“Neutrality in tech is a myth.”
Journalist and computer scientist Meredith Broussard uses this line to challenge a familiar story: that technology merely applies objective rules while human bias exists somewhere outside the machine. Every technical system begins with choices. Someone defines the problem, selects the data, decides what can be measured, sets acceptable error rates, and determines who must live with the result. Even a mathematically correct model can support an unjust process. Broussard is not claiming every engineer intends harm. Intentions still cannot erase the values built into design and deployment.
Data are shaped by history and institutions. A hiring dataset records earlier decisions about who was recruited, promoted, or excluded. Health records reflect differences in access, diagnosis, and treatment. Arrest data reflect policing practices, not a neutral census of wrongdoing. When these records become training material, an algorithm may reproduce their patterns with the authority of computation. Calling the output data-driven does not answer whether the target is appropriate, the categories are fair, or the underlying activity should be automated at all.
Interfaces also distribute power. A system that gives managers a risk score but offers workers no explanation privileges one perspective. A form that requires everyone to fit a fixed identity category converts an administrative choice into an apparent fact. Defaults, thresholds, and appeal routes determine whose inconvenience counts. Neutrality language can make these decisions difficult to see because it presents the finished tool as inevitable. Responsible practice surfaces them in documentation, testing, public participation, and procurement requirements before the system becomes infrastructure.
Broussard’s critique leads to better engineering rather than rejection of technology. Teams can examine data provenance, test errors across relevant groups, include people affected by the system, and provide meaningful human review. They can choose a simpler method when complexity adds opacity without value. They can decide not to automate a task that depends on contested social judgments. Evaluation should ask not only whether the model works as specified, but whether the specification serves a legitimate purpose and whether people have power to contest its effects. Broussard’s sentence replaces a comforting myth with a practical obligation. If technology carries choices, those choices can be documented, debated, tested, and changed before neutrality becomes an excuse for avoidable harm. The work includes examining who defines success, who is missing from the data, who can appeal, and who has authority to pause deployment. Institutional power must always remain publicly visible to affected communities.
Problem definitions, data sources, targets, categories, and acceptable errors all reflect human priorities.
A technically correct implementation can still enforce a poor or unjust specification.
Records of hiring, health care, education, or policing carry the institutions and inequalities that produced them.
Training on those records can reproduce patterns while making them appear objective or inevitable.
Documentation, participation, subgroup testing, simple alternatives, and appeal routes make embedded choices visible.
Some problems should be reframed or left unautomated when computation would weaken fairness and human judgment.
Explore more "Quotes of The Day"
Continue exploring how writers, thinkers, artists, and advocates have described the choices that shape an everyday life.