Researchers warn of "agentic flooding" of public services with AI generated communications

Researchers Chris Schmitz of the Hertie School and Lewis Hammond and Alan Chan of the Centre for the Governance of AI have identified 84 cases across 11 countries in which public bodies or credible observers attribute surges in requests to public use of AI, including freedom of information requests, planning consultation responses and civil court claims. They warned that governments are likely to respond with fees and other barriers rather than by building capacity.

The paper, Characterizing Agentic Flooding of Government Services, defines agentic flooding as a surge in the volume or complexity of requests to a service that is caused by AI use and substantially strains the body's capacity, and distinguishes quantitative flooding, where more requests arrive, from qualitative flooding, where each request becomes longer or more complex.

Schmitz and colleagues recommend three near-term actions for governments:

1. Audit services against a 13-factor risk matrix covering submission effort, expected benefit of a successful request, statutory processing obligations, identity verification and spare capacity.
2. Integrate digital identity into the most exposed services, enabling per-claimant rate limits and pre-population of known data.
3. Commission internal legal reviews to establish which responses, including fees, restrictions on free-text or digital channels and automated processing, are lawful in the jurisdiction, since uncertainty is itself a barrier to preparation.

Nine in ten cases in the dataset showed qualitative flooding, 60% showed quantitative flooding and half showed both. In one German social court case, submissions ran to more than 4,000 pages. In 87% of cases the mechanism was the same: large language models generating legally plausible text at negligible cost, which people then submitted through existing channels by hand.

The researchers found no evidence yet of autonomous agents navigating government portals. Services that accept free text of unlimited length through open digital channels, such as FOI portals, consultation platforms and public comment procedures, were the most exposed.

The lead example is Australia's federal FOI regime, where the government has considered reintroducing application fees after officials pointed to a wave of AI-generated requests.

Transparency and access to information was one of 13 service domains scanned in each country, alongside public participation and regulatory complaints. The two UK cases used as illustrations were Money Claims Online and Regulation 18 and 19 Local Plan consultations, both showing rises in volume and complexity.

The dataset was built using an LLM-assisted pipeline with human review at each stage and a strict inclusion rule: the affected body or a reputable secondary source must itself have attributed the change to AI. Of roughly 2,300 candidate services, fewer than one in 20 met the test. Officials asserted AI involvement in 69% of the included cases.

Schmitz said the method supports no causal or quantitative conclusions and that the true number of affected services is likely to be higher.

Governments responded in 56% of cases, mostly with narrow or non-binding measures such as guidance on AI use. In 17% they added friction: fees, identity checks, per-claimant limits or IP blocking. In 25% they introduced AI tools of their own, for example to analyse consultation responses in bulk.

The paper notes that friction is fast to deploy and has precedent from earlier demand surges, but disproportionately deters poorer and less digitally literate users, worsens trust in government and in some jurisdictions is unlawful.

Some measures will stop working as AI capabilities improve, as CAPTCHAs already have. The authors predict that demand suppression will become the default because service redesign and digital identity integration take years to deliver.

In the UK, the statutory levers available to FOI and EIR practitioners include the cost limit under section 12 of the Freedom of Information Act 2000 and the fees regulations made under section 9, the vexatious and repeated request provisions in section 14, and the manifestly unreasonable exception at regulation 12(4)(b) of the Environmental Information Regulations 2004.

None was designed with machine-generated requests in mind, and the ICO's guidance on vexatious requests focuses on the burden, motive and value of a request rather than how it was produced.