Thematic report · AI campaign regulation

The regulators all built the same thing, and it wasn’t the thing that was broken

AI campaign regulation across four democracies: what is in force, what it costs, and the one intervention the evidence supports.

Verified as of 20 September 2026 · Reviewed 4 November 2026

Correction, 20 September 2026. This report originally stated that Brazil’s Superior Electoral Court had not resolved the scope of its synthetic-media rule. The court fixed a binding definition on 1 September 2026, before this report’s verification date. The Brazil section and the comparison table have been updated, and the state count revised from a 29–31 range to 30. The report’s central argument is unchanged and is discussed against the new decision in the Brazil section.

Correction, 3 November 2026. The Hawaii ruling was dated 24 January 2026 in earlier versions of this page; the court ruled on 30 January 2026.

Correction, 4 November 2026. Earlier versions said Israel’s labelling rule was a committee directive and that no Knesset law had passed. The duty is in a temporary amendment to the Elections (Propaganda Methods) Law; the committee’s rules of 26 July set how labels must look. The Israel section and the comparison table are updated. The central argument is unchanged.

BRAZIL

4 Oct

Runoff 25 October

ISRAEL

27 Oct

Knesset election

UNITED STATES

3 Nov

Midterms

All three elections have now been held.

What is in force

Brazil

TSE Resolution 23.732/2024, in force since February 2024. All AI-generated synthetic campaign content must be labelled “explicitly, prominently, and accessibly.” Deepfakes are banned outright, whether the depiction is negative or positive. The penalty is cassation: loss of candidacy or of office already held. On 1 September 2026 the TSE fixed a binding definition of a deepfake by 5–2: synthetic content carrying a degree of realism sufficient to confuse the viewer, reproducing the image, voice or expression of a real, deceased or fictional person, where the content constitutes electoral propaganda. Stylised avatars, drawings, caricatures without verisimilitude, technical audio and image adjustment, vignettes and logos fall outside it; memes remain permitted.

Israel

The labelling duty is in statute: a temporary amendment to the Elections (Propaganda Methods) Law, section 2A2, requires election material made or substantially altered with AI to be labelled. On 26 July 2026 the Central Elections Committee, chaired by Justice Noam Sohlberg, published rules setting how the label must appear. Enforcement is real: an AI monitoring task force was established in January 2026, an order was issued against Likud in February 2026, and against Yashar / Gadi Eisenkot on 6 August 2026.

United States

Thirty states have election deepfake laws in effect, per Public Citizen’s tracker as of 8 September 2026, with no federal baseline. Counts of 29 and 31 circulated earlier in the year; the difference is one of criteria rather than fact: whether a struck-down law still counts, and whether a state whose specific statute fell still counts on the strength of surviving provisions. Hawaii is excluded here because Act 191 was struck on 30 January 2026 by Judge Shanlyn Park. California is included despite losing AB 2839 (October 2024) and AB 2655 (August 2025), because other provisions survive. Montana is included while its law remains in litigation. In September a federal judge found Montana’s law likely unconstitutional, because it treats identically deceptive content differently depending on whether it attacks or promotes a candidate; relief was limited to the plaintiffs and the law remains in force for everyone else.

South Korea

Article 82-8 of the Public Official Election Act (2023) prohibits virtual sounds, images and video “difficult to distinguish from reality” for the 90 days before an election, with penalties of up to seven years’ imprisonment or KRW 10–50 million. The AI Basic Act (January 2025) adds a general labelling duty with an artistic-works exception. The National Election Commission identified 129 violating deepfakes in 19 days across January–February 2024.

The shared flaw

“A mandatory disclaimer for parody or satire would kill the joke”, and the law’s vague standards risked discriminatory enforcement turning on “the value judgments and biases of the enforcement agency.”

Judge Shanlyn Park, striking Hawaii’s Act 191 · 30 January 2026

The American strikedowns are not checkpoint failures. The courts did not object to the existence of a gate; they objected to the impossibility of saying, in advance, what passes through it. Each law fell on classification.

Brazil has the most demanding instrument of the four and spent eighteen months without a settled answer to the same question. Its ministers divided on 18 August 2026 over what the rule covers. Fourteen days later, on 1 September, the court supplied a criterion: realism sufficient to confuse the viewer, with an explicit carve-out for caricature, technical editing and memes. In the same session it rejected, 4–3, a challenge to an AI-generated video of Jair Bolsonaro endorsing his son’s candidacy at a party convention, holding that a closed party event is not automatically electoral propaganda. Two ministers dissented.

That is the closest any of the four comes to drawing the boundary, and it arrived eighteen months after the rule it interprets and five weeks before the first round. It is also incomplete on its own terms: the decision publishes no measure of realism, and case-by-case determination falls to electoral judges without a standard. The court named the category. It did not supply the test.

Israel answers the classification question by assertion: content that has been materially altered must be labelled. The standard has not been tested in court.

Comparison

Checkpoint, classification and durability by jurisdiction
JurisdictionCheckpointClassificationDurability
BrazilStrong on paperCriterion fixed 1 Sept; no published metricJudicial rule, stable form
IsraelActive, enforcingAsserted, untestedTemporary statute and committee rules, cycle-bound
United StatesFragmented, partly struckFailed judicial review ×3Statutory where it survives
South KoreaCriminal, enforcedBounded by 90-day windowStatutory, persistent

Five costs

  1. 01

    The failed rule is worse than no rule.

    A statute struck down after a campaign has begun leaves candidates who complied disadvantaged against those who waited. The interval between enactment and invalidation is the period in which the rule shapes behaviour without ever being lawful.

  2. 02

    Enforcement asymmetry costs you even when the regulator is right.

    Orders land on the actors that are visible, documented and reachable. Anonymous and offshore distribution absorbs the same content at no cost, so the burden falls hardest on the parties already operating in the open.

  3. 03

    Severity plus uncertainty chills more than severity alone.

    Cassation or a custodial sentence attached to a standard nobody can apply in advance suppresses lawful speech the rule was never aimed at. Campaigns respond to the range of possible readings, not the intended one.

  4. 04

    Detection displaces rather than removes.

    Labelling and takedown move synthetic material into closed channels: private groups, messaging apps, unattributed accounts. The content survives the checkpoint; only the ability to observe it is lost.

  5. 05

    A rule can have no existence beyond the crisis that produced it.

    An instrument issued by directive for one electoral cycle expires with that cycle. What looks like a regulatory regime may be a single administrative act with no institutional memory behind it.

What the evidence supports

Viorela Dan’s 2026 study in The International Journal of Press/Politics (n=2,085) tested what actually undoes the damage of a political deepfake. Journalistic fact-checks substantially reduced it: attitudes and voting intentions mostly returned to control-group values. The effective intervention was reporting, not a gate.

The Brennan Center’s assessment (Goldstein and Lohn, 23 January 2024) points the same way by omission: it does not recommend deepfake laws at all, concentrating instead on detection capacity, election-official readiness and rapid public correction.

The finding

AI campaign regulation substitutes checkpoint action for sensing capacity, and the substitution is heaviest exactly where sensing capacity is weakest.

The PDF edition is being updated to reflect the 20 September correction and will return here once reissued.

Verified as of 20 September 2026 · Reviewed 4 November 2026

Sources

Brazil

  • Tribunal Superior Eleitoral, Resolution 23.732/2024 (February 2024).
  • TSE plenary session, 18 August 2026: ministers divided on the scope of the resolution.
  • TSE decision of 1 September 2026: deepfake tese fixed 5–2; challenge to the PL convention video rejected 4–3.

Israel

  • Elections (Propaganda Methods) Law, section 2A2 (temporary amendment, 2026); Central Elections Committee rules on disclosure, published 26 July 2026.
  • Labelling proposal presented 22 June 2026, later enacted.
  • CEC AI monitoring task force, January 2026; order against Likud, February 2026; order in the Yashar / Gadi Eisenkot matter, 6 August 2026.

United States

  • Public Citizen, Tracker: State Legislation on Deepfakes in Elections, as of 8 September 2026: thirty states.
  • California AB 2839, enjoined October 2024; California AB 2655, struck August 2025.
  • Hawaii Act 191, struck 30 January 2026, Judge Shanlyn Park.
  • Montana election deepfake statute, in litigation.

South Korea

  • Public Official Election Act, Article 82-8 (2023).
  • AI Basic Act (January 2025), general labelling duty with artistic-works exception.
  • National Election Commission enforcement data, January–February 2024 (129 violations in 19 days).

Evidence

  • Viorela Dan (2026), The International Journal of Press/Politics, n=2,085.
  • Brennan Center for Justice, Goldstein and Lohn, 23 January 2024.

Stage 0 disclosure

Stage 0 assessment: structured analytical opinion, not a certified audit. Findings rest on named public sources and are open to challenge on the record.

This is a thematic report, not a country diagnostic. It assesses one governance function across four jurisdictions and does not revise any country’s assessment. The Israel material inherits the July 2026 diagnostic’s scope, which excludes Gaza and the West Bank.

← All reports