Algorithmic welfare

An algorithmic welfare system, as I use the term, administers access to entitlements through three jointly operative features: automated decisioning, in which eligibility or authentication is resolved by the system rather than an official’s situated judgement; criterion-enforcement without discretion, in which the operative criterion, such as a biometric match or a database flag, executes itself at the point of application; and integration across databases, in which entitlement depends on records held and cross-verified elsewhere.

The definition fixes the scope by architecture rather than technique, and is deliberately agnostic about machine learning. The authentication layer of Aadhaar, my paradigm case, is biometric matching and database linkage rather than learning in the contemporary sense, but it has all three features.

Does the algorithmic character do any normative work, or are these the old pathologies of bureaucracy? I think it does, at a precise point. In a discretionary bureaucracy the criterion is evidence an official weighs, and the official can recognise a manifestly entitled person across its failure: the clerk who has known a widow for a decade can honour her pension claim though her papers are lost. A biometrically gated system removes that standpoint. The fingerprint reads or it does not, and manifest entitlement is not among the things the architecture can read. Lipsky showed that front-line discretion is where policy is made in practice; Bovens and Zouridis traced its migration into software. My claim is that, for these systems, the same enforcement also produces the subject’s operative standing, the condition I call constitutive dependence.

Discretion is also the medium of arbitrariness and prejudice, so the point is structural, not sentimental. Scale and speed make the foreclosure consequential without constituting it: a small, slow but rigidly architectural system would exhibit it, a vast, fast but discretion-preserving one would not. Systems such as the Dutch SyRI and Australia’s Robodebt, which acted on people whose standing they presupposed, belong to a neighbouring family, and whether any given system is a case of constitutive dependence needs its own analysis. What such systems owe is set out in the Constitutive Obligation Principle.