Model Audit Provisions for Automated Decisions
Model wording a regulator or ministry can draft from directly, with notes on which choices are load-bearing.
Rights that cannot be exercised in practice are commitments rather than protections. Most digital rights frameworks now in force are, on paper, reasonably strong: there is a right to explanation, a right to object, a right to human review, a right to erasure. The distance between those statements and what a person can actually obtain is where our work sits.
Consider what exercising them involves. A right to explanation is satisfied by a paragraph of generic text describing the category of factors considered. A right to object requires identifying which of several contracted processors holds the relevant record. A right to human review produces a review of whether the correct process was followed rather than of whether the outcome was right. Each step is formally compliant and practically empty, and the burden of discovering this falls on the person least equipped to carry it — typically someone already dealing with the consequence of the decision they are trying to contest.
We work on the operational distance between a stated right and an exercisable one. Whether the reasons given are actually reasons — specific enough that a person could identify an error in them. Whether an appeal reaches someone with authority to reverse. Whether the process is navigable by the person most likely to need it, tested with those people rather than assumed. Whether refusal or delay carries any consequence for the institution. Our interest is deliberately mechanical, because the mechanism is where rights are either honoured or quietly hollowed out while the text remains unchanged.
Model wording a regulator or ministry can draft from directly, with notes on which choices are load-bearing.
A controlled trial of whether removing throughput pressure restores meaningful human override in a deployed decision-support system — with the result published…
Developing an auditable standard for the review of automated decisions affecting individuals, drafted with the regulators and engineers who would have to…
A working session on the model audit provisions, with supervisory staff, engineers and caseworkers in the same room.
Most audit obligations in current legislation can be satisfied without anyone examining whether a system produces defensible decisions. This paper sets out…
Our policy paper on audit obligations is published, with model provisions a regulator can draft from directly.
Individual agreement was designed to authorise a specific use of specific information. It is now the legal basis for inferences nobody could…
Our full response to a public consultation, published as submitted. We supported three proposals, opposed one, and identified an omission we consider…