No one has just one identity: age, gender, race, class, culture, all at once. These identity factors overlap and combine, which is intersectionality.
Crenshaw (1989) coined it studying Black women who lost discrimination cases because the law treated race and sex as separate boxes; their experience fell in the gap between the two.
Factors interact rather than stacking up neatly. A young disabled woman's experience online is its own thing, often worse at the overlap than any single factor would predict.
Matters because digital systems, like Crenshaw's courts, often see one factor at a time or none: a system built for a typical user works against those at the overlap, the people least likely to have been in the test group.
Identity Factors Shape Online Experience
Two people on the same platform can have different experiences, because identity factors change what the system offers and how safe it feels.
Language: a widely supported language gets the full interface, help, and moderation; a minority language gets rough machine translation and little moderation, so abuse can slip through unnoticed.
Ability: a blind user can use an app only if it follows accessibility standards (labelled buttons and images for screen readers). Age shapes the experience too.
Through power: whoever designs the system decides, often unintentionally, who moves through it easily and who gets stuck. A choice that feels neutral still leaves some users out.
Example
Example
A screen-reader user can navigate an app easily only if it was built with accessibility in mind.
With unlabelled buttons and images the same app is almost unusable. The code has not changed, only who can get through it.
Representation: Who Gets Seen
Representation = whose faces, voices, and stories appear, both in media and in the training data. Underrepresented groups get treated as edge cases.
Not only on-screen: it is also who is in the data and who is in the room. A team that shares one background may never notice how a product fails people unlike them.
Example: early facial systems trained on lighter-skinned faces misread darker-skinned users, and some cameras asked Asian users if they had blinked, treating whole groups as edge cases.
Generalises beyond faces: stereotyped image-search results, accent gaps in voice assistants. Uneven data means an uneven system.
Case study
System: facial recognition identifies or verifies people using algorithms trained on large sets of face images.
Specifics: the 2018 Gender Shades study (Buolamwini and Gebru) tested IBM, Microsoft, and Face++, finding error rates under 1% for lighter-skinned men but up to ~35% for darker-skinned women, because the training data underrepresented that group.
Impacts and implications: the least-represented group was far more likely to be misidentified, which matters for policing, unlocking devices, and border control; the study pushed several companies to improve or pause their systems.
Concepts: identity (a face treated as a key), power (those least represented carry the most risk), values and ethics around fairness; social and political contexts.
Algorithmic Bias and Marginalised Identities
Algorithmic bias = systematically unfair outcomes for certain groups, usually from biased or unbalanced data rather than a biased programmer, repeated at scale.
Hits marginalised identities hardest: Amazon scrapped a 2018 hiring tool that downgraded CVs with the word "women's". A proxy variable (postcode standing in for race) lets a system discriminate without being told the protected trait.
COMPAS (ProPublica, 2016): a US court reoffending algorithm reportedly flagged Black defendants high-risk far more often than white defendants; the company disputed it, and it became a landmark on what algorithmic fairness even means.
Through power: a biased human decision can be questioned, but a biased algorithm is trusted as objective and applied to millions at once. Because the same model runs on millions of cases, one hidden bias is applied to everyone at once.
Common Mistake
Common mistake
An algorithm is not fair just because it is mathematical.
It encodes the gaps and bias of its training data and can hide bias in proxy variables; calling a decision data-driven can disguise who it disadvantages.
The Digital Divide and Identity
Before a system can treat you at all, you have to reach it. The digital divide is the gap between those with reliable access and skills and those without, and it tracks identity closely.
Access thins along income, age, rural location, and disability; because these overlap, the divide is itself intersectional, hitting hardest those at several disadvantages at once.
COVID-19 (2020): when schooling moved online, students without a home computer or reliable internet (the homework gap) fell behind their better-connected classmates. Moving school online widened a gap that was already there.
Not only owning a device: it includes connection speed and reliability, the cost of data, skills and confidence, and accessible design. Closing the gap means more than handing out hardware.
Active recall
Self review
What is intersectionality, who coined it, and why do identity factors combine rather than add up?
Give two ways identity factors can change a person's experience of the same platform.
What did the Gender Shades study show about representation in training data?
What is algorithmic bias, where does it come from, and what is a proxy variable?
Why is the digital divide described as intersectional, and what did the 2020 homework gap reveal?
