Are the fairness risks of learning analytics being underestimated?

Are fairness risks actually measurable, or just theoretical?

They are measurable, and a 2026 study provides the clearest numbers. Researchers built a student success prediction system using the Open University Learning Analytics Dataset (OULAD) and tested four machine learning models, with XGBoost achieving the highest accuracy (AUC = 0.947, meaning it correctly ranked at-risk students 94.7% of the time) [1]. But when they checked for fairness using two standard metrics — Demographic Parity Difference and Equal Opportunity Difference — they found clear bias against certain student groups. After applying mitigation techniques, they reduced those fairness gaps to near zero while losing less than 1% of predictive accuracy [1]. This shows that fairness risks are not only real but also fixable without sacrificing performance.

The same study also expanded fairness analysis beyond traditional demographics like race or gender. It included 'contextual fairness' proxies such as late-night study patterns and inactivity streaks, which can reflect socioeconomic pressures on international and working students [1]. This suggests that standard fairness audits may miss important sources of bias if they only look at obvious categories.

Are there hidden statistical traps that make fairness risks worse than they appear?

Yes, and a 2022 paper identifies three specific pitfalls that can silently distort learning analytics results: confounding bias, overcontrol bias, and collider bias [2]. Collider bias is especially dangerous because it can create fake correlations or hide real ones when you filter or condition on a variable that is affected by both the cause and the outcome you are studying. For example, if you only analyze students who received an intervention, you might mistakenly conclude that a certain behavior is harmless when it actually signals risk for the broader population [2]. These biases are easy to overlook in observational data, which is the norm in learning analytics since randomized experiments are often impractical.

This means that even well-intentioned fairness corrections could be undermined if the underlying statistical model is biased from the start. The paper recommends using directed acyclic graphs (DAGs) to visually map out causal assumptions and spot these traps before running analyses [2]. Without such precautions, fairness risks may be systematically underestimated.

What can institutions actually do to avoid underestimating fairness risks?

Multiple studies converge on the same answer: fairness must be built into the system from the design phase, not added as an afterthought. A 2023 editorial reviewing the field concluded that learning analytics can either 'exacerbate existing inequalities' or work to overcome them, depending on the assumptions and rules developers set [3]. The authors stress that there are 'no simple answers' and that ensuring fairness requires ongoing research and embedded practice, not a one-time checklist [3].

A 2025 study from a rural South African university adds practical recommendations: institutions should establish clear ethical guidelines and policies for AI and learning analytics, and provide training to improve faculty data literacy [4]. Students in that study reported that perceived risk and performance expectations influenced their engagement with AI tools, suggesting that transparency and trust are critical for fairness to work in practice [4].

Finally, a 2026 simulation study compared different AI disclosure policies for student assignments. It found that restorative, education-oriented policies led to stronger norm internalization and narrower fairness gaps, while strict punitive policies produced the widest gaps and lower performance [5]. This suggests that how institutions enforce fairness rules matters as much as the rules themselves.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2022 to 2026, 3 from 2024 or later, 2 in Q1 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 61 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions