The Digital Poorhouse

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes these systems as mirrors that reflect and amplify intention. The digital poorhouse is the proof, at documentary scale, of what that amplification produces when the intention is institutional management of the poor. A system trained on historical administrative data inherits the sedimented record of every past discriminatory decision the institution made—every investigation, every denial, every removal—and projects that record forward as a prediction, which the institution then uses to justify the same patterns of surveillance and exclusion. The mirror does not flatter. It reflects, faithfully and at scale, the old answer to the question of what we owe the poor: which is management rather than justice, surveillance rather than trust.

Eubanks’s documentation of the digital poorhouse does the work for the cycle that no purely technical account can do: it names the political choice inside the algorithm. Algorithmic governance presents itself as neutral; the digital poorhouse concept strips the presentation, exposing the value-laden decisions about whose data to train on, what to optimize for, and who bears the cost of error. The concept therefore functions, within the cycle, as a discipline for reading any automated system: ask what historical institution it descends from, ask what that institution was designed to do, and ask whether the new tool was engineered to break that design or merely to automate it.

Origin

The concept emerged from Eubanks’s fieldwork across three jurisdictions reported in Automating Inequality (2018), and it was grounded in a historical argument about the institutional ancestors of digital systems. The county poorhouse—the real nineteenth-century institution of confinement and moral assessment—gave way to scientific charity, the late-nineteenth-century movement that sent caseworkers into the homes of the poor to investigate and morally assess them before authorizing relief. Scientific charity gave way to the eugenics movement, which used the same registries and the same sorting logic to decide whose reproduction should be encouraged. Each transition automated and scaled the functions of the one before; each passed on the accumulated data of past decisions as the substrate of new ones.

Eubanks’s historical argument is that predictive models are, in the most literal sense, trained on this lineage. A model learning from decades of administrative records learns which families were investigated, which applications were denied, which neighborhoods were surveilled—patterns that reflect not the actual distribution of need or risk but the historical distribution of institutional attention, which was never neutral. The eugenic registry does not need to survive as an explicit instruction; it survives as the statistical shadow it cast across the records, which the model faithfully reconstructs and calls a score. This is what makes the digital poorhouse a claim about software architecture rather than merely a rhetorical provocation.

The concept gained wide circulation after Eubanks published a preview essay in Harper’s Magazine in January 2018, months before the book. It was quickly adopted in computer science education, policy schools, and civil rights organizations as a framework for understanding automated systems that claimed technical neutrality while producing discriminatory outcomes.

Key Ideas

Continuity, not novelty. The digital poorhouse refuses the AI moment’s favorite story about itself—that something unprecedented has happened, that a threshold has been crossed, that the systems of the past are behind us. The capability is new; the purpose is old. Automated denial, automated surveillance, automated moral sorting of the poor are not innovations but continuations, and unless we examine the old purpose, the new capability executes it faster. This continuity argument is the concept’s sharpest edge, because it makes the ethics of a new system a historical question: what institution does this tool descend from, and what was that institution for?

The bias is built in, not introduced. Standard accounts of algorithmic bias treat discriminatory outcomes as defects in an otherwise neutral tool—flaws in the training data, miscalibrated thresholds, something a better engineer could fix. Eubanks locates the bias not in some corrupting influence but in the constitution of the tool itself: in the decision to build a fraud-detection system rather than a benefits-access system, in the choice to train on data generated by an institution that historically surveilled certain families more than others, in the optimization target that rewards denial over access. The bias lives in the design, and no amount of debiasing the data reaches it.

The management-versus-eradication distinction. The digital poorhouse’s deepest function is to make poverty manageable so that the political will to end it can safely dissipate. A faster eligibility algorithm, a more precise risk model, a better-coordinated entry system—all of these make the administration of poverty more efficient without requiring the political question of why poverty exists to be answered. As AI systems grow more powerful, Eubanks warns, the danger is not that they will fail to manage poverty. It is that they will succeed—that they will make poverty so smoothly, invisibly, and efficiently managed that the moral urgency to end it quietly disappears, replaced by a dashboard showing that the system is running well.

The reciprocity test. The concept generates a diagnostic any evaluator can apply without understanding the mathematics: if this system were aimed at anyone other than poor and working people, would it survive? The vulnerability assessments, the predictive scores, the automated denial machinery—all would be politically intolerable if applied to populations with the power to refuse them. The question does not require technical expertise; it requires only the willingness to imagine standing on the other side.