Statistics, long considered the bedrock of academic inquiry, has evolved from a tool for understanding populations to an engine driving artificial intelligence and shaping societal structures. While statistical methods are essential for validating research, balancing educational metrics, and guiding critical decisions, their inherent reliance on averages and dominant patterns can inadvertently create and exacerbate inequalities, particularly for those who fall outside the statistical norm. This reliance, amplified by the pervasive integration of artificial intelligence, risks further marginalizing already disadvantaged groups.
The Statistical Foundation of Knowledge and AI
At its core, statistics provides the quantitative framework that lends credibility to research findings. It allows scholars to extrapolate insights from sample groups to broader populations, transforming anecdotal observations into generalizable evidence and distinguishing conjecture from demonstrable truth. Beyond the ivory tower, statistical principles are routinely applied in academic institutions to standardize grading, identify high-achieving students and faculty, and inform institutional strategy.
However, the influence of statistics extends far beyond academia. It is the fundamental engine powering artificial intelligence. AI decision-making systems leverage statistical models to identify patterns, make predictions, and match data to specific targets. In the realm of natural language processing, statistics predicts the likelihood of subsequent words or phrases in large language models, influencing everything from search engine results to content recommendations, optimizing delivery based on engagement metrics like attention and popularity. In essence, at its most fundamental level, AI operates as a sophisticated statistical replicator, learning from and reproducing patterns found in vast datasets.
The "Out of Distribution" Problem: When Data Deviates
Despite the ubiquitous application and perceived objectivity of statistical analysis, a significant concern arises from its inherent bias towards the majority or the "average." Jutta Treviranus, Director of the Inclusive Design Research Centre and Professor of Inclusive Design at OCAD University, has dedicated her 45-year career to inclusive design and higher education, with her research consistently guided by individuals who are "out of distribution." These are the people and communities whose data points are often considered noise or outliers within population datasets. Their circumstances are frequently too complex, heterogeneous, or unpredictable to achieve statistical significance or to isolate causal factors within traditional analytical frameworks. Alarmingly, these individuals and communities constitute a substantial and growing proportion of the global population, a trend exacerbated by recurring human crises.
Treviranus’s observations predate the widespread adoption of AI, yet her experiences highlight how statistical methodologies, even in their nascent stages, have historically disadvantaged marginalized minorities while benefiting the majority, thereby contributing to persistent disparities. For decades, she has meticulously collected data by posing a simple yet profound open-ended question to those she encounters: "What do you need to thrive and participate fully?" The resulting dataset, by its very nature, is multi-faceted and defies easy categorization. To visualize this complexity, Treviranus employs a high-dimensional multivariate scatterplot, which she terms the "human starburst."
The "Human Starburst" and the Concentration of Needs
The "human starburst" reveals a stark distribution pattern: approximately 80% of the data points are clustered within the central 20% of the plotted space, representing the needs and circumstances of the majority. The remaining 20% of the data points are widely dispersed across the peripheral 80% of the space. It is in these peripheral regions, where data points become increasingly sparse, distant, and heterogeneous, that the needs of individuals with disabilities, those facing intersectional barriers, and people experiencing significant challenges are found.
This distribution pattern has profound implications for resource allocation and societal design. Due to economies of scale, most systems, services, and products are engineered to cater to the needs concentrated in the central cluster. Consequently, they function adequately for the majority. However, as an individual’s needs diverge from this central norm, the efficacy of these systems diminishes. For those at the far periphery, whose needs are outliers, these systems often fail entirely. This pattern permeates every facet of society, dictating priorities and value judgments across markets, service provision, education, employment, and media. Crucially, it also shapes what is considered scientific truth about a population. A statistically based finding, while accurate for the hypothetical "average" person, becomes increasingly inaccurate as one moves away from that average and is fundamentally wrong for those at the edges of the data distribution.
Historical Roots of the "Average Man"
The statistical reliance on the average individual is not a recent phenomenon. Its theoretical underpinnings can be traced back to the 19th-century French mathematician Adolphe Quetelet. Quetelet, whose work on "average man" theories was later infamously co-opted to justify eugenics, posited that an individual embodying all the qualities of the "Average Man" would represent "all that is great, good, or beautiful." Conversely, he argued that "everything differing from the Average Man’s proportions and condition, would constitute deformity and disease." This historical framing reveals a deep-seated bias within statistical thought, one that inherently pathologizes deviation from the norm.
AI: Amplifying Existing Inequities
The advent of artificial intelligence, rather than correcting these statistical biases, has the potential to amplify, accelerate, and automate them. In educational settings, as in virtually all other domains, AI is poised to exacerbate the challenges faced by those already struggling while further benefiting those who are already well-positioned.
The integration of AI into educational technology is already pervasive. Learning management systems, admissions platforms, proctoring software, plagiarism detectors, and productivity tools all increasingly incorporate AI functionalities. In their current iterations, these systems are designed to identify and reinforce optimal patterns. Consequently, any deviation from these statistically determined optima, which inherently includes diversity, is likely to be discouraged and eventually eliminated. Differences from target patterns are often flagged as suspicious.
Consider the implications: AI-powered instructional tutors may reshape divergent learners to conform to statistically determined optimal learning paths. Proctoring systems can flag non-standard behavior as suspicious, potentially penalizing students with unique learning styles or those experiencing anxiety. Students might be presented with content predicted to be most effective based on aggregate data, potentially bypassing novel or less common learning needs. Admissions departments can utilize AI to identify applicants who mirror historical patterns of success, potentially overlooking unconventional but promising candidates. Student support services might receive AI-generated responses that are optimized for the average student, neglecting those with more complex or atypical needs. AI-driven hiring tools can filter out candidates whose profiles differ from those of past "optimal" employees, thereby perpetuating workforce homogeneity. Productivity monitors can penalize deviations from prescribed optimal patterns, even when such deviations are necessary to accommodate students with anomalous needs. An education system thoroughly infused with current AI tools risks mechanizing and perpetuating the discriminatory ideals of the eugenicist past.
The Paradox of Ethical AI and Statistical Insignificance
Even systems designed to monitor and certify ethical AI practices are often reliant on statistical frameworks. In this context, risks and harms experienced by outliers or marginalized minorities are frequently dismissed as statistically insignificant or relegated to the realm of anecdotal evidence. This creates a paradoxical situation where the very tools intended to ensure fairness are blind to the systemic disadvantages faced by those outside the statistical norm.
Reimagining AI: A Call for Inclusive Design
The pervasive influence of AI, acting as a magnifying mirror of our societal conventions and assumptions, presents a critical opportunity to reflect on and rebalance pervasive inequities. At the Inclusive Design Research Centre, a collaborative approach involving the disability community and other partners is central to identifying and addressing key accessibility challenges. The fundamental premise is that humans remain in control of AI development and deployment.
This control offers the potential to design AI systems that actively value difference. Instead of algorithms driven by data exploitation, the focus can shift to data exploration, enabling AI to uncover missing perspectives, particularly in areas like admissions and hiring. By adjusting our metrics and algorithmic priorities, we can begin to value the "human edges"—the periphery of the data distribution. It is at these edges where early warning signs of societal crises often emerge, where the greatest diversity of human experience resides, and where the most generative ideas for truly innovative change are likely to be found.
The challenge lies in moving beyond a statistical paradigm that prioritizes the average and instead embracing a design philosophy that recognizes and leverages the richness of human variation. This requires a conscious effort to build AI systems that are not merely efficient for the majority but are inclusive and equitable for all, including those who have historically been rendered invisible by the tyranny of the average. The future of AI, and indeed of society, depends on our ability to shift from a model of statistical replication to one of inclusive exploration and amplification of diverse human experiences. The opportunity is present to create AI that serves humanity in its entirety, rather than merely optimizing for a statistically defined ideal.
