions should not be adopted merely because AI suggests it.
Each restriction should be supported by the research objective, relevant evidence, or a documented ethical or operational requirement.
Why Can Excluding Nonusers Distort an AI Adoption Study?
Suppose a researcher wants to estimate the proportion of university students who use generative AI for academic activities.
If the inclusion criteria require previous AI use, the study excludes students who have never used the technology.
The resulting sample may provide information about users' experiences, but it cannot directly estimate the prevalence of use among all eligible students.
This is not merely a wording problem. The eligibility criteria have changed the population represented by the study.
AI should therefore distinguish between characteristics that define the target population and variables that the study intends to measure.
Should Researchers Exclude Participants With Limited Technology Access?
Not automatically. Technology access may be an important characteristic of the population or a factor associated with the outcome.
For example, a study examining inequalities in generative AI adoption would be seriously limited if students with poor connectivity were excluded simply because online recruitment was easier.
In an intervention study requiring access to a particular platform, researchers may need to determine whether suitable devices or connectivity can be provided.
Exclusion may sometimes be unavoidable, but its rationale and implications should be documented.
The researcher should consider whether the restriction changes the target population, introduces selection problems, or limits the applicability of findings.
Can Eligibility Criteria Introduce Selection Bias?
Yes, depending on the study's target inference and the mechanisms through which eligibility is determined.
For example, a study comparing academic outcomes among AI users and nonusers might exclude students with incomplete academic records.
If record completeness is related to AI use, achievement, or their common causes, the exclusion could distort the observed association.
However, not every eligibility restriction is selection bias in the same technical sense. Some restrictions primarily narrow the target population or affect generalizability.
Researchers should examine the mechanism rather than assume that every exclusion automatically creates bias.
A focused assessment of sampling bias can help evaluate these consequences.
Should Eligibility Criteria Be the Same in Experimental and Observational Research?
Not necessarily. Different designs may require different eligibility considerations.
In an experimental study, criteria may address whether participants can appropriately receive the intervention, whether the comparison conditions are meaningful, and whether safety or ethical restrictions apply.
In observational research, criteria may focus on population membership, exposure definitions, relevant observation periods, and the availability of essential information.
However, researchers should be cautious about defining eligibility using characteristics that occur after the exposure or intervention.
Post-exposure selection may introduce bias when eligibility depends on consequences of the exposure or factors related to the outcome.
Hernán and Robins (2020) emphasize the importance of defining eligibility and time zero consistently when specifying causal questions and emulating target trials.
Why Is the Timing of Eligibility Assessment Important?
Eligibility criteria should be evaluated at a clearly defined point appropriate to the study design.
For example, a trial might establish eligibility before random assignment and then follow participants for subsequent outcomes.
Excluding participants after assignment because they did not adhere to the intervention can change the comparison and potentially introduce bias.
Similarly, an observational study may create selection problems if eligibility is defined using information that becomes available only after the outcome process has begun.
AI should therefore help researchers distinguish baseline eligibility from post-enrollment events, withdrawal, missing data, and analytical exclusions.
These are related but methodologically different decisions.
Are Consent Requirements the Same as Inclusion Criteria?
Consent and eligibility are related but distinct.
Eligibility concerns whether an individual satisfies the study's defined criteria. Informed consent concerns whether participation is authorized through an appropriate ethical process.
For example, a student may meet all scientific eligibility requirements but decline participation.
The student remains eligible in principle but does not become an enrolled participant.
Researchers should avoid treating refusal to consent as though it were a scientific characteristic defining the target population.
Consent procedures must follow applicable ethical and institutional requirements, including additional safeguards when participants are minors or otherwise require special protections.
Can AI Recommend Excluding Participants With Disabilities or Particular Personal Characteristics?
AI may suggest exclusions based on assumptions about participant capability, intervention suitability, or measurement convenience.
Such recommendations require careful scrutiny.
Excluding participants with disabilities or other personal characteristics without a legitimate scientific or ethical rationale may create unjustified restrictions and undermine equitable participation.
Researchers should consider whether reasonable accommodations or alternative procedures can enable participation while preserving the study's methodological requirements.
The Belmont Report identifies justice as an important principle in participant selection, alongside respect for persons and beneficence.
Eligibility decisions should therefore consider fairness and participant protection, not simply operational convenience.
Should Eligibility Criteria Be Written Before Recruitment Begins?
Generally, yes. Prespecifying criteria can improve consistency, transparency, and the interpretation of the resulting sample.
For intervention trials, SPIRIT 2025 provides guidance on specifying eligibility criteria within trial protocols.
Clear criteria allow researchers to establish screening procedures and document why individuals were included or excluded.
Changes may sometimes be necessary, but they should be justified, documented, and reviewed under applicable protocol and ethics requirements.
Researchers should not revise eligibility rules after seeing outcomes merely to obtain more favorable results.
How Can AI Help Make Eligibility Criteria Operational?
AI can help convert vague population descriptions into criteria that can be applied consistently.
| Vague criterion |
Potential problem |
More operational approach |
| University students |
Educational level and enrollment period are unclear. |
Specify current enrollment status, institution, and relevant academic level. |
| Experienced AI users |
Experience has no defined meaning. |
Specify the relevant activity and justify any duration or frequency requirement. |
| Students with good internet access |
Good access is subjective and may introduce unnecessary exclusions. |
Define the actual technical requirement and consider whether access can be provided. |
| Students with complete data |
Completeness may be determined after enrollment or exposure. |
Separate baseline eligibility from missing-data handling and analytical decisions. |
| Students willing to participate |
Consent is being confused with population membership. |
Specify scientific eligibility separately from informed consent and enrollment. |
Operational clarity is valuable, but every criterion still requires a substantive rationale.
Can AI Verify That Eligibility Criteria Are Ethically Appropriate?
AI can identify possible concerns involving fairness, participant burden, privacy, and the justification for exclusions.
However, it cannot independently establish ethical acceptability or replace institutional review.
For example, restricting an educational intervention to students with particular academic characteristics may be appropriate for a narrowly defined research question but inappropriate if it unfairly excludes individuals without scientific justification.
Researchers should evaluate the criteria against applicable ethical principles, institutional requirements, and the intended population.
Where relevant, independent methodological review should examine whether eligibility rules introduce unnecessary limitations.
Watch Out
Do not adopt AI-generated eligibility criteria merely because they are specific, measurable, or conventionally worded. An arbitrary exclusion can be operationally precise while remaining scientifically unjustified or ethically problematic.
How Can You Audit AI-Generated Inclusion and Exclusion Criteria?
1. Define the target population. Establish who or what the research intends to study before drafting eligibility rules.
2. Identify essential inclusion conditions. Specify characteristics required by the research question and design.
3. Examine proposed exclusions. Require a methodological, ethical, or operational justification for each restriction.
4. Check operational clarity. Determine whether eligibility can be assessed consistently using available information.
5. Evaluate unintended consequences. Examine whether the criteria exclude important groups or introduce selection problems.
6. Separate eligibility from enrollment and analysis. Distinguish screening criteria, consent, withdrawal, missing data, and post-enrollment exclusions.
7. Verify ethical requirements. Review participant protections and institutional obligations.
8. Document and prespecify the criteria. Record the rationale and obtain appropriate approval before recruitment.
The resulting criteria should be sufficiently clear for consistent screening without narrowing the population beyond what the research requires.
