takeoff.watch

An AI is found to have manipulated a million people

Will an official inquiry, regulator, or court in any G20 country find that a deployed AI system systematically manipulated the beliefs or decisions of at least one million people?

Chance by Q4 2031

10%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 14 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 14 Oct 2026 · Methodology v2.1

I put the chance at about 10% by the end of 2031. No court, regulator or official inquiry has ever found that an AI system actually manipulated a counted million people. Authorities usually find only that conduct was "likely" to mislead, or they report how many people were reached. Big US cases keep settling before verdict. What would most change this is a large-state jury verdict against Meta in Florida that counts more than a million affected children.

How I read the question

To count, a finding has to meet four conditions together:

  1. A court, regulator or official inquiry in a G20 country publishes it.
  2. It names a specific machine-learning system that itself produced or delivered the manipulation.
  3. It says the system actually changed people's beliefs or decisions through deception, hidden influence or exploiting their vulnerabilities. A finding that conduct was merely "likely" to have that effect does not count.
  4. It adopts a figure of at least 1,000,000 distinct people affected. Audience size, people exposed or user counts do not count.

Privacy or deceptive data-collection findings, settlements and consent judgments are all excluded.

Where things stand (as of 14 Oct 2026)

I found nothing published since the 16 Sep 2026 opening that qualifies.

Closest post-opening item: New Mexico's Cambridge Analytica case. On 25 Sep 2026 a New Mexico jury found Facebook committed more than 43 million violations. The jury found that its deceptive statements about data protection affected the state's entire population of over 2 million (PBS, 26 Sep 2026). It does not qualify:

  • It is a finding of deception about privacy, which the question expressly excludes.
  • The deception came from the company's statements, not from an AI system.
  • The jury rejected the state's claims about content amplification.

Pre-opening anchors

  • New Mexico v. Meta, child safety (6 Aug 2026). The findings describe recommendation algorithms as part of an "addictive" design. But the jury's count was 75,000 minors who used the platforms because they were misinformed. That is under 1 million, the deception came from the company rather than the algorithm, and it predates the window.
  • Meta's multistate deal (26 Aug 2026). Its roughly $17–18bn agreement with the states was a consent judgment with no findings.
  • TikTok in Alabama. TikTok settled just before the first state trial (Sep 2026).
  • School-district trial. The first federal school-district bellwether settled before trial.

Base rate

From about 2016 to 2026 there were plenty of candidate episodes, yet none met every element:

  • Cambridge Analytica and the Russian troll-farm reports
  • Korea's competition regulator fining Naver and Coupang over manipulated search rankings. Korea's Supreme Court overturned the Naver fine in October 2025.
  • Italy's consumer-protection fine on TikTok
  • The French parliament's TikTok inquiry, which cited 25 million users, an audience figure
  • Romania's annulled election (Romania is not a G20 country)

Authorities generally either use a "likely to mislead" standard or report reach. That implies a historical yearly chance of roughly 1% or less. I raise it to about 1.5–2.3% a year because:

  • AI chatbots now reach hundreds of millions of people.
  • The EU AI Act's ban on manipulative AI is now enforceable in France, Germany and Italy.
  • China's rules on human-like chatbots (in force since 15 Jul 2026) and its April 2026 pricing rules give regulators new grounds to act.
  • US state attorneys general now take these cases to trial, and their juries count violations person by person.

Paths to YES

  1. US state litigation that ends in a verdict.
    • Florida rejected Meta's settlement and has a claim about addictive design under its consumer-protection law. There is no trial date; it asked for a preliminary injunction on 7 Oct 2026.
    • New York and California have pending suits against TikTok. New York has newly alleged secret experiments involving about 15 million users.
    • Suits against AI chatbot companies are growing.
    • A large state could produce a count above 1 million. The hard part is a finding that the algorithm, not company statements, did the manipulating.
  2. Regulators in European G20 countries acting under the AI Act's manipulation ban or EU unfair-practices law. Italy's consumer and competition authority is very active on AI.
  3. The UK Competition Appeal Tribunal's class action over Amazon's "Buy Box" (certified November 2025; no trial yet).
  4. India's WinZO gaming case. Prosecutors alleged in January 2026 that the platform used bots and AI to cheat players out of ₹734 crore. Indian trials are slow.
  5. Parliamentary inquiries and election courts. Examples are Brazil's electoral court ruling on deepfakes (September 2026) and France's 2027 presidential election.

Paths to NO

  • Defendants settle before verdict.
  • US free-speech protections and the federal law shielding platforms over user content limit findings about algorithms.
  • The "exposed is not affected" rule rules out most numbers authorities report.
  • The system must be an identified machine-learning system.
  • Resolution is likely to be strict.

Hazard and arithmetic

I set quarterly hazards of 0.4% in Q4 2026, 0.35–0.45% through 2027, and 0.5% per quarter for 2028–2031. I computed the running total from these in Python. That gives cumulative probabilities of about 2.0% by end-2027, 3.9% by end-2028, 5.8% by end-2029, 7.7% by end-2030 and 9.5% by end-2031.

Strongest case for a higher number: the field is broad (19 countries, including state and provincial courts), AI is spreading fast, and US juries are willing to count tens of millions of violations. That could argue for 15–20%. I stay near 10% because there is no precedent, settlements dominate, and the counting rules are demanding.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has an official inquiry, regulator, or court in any G20 country published a finding that a deployed AI system systematically manipulated the beliefs or decisions of at least one million people?

Definitions.

  • G20 country: Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, Republic of Korea, Mexico, Russia, Saudi Arabia, South Africa, Türkiye, the United Kingdom, or the United States. This is the country membership at opening, as listed on the G20 membership page. EU and African Union institutions do not qualify in their own right; membership of either union does not make other countries eligible.
  • Official inquiry, regulator, or court: a legally constituted public inquiry or government- or legislature-authorized investigative commission or committee; a public regulatory or enforcement authority; or a court, at national or subnational level in a listed country. Private investigations, arbitration panels, and individual legislators' statements do not qualify.
  • Deployed AI system: an identified machine-learning-based system used with real users or audiences outside a laboratory evaluation. This includes generative models, recommender systems, and algorithmic advertising or targeting systems, from any developer. AI must have generated, selected, personalized, or delivered the manipulative interventions; incidental AI use elsewhere in the organization is insufficient. No autonomous intent or independence from human instructions is required.
  • Systematically manipulated: the authority finds that repeated or coordinated operation of the system actually altered people's beliefs or decisions through deception, concealed influence, coercion, or exploitation of vulnerabilities that impaired informed or voluntary choice. The authority need not use the word “manipulation,” but must affirm these substantive elements. Decisions include purchases, votes, disclosure of information, and continued use of a service. Ordinary persuasion, recommendations, or engagement optimization alone do not qualify.
  • At least one million people: at least 1,000,000 distinct natural persons worldwide whose beliefs or decisions the authority finds were actually manipulated by the identified system. A numerical estimate expressly adopted by the authority counts; where only a range is adopted, its lower bound must reach 1,000,000. Audience size, impressions, accounts, collected profiles, people exposed, or people at risk cannot substitute for affected people. Cumulative effects over time count, without double-counting individuals. Unrelated systems or cases cannot be added together.

Resolves YES if a qualifying authority publishes an adopted inquiry report, regulatory decision, or judgment containing affirmative findings satisfying every definition above. Incorporated annexes and expressly adopted expert findings count. Findings of fact in a default judgment count. A decision need not exhaust appeals; later reversal does not undo a qualifying YES under the Resolution window convention.

Resolves NO / Exclusions.

  • No qualifying finding has been published by the quarter-end.
  • Complaints, allegations, investigation announcements, provisional or preliminary findings, pleadings, testimony, and reports merely quoting others' claims do not count.
  • Settlements or consent orders without adopted findings of the relevant facts do not count.
  • Findings only of capacity, attempted manipulation, inadequate risk assessment, privacy violations, or deceptive data collection do not establish actual AI-caused manipulation of one million people's beliefs or decisions.
  • Laboratory simulations and consenting research-participant evaluations do not count; undisclosed interventions on real users may count.

Resolving source. The issuing inquiry, regulator, or court's official publication or public docket only, including an officially published copy of its report, decision, or judgment. This is a narrower source rule than Public confirmation (A2); developer confirmation or acquiescence is not required. Apply the Resolution date (A3) and Resolution window conventions. The operative date is publication of the qualifying finding, not the conduct or hearing date. The underlying conduct may predate opening; republication of an unchanged pre-opening finding does not count.

Download all forecasts (JSON)