03 · What You Need to Know
What Researchers Should Include When Reporting AI Fairness Limitations
Fairness Reporting Is More Than Acknowledging That AI Can Be Biased
Statements such as "AI systems may contain bias" are common but methodologically incomplete.
They do not explain which bias was investigated, whether it was observed, or how it could affect the research.
For example, an educational prediction model may exhibit lower sensitivity among students in asynchronous programs.
This is a specific performance limitation that readers can evaluate.
By contrast, a general warning about algorithmic bias provides no information about the model's actual behavior.
Researchers should connect fairness limitations to the evidence generated during the study.
Describe the AI System and Its Intended Research Function
Fairness findings cannot be interpreted independently of the task performed by AI.
Researchers should identify whether the system classified participants, predicted outcomes, recommended literature, translated materials, or supported another analytical activity.
Relevant technical information may include the model name or identifier, version, provider, configuration, evaluation date, and important preprocessing procedures.
For generative AI, consequential prompts and instructions may also need to be documented.
Researchers should distinguish the model's general capabilities from the specific function evaluated in the study.
A fairness assessment for one task should not be presented as evidence of fairness for unrelated applications.
Identify the Populations Included in the Evaluation
Reports should describe the populations represented in the development and evaluation datasets.
Relevant characteristics may include language, institutional setting, age, educational program, or other scientifically justified variables.
For example, a model developed using conventional university students may have limited validation evidence for working learners enrolled in asynchronous courses.
Researchers should report the subgroup definitions and explain why those comparisons matter.
They should also identify important populations that were not evaluated.
The absence of subgroup evidence should not be interpreted as evidence of equivalent performance.
Explain Why Particular Fairness Metrics Were Selected
Fairness metrics evaluate different properties of a system.
Demographic parity concerns positive classification rates, equal opportunity concerns true-positive-rate parity, and equalized odds examines both true-positive and false-positive rates.
Calibration evaluates the relationship between predicted probabilities and observed outcomes.
Researchers should explain which metrics were selected and why they are appropriate for the intended application.
For example, sensitivity may be especially important when failing to identify students who need support could prevent them from receiving assistance.
The choice should be justified by the scientific purpose and consequences of errors rather than by whichever metric produces the most favorable result.
Report Subgroup Performance Alongside Aggregate Results
Overall accuracy can conceal poor performance among smaller populations.
Researchers should therefore report relevant subgroup results when differential performance could affect scientific conclusions or consequential decisions.
For classification, this may involve sensitivity, specificity, false-positive rates, false-negative rates, or precision.
For probabilistic prediction, subgroup calibration and discrimination may be important.
Relevant subgroup sample sizes and outcome counts should accompany performance estimates.
The need to examine AI performance across population subgroups is particularly important when aggregate metrics are dominated by a large majority group.
Report Absolute Differences and Underlying Counts
Percentages can obscure the number of observations supporting an estimate.
Suppose a model has a false-negative rate of 10% in one group and 20% in another.
The absolute difference is ten percentage points, while the second group's rate is twice that of the first.
Both descriptions may be useful, but neither communicates the precision of the estimate without information about sample size.
Researchers should report underlying counts where appropriate.
This is especially important when subgroup evaluations contain relatively few positive or negative cases.
Statistical Uncertainty Should Be Explicit
Subgroup performance estimates are subject to sampling uncertainty.
Researchers should report suitable confidence intervals or other uncertainty measures for relevant estimates and differences.
A small observed disparity may be imprecisely estimated, while a substantial difference may remain scientifically important even when conventional statistical significance is not achieved.
Researchers should avoid interpreting a nonsignificant difference as proof of equal performance.
When the evaluation sample is insufficient to establish equivalence, that limitation should be stated directly.
Uncertainty should be considered alongside the magnitude and potential consequences of the disparity.
Distinguish Observed Disparities From Their Possible Causes
An observed difference in performance does not independently establish why the difference occurred.
Possible contributors include development-data representation, measurement validity, label quality, language differences, or deployment conditions.
For example, lower classification accuracy for Filipino-language responses may reflect difficulty interpreting code-switching, differences in response characteristics, or inadequate reference labeling.
Researchers should distinguish causes supported by empirical investigation from plausible explanations requiring further testing.
Statements that a model discriminates against a group should be supported by an appropriate fairness analysis rather than inferred solely from one performance difference.
Explain Whether the Reference Outcome Is Valid
Fairness metrics depend on the outcomes or labels used as reference standards.
If the reference outcome is inappropriate, numerical performance comparisons may be misleading.
For example, defining student engagement through login frequency may inadequately represent learners who study offline.
Researchers should describe how reference outcomes were established and whether their meanings are comparable across populations.
Where measurement equivalence has not been evaluated, the limitation should be acknowledged.
Equal accuracy against an invalid reference does not establish fair or scientifically meaningful classification.
Fairness Metrics May Conflict and Should Be Reported Accordingly
Different fairness criteria may produce different assessments of the same model.
For example, a model may be calibrated across groups while exhibiting unequal classification error rates.
Under particular conditions, some fairness properties cannot all be satisfied simultaneously.
Researchers should explain relevant trade-offs rather than reporting only the criterion that the model satisfies.
The implications of conflicting AI fairness metrics should be discussed when they affect model selection or interpretation.
A transparent report identifies what was prioritized and what remained unresolved.
Explain Who Selected the Fairness Criteria
The selection of fairness criteria involves scientific and ethical judgments.
Researchers should describe how relevant metrics and performance requirements were chosen.
For consequential applications, this may involve methodological specialists, research leadership, ethics oversight, and consultation with affected populations.
Researchers should distinguish advisory participation from formal decision authority.
Where competing perspectives influenced the choice, the rationale should be documented.
The process discussed in choosing appropriate AI fairness definitions can support transparent reporting of these decisions.
Report Fairness Limitations in the Methods Section
The methods section should explain how fairness was evaluated.
Relevant information includes subgroup selection, reference-label procedures, evaluation samples, performance metrics, uncertainty estimation, and any predefined thresholds.
Researchers should identify whether comparisons were planned or exploratory.
For generative AI, the methods should describe the prompts or evaluation scenarios used to investigate differential behavior.
These details help readers determine whether the fairness assessment is reproducible and appropriate to the research question.
Report Observed Disparities in the Results Section
Material fairness findings belong in the results rather than being mentioned only as general limitations.
For example, a study should report a substantial subgroup sensitivity difference when it directly affects the evaluation of the proposed model.
The results should distinguish estimates from interpretations and provide sufficient numerical information.
Where subgroup comparisons were exploratory, that status should be made clear.
Researchers should avoid selectively reporting only favorable metrics.
Unfavorable but scientifically relevant results are part of the evidence needed to assess model suitability.
Discuss the Consequences in the Discussion Section
The discussion should explain what fairness limitations mean for the research findings and intended application.
For example, lower sensitivity for working students may limit confidence in using a model to identify academic support needs within that population.
Researchers should consider whether the limitation affects prevalence estimates, group comparisons, or the validity of recommendations.
They should also examine whether alternative procedures or additional verification could reduce the consequences.
The interpretation should remain proportionate to the available evidence.
Limitations Should Distinguish What Was Tested From What Was Not
A fairness evaluation may examine only selected populations or metrics.
For example, a study may evaluate gender-related performance but lack sufficient data to examine disability-related or intersectional differences.
Researchers should identify these boundaries without implying that every untested group necessarily experiences poor performance.
Likewise, favorable results for evaluated groups should not be generalized to populations absent from the validation sample.
Specific statements about untested conditions are more informative than broad disclaimers.
Report Mitigation Procedures and Their Effects
Researchers may attempt to reduce disparities through revised training data, threshold adjustments, recalibration, or additional human verification.
Reports should explain which problem each intervention was intended to address.
They should also provide evidence of whether the intervention improved relevant performance.
For example, lowering a classification threshold may increase sensitivity while also increasing false positives.
Researchers should report these trade-offs rather than describing mitigation as successful solely because one metric improved.
Evaluation should use suitable validation data and distinguish development-stage improvements from independent confirmation.
Disclose When Fairness Mitigation Was Not Successful
Not every fairness-related problem can be corrected within the available resources or research design.
Researchers may find that a model remains unreliable for a particular population despite attempted adjustments.
Such findings should be reported accurately.
Depending on the consequences, the appropriate response may involve restricting the model's use, obtaining additional data, or selecting another procedure.
Researchers should not describe unresolved limitations as minor merely because the overall model performance remains strong.
Fairness Reporting Should Address Downstream Decisions
Model performance is only one component of the broader research or institutional workflow.
For example, an AI model may identify students needing support with comparable sensitivity across groups while the resulting services remain inaccessible to learners with employment obligations.
The classification may satisfy a selected fairness metric while the intervention process produces unequal benefits.
Researchers should report relevant limitations associated with the intended use, not only the numerical properties of the model.
Where downstream consequences were not evaluated, that boundary should be stated.
Generative AI Fairness Requires Task-Specific Reporting
Not every AI-assisted research task involves classification or prediction.
Generative AI may provide methodological advice, summarize research literature, translate participant responses, or support qualitative coding.
Fairness concerns in these tasks may involve systematic omissions, stereotypical assumptions, or unequal informational quality.
Researchers should describe how such differences were evaluated.
For example, matched-prompt testing may be appropriate when investigating whether AI gives different methodological advice to otherwise comparable hypothetical researchers.
Researchers should not apply classification fairness metrics to tasks for which they are conceptually inappropriate.
Model Cards and Structured Documentation Can Support Transparency
Model cards provide a framework for describing intended uses, evaluation procedures, performance characteristics, and limitations.
They can help researchers document subgroup findings and communicate appropriate boundaries of model use.
However, a completed model card does not independently establish fairness.
Its value depends on the quality and completeness of the evidence included.
Researchers should select documentation practices appropriate to their discipline and application.
Relevant Reporting Guidelines Should Be Used Where Applicable
Different research fields have developed reporting guidance for AI and prediction models.
For example, TRIPOD+AI provides guidance for reporting clinical prediction models developed or evaluated using regression or machine-learning methods.
Other disciplines may have different requirements.
Researchers should follow applicable journal, funder, institutional, and disciplinary guidance.
They should not assume that a reporting framework developed for one domain automatically applies in full to every AI-assisted research project.
Fairness Limitations Should Influence the Conclusion
A manuscript should not acknowledge substantial fairness limitations and then recommend unrestricted deployment without addressing them.
For example, if a student-risk model has not been adequately validated for asynchronous learners, the conclusion should not claim that it is suitable for all university students.
Researchers should align their recommendations with the populations and conditions supported by the evaluation.
Where uncertainty remains substantial, further validation may be the most defensible recommendation.
Transparent limitations should meaningfully constrain scientific claims rather than function as a routine paragraph before an otherwise unqualified conclusion.
