Paper maps six deployed AI systems that enable authoritarianism

The paper is a taxonomy exercise, and that is what makes it useful. Three policy researchers, Jeba Sania, Marta Ziosi, and Fazl Barez, took six real deployed AI systems from very different political contexts and put them through a common frame, and Tech Policy Press walks through the results.

The six systems are worth naming, because they are not all what you would expect. FlockSafety's neighborhood license plate readers in the US, live facial recognition vans used by 13 local police departments in the UK, Amsterdam's SlimmeCheck welfare fraud detector, the Israeli Defense Force Unit 8200's Lavender militant identification system, the Xinjiang Public Security Bureau's IJOP, and Moscow's Sfera biometric payment system on the metro. A neighborhood ALPR sits in the same table as a war-zone targeting system, and that juxtaposition is the argument.

Two of the descriptive findings stood out. First, a 'preceding security crisis' was, per the researchers, influential in the introduction of four of the systems, which lines up with a long-standing pattern in surveillance policy. Second, 'Four of the six systems were initially developed by private companies or through public-private partnerships,' which puts procurement, not just legislation, at the center of the accountability story.

The paper's normative move is the part vendors and regulators should read carefully. The authors argue safeguards 'must be embedded in the technical systems, not reliant on operational restrictions,' and single out the testing stage as 'a promising point of intervention.' They also make a transparency observation that is quietly damning, that the researchers had more access to the technical details of IJOP in Xinjiang than to FlockSafety's systems in the US.

The honest caveat, one the authors themselves make, is that six systems is a small sample and the conclusions are 'somewhat tentative due to gaps in available information.' What the reporting does not give you is any indication that operators have been forced to change course, or that any of the six have been meaningfully rolled back. But the framework itself, the six characteristics and the six feature clusters, is the kind of scaffolding that legislators drafting AI accountability rules and auditors pricing this work have been asking for.