How I’m Fighting Bias in Algorithms

Joy Buolamwini’s TED talk begins with a small technical failure that is difficult to dismiss as merely amusing: the Aspire Mirror she built could not recognize her face until she put on a white mask. She also recalls an earlier human–computer interaction project that failed to detect her. These are not simply examples of software behaving unpredictably. They make visible a boundary in the system’s field of vision: some people are treated as ordinary inputs, while others fall outside what it reliably recognizes. Buolamwini calls this the “coded gaze.” Technology does not look from a neutral nowhere; the people who build a system, choose its training material and define success help determine whom it can see.

The talk connects that experience to a wider chain of algorithmic decisions. When facial analysis or recognition is used in commercial settings, surveillance or public safety, an error is not just a frustrating interface. It can shape identity checks, suspicion and exposure to observation. The feminist question is therefore larger than whether a dataset appears more balanced. Which institutions gain the power to watch and classify people? Do those classified know how a system works? Can they challenge an error, and have affected communities helped set the purpose and limits of its use?

As a Black woman researcher, Buolamwini brings the intersection of race and gender into algorithm evaluation rather than treating them as unrelated test categories. The TED talk uses personal narrative to make that question legible to a public audience; her related research offers a route toward more systematic assessment. Gender Shades is especially useful alongside the talk: the presentation shows how one person encounters the limits of a technology’s vision, while the research turns differential error into a question that can be examined through evaluation. These works do different things; a public talk is not a substitute for a research paper.

The talk’s strength is to make “technology bias” a concrete human–machine relation and invite viewers without an engineering background to ask who is responsible for design. Its limits matter too. An approximately eight-minute public presentation published in 2017 cannot establish the risks of every model, dataset or application in use today. Nor can it replace current independent audits, constraints on public procurement, transparency requirements or routes of appeal. Watch it as an entrance rather than a final diagnosis: starting from a failed recognition, ask who sets a system’s goals, who bears its costs and who has the power to demand that it stop.

Watch and read further: TED talk and transcript; Gender Shades: intersectional accuracy research.