Pick a specific automated decision made by a public agency — a benefits algorithm, a visa-streaming tool, an exam grade, a welfare-fraud predictor — and bring a judicial-review or constitutional case against it on behalf of affected people. The case names a real harmed person, attaches evidence of disparate impact or unlawful process, and asks the court to suspend the system or rewrite the rules around its use.
An advocate chooses litigation when a system is already deployed and the political route is blocked: courts can compel disclosure of how a system works, force suspension faster than legislatures move, and produce a precedent that travels to every other agency running anything similar. The strategy converts an opaque administrative practice into a public record reviewed under a regime — administrative law, equality law, data-protection law — that pre-dates the AI.
It trades off speed of impact for narrowness of remedy. A win usually rewrites that deployment, not the class of deployment; an unsympathetic plaintiff or a thin record can produce precedent the movement then has to live with. And it depends on a competent legal aid sector — in jurisdictions without one, the tactic does not transfer.
Verdict — good strategy, conditional
Strong, under named conditions. Where the strategy applies — a developed legal-aid sector, an already-deployed and opaque public-sector algorithmic system, an affected class with standing — it has produced fast, concrete wins: a deployed system suspended, a court order forcing disclosure, a precedent other jurisdictions then cite. The four effects: entries above are wins where the strategy is the proximate driver, not adjacent to it. SyRI in the Netherlands is the strongest of the four — a court-of-law precedent that the system itself was unlawful, not merely badly run.
The strategy is bad when used as the only tool. Its remedies are narrow: SyRI was replaced by a successor with similar logic; the Home Office's "interim process" was opaque about whether the next visa-decision system would be algorithmic. A movement that relied on litigation alone would win and re-win the same kind of case in series while the underlying state appetite for algorithmic decision-making continued unaltered. The strategy works best paired with empirical audit and expose (which produces the evidentiary record) and with coalition lobbying of binding regional regulation (which constrains the next deployment before it ships). Adopters who run only the litigation arm produce reversals; adopters running the trio produce regime change.
The cases above also expose a second, sharper good-vs-bad axis: target choice. Cases brought against a named already-deployed system with a documented harmed class (SyRI, visa-streaming, Ofqual, the Amsterdam Uber rulings) won decisively. Cases against a category of practice in the abstract are far harder to win and far easier to lose into a precedent the movement does not want to live with — a real risk the strategy carries that adopters routinely manage by case selection.
Ecology
This strategy is fed by empirical audit and expose — an audit is often the evidentiary record a case rests on, and an audit alone rarely forces action without a legal frame around it. It is fed by survivor-led testimony as evidence — the named plaintiff is the survivor, and survivor testimony at hearing converts an abstract harm into a justiciable one. It sits beside class-action litigation against private AI as the public-sector sibling of the same legal-disciplinary form; where the targets are private platforms providing services to public agencies, the strategies fuse (the Worker Info Exchange / Uber arc is one such fusion). It overlaps with organising the workers in the AI supply chain when the affected class is workers themselves — the litigation then doubles as union-style leverage.