Synthetic agents in government

Disclosure and accountability formats, and what governments are doing.

Disclosure formats for AI in government exist at five grains, and none records, for one output, who or what produced each part and what the human did with it. Tool registers (UK ATRS, 153 records on 14 October 2026; a registry of AI agents under construction in Singapore; an accountable owner for each agent asked for in New South Wales and by the Australian DTA) give an owner and describe human involvement in words. Assessments (Canada's AIA, the AI Act's fundamental-rights assessment) scale duties to risk. Decision notices (France since 2017, Italian case law) say an algorithm contributed and how much. A document rule (AI Act Art. 50(4)) exempts AI-generated public-interest text only if a human holds editorial responsibility. PROV-DM models agents and delegation. The human response is the weak point. Dutch experiments find no general automation bias but selective use of advice that fits a prior view (govai-03, govai-11); oversight policies assume a capacity people may lack (govai-04); the reviewer can become the crumple zone (govai-05). Deployments are real but reported in aggregates: UK Consult (F1 0.76) and Extract (two-thirds of cases needed minor edits), Singapore, Estonia, the UAE's Regulatory Intelligence Office and Davos whitepaper, Dubai's 22 Chief AI Officers. For Mandates the test is whether the named human could decide, in time and with authority, and whether anyone able to change the decision hears back.

Sources

What it implies for the proposals

Ideas to take

A hands card shaped like a provenance record, with a human-response block.

Changes: The GRAM page can specify a card for each register of an Entry and for any document built from the log: the part produced, the agent type (person, software agent, organisation), who it acted on behalf of, its role (Alter, Adversary or none), degree and mode of contribution in the French wording, then a human-response block: who reviewed, with what authority and time, and what changed. The log can be exported in a form other tools read.

Risk: Free-text disclosure drifts (govai-09 reports wide variation in Australian statements) and a rich model invites completeness theatre. Keep fields few and fixed; a card nobody fills in is worse than a short one.

Authority and capacity in place of 'human in the loop'.

Changes: Every human response records whether the person could override, stop or escalate, and had the information and time to do so (Art. 14(4), 26(2)). The failure log gets a category, crumple zone: blame landing on a person without control. Mandates gets a standard question for any claim that a human decides.

Risk: Authority is partly self-reported and partly a legal question that a literary reading will not settle. Naming roles can expose individuals; use role titles and ask an outside reader.

Measure the human response and treat absence as 'unchecked'.

Changes: Any Trace saying officers review everything gets three follow-up Traces: edit rate, override rate, time per item. If none is published the Test result is unchecked, logged as a finding about the institution and not as a gap in the study.

Risk: A low edit rate is ambiguous: a good tool or a rubber stamp. The method must keep the ambiguity and not turn a rate into a verdict; metrics can also be gamed once published.

Reading before reply, as a small embedded design.

Changes: The protocol fixes the order: the human writes the Reading, then sees the Adversary's Test, and the log records whether the Reading changed and whether the Adversary agreed or disagreed. Over a few dozen Entries this gives a rough, honest measure of selective use inside the research team.

Risk: Too few Entries for inference, and the researcher is both subject and author; present it as a log, not a result. The Dutch findings come from vignette and simulation studies and may not transfer.

Outside-in reconstruction as the standard Test for institutional claims.

Changes: The Test register gets a template: artefacts obtained, one factor varied, result, and a list of what could not be obtained. For Mandates the artefacts are public records (ATRS, AIA results, published evaluations); a polity-level inquiry such as the Robodebt Commission is the pattern for testing act-level documents.

Risk: The first study is public material only, so the artefacts will rarely allow a model to be re-run and the method may shrink to document comparison. Released data can carry personal information.

Make a stated principle checkable: humans decide at critical stages.

Changes: Gives Mandates a distinct study for layered jurisdictions: ask each level to list the critical stages in its own documents and name the person with authority at each; the disagreement between levels is the finding. For governance-innovation practitioners it reads as an offer to operationalise a named principle, not a critique.

Risk: Political sensitivity, and the whitepaper text has not been found; a government reader may take any gap as an accusation. Work from the government's own words and agree an outside reader from inside government before starting.

A registry for the synthetic agents used in a study.

Changes: The protocol adds a short registry before any synthetic agent is used: owner, role (Alter or Adversary), permitted actions, log location. The Estonian image of the kratt, an agent that must always be given work, can name the care rule that every agent has a stated job and anything outside it goes to the failure log.

Risk: Formalism for a three-week study; the folklore framing may read as frivolous to some audiences; agent registries are announced, not evaluated.

Disclose before adoption: a hands map travels with any drafted text.

Changes: Any text an institution could adopt (draft clauses, plan text, terms of reference) circulates with a map of which parts a synthetic agent produced and who reviewed them, from first circulation. Fiction stays labelled and outside the text. This gives governance labs a practical product and ties to Art. 50(4).

Risk: Provenance of iterative drafting is hard to keep at clause level and false precision is likely. Drafters may stop using the tool openly to avoid the map.

Counter-arguments

Gaps