Race After Technology

Benjamin, a Princeton sociologist, argues that discriminatory design does not require discriminatory intent. Systems trained on historical data inherit historical outcomes, and a claim to objectivity makes that inheritance harder to contest than open prejudice.

Her term, the New Jim Code, deliberately echoes Michelle Alexander's The New Jim Crow and makes the same structural point about a different mechanism.

A system that cannot be accused of prejudice is harder to challenge than a person who can.

Devmark’s reading — not a quotation from the book

She works through predictive policing, risk scores, hiring algorithms, image recognition, and the more mundane cases — soap dispensers that do not detect darker skin, name-based filtering in recruitment — where the failure is visible without any statistics.

The consistent pattern is that the harm is downstream of a design decision nobody recorded as a decision.

Benjamin's closing argument is for what she calls abolitionist tools: technology built with the explicit aim of dismantling hierarchy rather than being neutral about it, on the grounds that neutrality in an unequal system preserves the inequality.

It is an academic book written accessibly. The theoretical apparatus is present but not required to follow the argument.

Benjamin's insistence on the word design is deliberate. Calling an outcome a bias invites a technical fix — retrain the model, rebalance the dataset — while calling it a design decision locates responsibility with people who made choices and could have made others.

That distinction is the book's practical contribution. It changes the question from what went wrong with the system to who decided what the system was for.

Benjamin separates the mechanisms rather than treating bias as a single phenomenon: systems deliberately built to discriminate; systems that discriminate by inheriting an unequal world as their default; systems that expose some people to scrutiny while rendering others invisible; and systems marketed as remedies that extend control. The four require different responses, which is the point of distinguishing them.

Her most useful category covers outcomes nobody chose: a sensor calibrated on a narrow range of skin tones, a model trained on historical decisions that encodes them as ground truth, a name-matching process that inherits existing hiring patterns. No intent is required, the outcome is reproduced at scale, and the automation supplies deniability.

Nobody wrote the rule. The system learned it from the record and now applies it faster.

Devmark’s reading — not a quotation from the book

The sharpest section examines tools presented as humane alternatives, drawing a contrast between an application that redirects small sums toward bail funds and one that offers electronic monitoring as an alternative to detention. Her argument is that the second expands the reach of the carceral system while describing itself as a reduction of it.

Underneath the examples is a theoretical claim: that racial categories are themselves a sorting technology, built historically to allocate resources and risk, and that computational systems are the current instantiation rather than a new departure. It explains why she treats the word design as load-bearing.

Her constructive chapter asks what a tool built by and for the people it affects would look like, and insists that the question is about power and governance rather than about accuracy. Fairer prediction of an unjust outcome is, on her account, not an improvement, which is the strongest challenge she puts to the algorithmic-fairness literature.

It is written from within critical race studies and science and technology studies, and it assumes that vocabulary. Readers arriving from engineering will find the theoretical apparatus heavy and some examples treated briefly where they wanted the technical detail. The argument rewards the effort; the book does not meet a general reader halfway.

The reusable move is to stop asking whether a system is biased and start asking what it was built to sort, on whose behalf, and what happens to the people it sorts downward. That question applies to systems with no demographic variable anywhere in them, which is exactly where the fairness checklists stop working.