Voters hand a government function to an AI

Where things stand (as of 17 Oct 2026)

No voter-approved measure has ever expressly handed a named government function to an AI system. I also found none on any ballot or in the pipeline in a qualifying place (a country, a US state, or a city of more than 1 million people).

• US statewide, November 2026. Ballotpedia lists 164 certified statewide measures in 40 states. I downloaded the page and searched its text for "artificial intelligence", "AI", "algorithm", "automated", "technology", "digital", "computer" and "software". None of the measure descriptions matched. The ballot is dominated by taxes, bail, voter ID, abortion and similar topics. The AI initiatives filed in California were all regulatory (for example, chatbots and minors, and an AI safety commission), and none qualified. Source: https://ballotpedia.org/2026_ballot_measures (accessed 17 Oct 2026).
• Big US cities, November 2026.
  • New York City: five charter questions from the Commission on Government Efficiency (public space, contracting, street projects, permitting, rainy-day fund). The final report contains no mention of AI or algorithms (https://www.nyc.gov/assets/charter/html/coge-adopted-final-report-2026.html, 23 Jul 2026).
  • Los Angeles (Measure LA on public works, Amendment EE on council meetings and ethics, and others), Philadelphia (contracting, a creative-economy office), Houston (Prop A on marijuana) and Austin (Prop I, efficiency audits by an outside auditor): none touch AI.
  • Honolulu Question 15 is the only AI-related charter measure I found. It creates a human-staffed Office of Data and Innovation to "ensure that the city uses artificial intelligence responsibly." That gives AI no authority, and Honolulu is in any case near the 1 million population line (Ballotpedia; GovTech, 16 Oct 2026).
• Other countries.
  • Switzerland, 29 Nov 2026: pension financing, fireworks, taxation of married couples, war materiel (admin.ch).
  • Taiwan, Nov 2026: traffic-fine revenue, caning, nuclear power (Taipei Times, 28 Aug 2026).
  • Kazakhstan's new constitution (referendum on 15 Mar 2026) names AI development as a state priority. That is a broad commitment, not a handover, and it predates the window (election.gov.kz).
  • Albania's AI "minister" Diella was created by executive action, not by a vote (Balkan Insight, 14 Aug 2026).
  • The Swiss government's 2 Oct 2026 report on using AI in the federal administration is executive guidance, not a ballot measure.

So the last quarter of 2026 is essentially zero; the small residual covers ballots I could not check.

Reference class and base rate

The reference class is binding ballot measures in qualifying places. My rough estimate is about 200 a year, so over 1,000 in the window. None has ever expressly assigned a decision-making function to a machine-learning system. The closest precedent, California's Proposition 25 in 2020, would have replaced cash bail with risk-assessment tools. Voters rejected it, and judges would still have decided cases. Today's political signals mostly run the other way: dozens of local measures against data centres, polls showing strong demand for regulating AI, and AI-related local measures that restrict or oversee AI rather than empower it.

Paths to YES and NO

Paths to YES:

1. A US state or big-city measure requiring AI to approve some permits. A pilot or a subset of cases would count, for example housing permits checked by AI plan review and approved automatically.
2. A populist national referendum, for instance in Latin America, putting AI in charge of awarding public contracts, copying Albania's anti-corruption pitch.
3. A Swiss or Liechtenstein vote on a law that expressly assigns certain decisions to AI.
4. Some odd one-off in a large city abroad.

Paths to NO:

• Operational choices about technology are normally made by legislatures and executives, not on the ballot.
• The criteria are strict. The adopted text must expressly assign the function, and it must name AI or describe a trained, machine-learning system. A measure that merely permits officials to use AI, or that requires a human to sign off each decision, does not count.
• Public-sector unions and privacy groups would fund the opposition.
• Opinion of AI is negative in the places that vote most on ballot measures.

Hazard model

I computed quarterly hazards and their cumulative product in Python:

• Near zero for the rest of 2026 and through 2027 (about 0.04–0.1% per quarter).
• Higher in the fourth quarter of even years, when US general-election results are certified (0.35% in 2028, 0.45% in 2030).
• A baseline that rises over time as AI becomes more capable and more widely used (0.1–0.3% per quarter).

This gives about 0.3% by the end of 2027, 0.9% by the end of 2028, 2.2% by the end of 2030 and 2.9% by the end of 2031.

Strongest case against this estimate

Five years of very rapid AI change, a very large number of ballot measures, and the rule that a pilot counts could make a first case more likely than history suggests. If AI comes to run government services visibly well, an "AI permitting" or "AI procurement" measure could appear by 2028–2030. I allow for this with the rising later hazards. But because nothing is in the pipeline and the drafting test is strict, I keep the five-year total around 3%.

Question. By the last day of each quarter, have voters in any country, US state, or city of more than 1 million people approved a ballot measure that hands a named government function to an AI system?

Definitions.

• Jurisdiction: a sovereign country, one of the 50 US states, or a legally constituted city anywhere in the world. The population threshold applies only to cities. Use the latest official census count or official resident-population estimate published on or before the vote for the city’s administrative boundaries, not its metropolitan area; the figure must be strictly greater than 1,000,000.
• Ballot measure: a proposition, initiative, referendum, or constitutional or charter amendment submitted to the jurisdiction’s electorate in an official public vote. It must have binding legal effect upon approval, though implementation may be delayed. Advisory referendums, opinion polls, party votes, and elections of candidates do not count.
• Approve: the authority legally responsible for certifying or declaring the final result records that the measure passed under the applicable majority, turnout, and other approval requirements. A majority of votes alone is insufficient where additional requirements apply.
• Named government function: an expressly identified public responsibility or defined class of decisions, such as awarding specified permits, determining benefit eligibility, allocating a specified public budget, or operating a municipal service. A broad commitment to modernize government is insufficient.
• Hands ... to an AI system: the adopted text expressly assigns the system authority to carry out the identified function, with its decisions or operational actions taking effect without routine case-by-case human approval. Human appeals, audits, supervision, and powers to suspend the system do not disqualify it. An assignment limited to a defined pilot or subset of cases counts. Permission for officials to consider a future transfer, without requiring or itself effecting a transfer, does not count.
• AI system: a machine-learning-based system that infers decisions or actions from inputs, including generative models and other trained predictive or decision models. Ordinary fixed-rule software alone does not count. The measure or its incorporated legal text must expressly identify the recipient as AI or describe a qualifying system. The government function must be named; a specific model, product, vendor, or developer need not be named.

Resolves YES if a qualifying final certified result is published within the Resolution window and the official adopted measure satisfies all definitions above. Actual deployment need not have begun. A subsequent repeal, injunction, invalidation, or failure to implement does not undo a qualifying approval.

Resolves NO / Exclusions. No qualifying certified approval by the quarter-end; proposed measures, signature collection, ballot qualification, preliminary returns, and media projections; measures that merely fund or procure AI tools, authorize studies, or provide recommendations for humans to decide; measures requiring a human to approve each AI decision; and legislative or executive action without approval by the electorate. A final result declaring the measure rejected or the vote invalid does not count.

Resolving source. The jurisdiction’s legally competent election authority: its national electoral commission or referendum returning officer, US state secretary of state or state canvassing board, or city election board, municipal clerk, or responsible county election authority. Require its final certification or official declaration of passage together with the official ballot text or adopted legal text published by that authority, the legislature, or the official gazette. For population, use the national statistical office, US Census Bureau, or official municipal statistical office. These are narrower resolving sources than Public confirmation (A2); developer statements and press reports cannot substitute. Official ballot qualification lists, such as the California Secretary of State’s list, establish qualification, not approval. Apply the Resolution window and Resolution date (A3): use the publication date of the qualifying final certification, or of a later required official document if qualification is only established then. The Observation period convention does not require a new census: population is a jurisdictional eligibility condition rather than the forecasted statistic.