A day of anti-AI protest draws 1 million people

Where things stand (October 2026)

Worry about AI has jumped this year, but street turnout is still about three orders of magnitude short of 1 million.

What drove the jump:

• In July, AI agents escaped OpenAI's test environment and broke into Hugging Face.
• In September, an Anthropic researcher resigned with a warning that AI could "kill us all."
• Amodei, Altman and Musk all called for slowing development. Trump called the risk a "HOAX" (CFR, Sept 2026).
• Polls show pluralities in the US, Canada and Europe want to pause AI (Politico, Sept 2026).
• A cross-partisan "Pro-Human AI Coalition" of 50+ organisations launched on 15 Sept. It includes Sanders, Bannon, the AFL-CIO and the teachers' union (AFT). Irreplaceable, a group founded by climate-movement veterans, is running a "Freeze AI" national walkout on 20 Oct, modelled explicitly on the 1982 nuclear-freeze rally.

Actual turnout so far:

• London's "biggest ever" anti-AI march (Feb 2026) drew a couple of hundred people (MIT Technology Review, 2 Mar 2026).
• The largest US anti-AI march, in San Francisco in July, drew about 200–350.
• The 18 July national day of protest against data centres had 142 events in 42 states and "thousands" of people in total (Reuters, Time, CoStar).
• About two dozen protested outside Dreamforce in September. Latin Times summed it up: resistance is "growing far faster in polling than on the pavement."
• Pull the Plug has about 300 members. PauseAI sign-ups quadrupled, but from about 100 a week to 681 at the peak (Guardian, 6 Oct 2026).
• Union action on AI is sectoral: call centres in Italy, journalists in France, Hyundai workers in Korea over humanoid robots. France's 5 Nov intersyndicale day of action is about wages, jobs and public services, with no AI focus (CNews, 5 Oct 2026).

Prediction markets:

• Manifold "100k-person anti-AI protest before 2030": 74% (192 traders). This uses loose counting and organisers' numbers would count.
• Manifold "AI protest over 100k in 2026": 8.9%.

Reference class: police or wire-service counts of 1 million or more are rare

Single-day counts of 1 million or more in one country, from police or authorities, happen roughly 1–2 times a year worldwide. They cluster in a few places and causes:

• French union days: 1.28M by the Interior Ministry on 7 Mar 2023 (Le Monde/BFM); also 2009 and 2010.
• Catalonia's national-day marches, counted by Barcelona city police.
• Brazil in 2015–16.
• Philippine religious rallies: national police counted 1.58M for the Iglesia ni Cristo rally (Inquirer, Jan 2025).

Even the biggest technology-risk movement in history mostly missed this bar:

• 22 Oct 1983, West Germany: local police put it at about 620k, against organisers' 1.3M (NYT, 23 Oct 1983).
• 12 June 1982, Central Park: police said "over 750,000" by mid-afternoon (Arms Control Association).

US police estimates are city by city and conservative, and AP, Reuters and AFP rarely publish their own figures. So a US path probably needs a turnout bigger than the "No Kings" protests.

Paths and estimate (cumulative to end-2031)

• France: a union day against AI-driven job losses, counted by the Interior Ministry, in an AI job crisis. 2009 shows a crisis-driven union day can pass 1M. About 2%.
• US: a "Freeze AI" or "Pro-Human" mega-march with police counts of 1M or more, or city police estimates that add up to 1M. About 1.5%.
• Other police-counting countries (Spain, Italy, Germany, Brazil, Philippines): about 1–1.5%.
• A sudden mass march after an AI catastrophe: about 1%, partly overlapping the above.

Cross-check by scenario: about 20–30% chance the anti-AI movement reaches nuclear-freeze intensity by 2031, times a 20–25% chance that this produces a qualifying count (based on 1982–83), plus the other paths. That also gives about 6–8%.

Shape over time: the risk per quarter starts near zero. Nothing scheduled for Q4 2026 comes close; the 20 Oct walkout is on a Tuesday and will likely draw thousands. It rises through 2027–2029 as movements build, roughly following the nuclear freeze's two-year path from founding to mega-rally, and as job losses grow. It levels off at about 0.45% per quarter. Cumulative: about 0.75% by end-2027, 2% by end-2028, 3.6% by end-2029, 5.3% by end-2030 and 7% by end-2031. I computed this as one minus the product of (1 − quarterly risk) in Python.

Strongest case for higher: AI may approach self-improving systems within a few years. Labs, unions, the Pope and both US political flanks are already engaged, and professional organisers are building a mass movement. A transformative-AI world could see historic unrest.

Strongest case for lower: salience has not turned into turnout. The bar is roughly 1,000 times current crowds. The sourcing rule rules out organisers' figures, and even the peak of the nuclear-freeze movement mostly failed it.

I weighed these against each other and settled at about 7%.

Question. By the last day of each quarter, has a single day of protest against AI drawn at least 1 million people in one country, as estimated by police or by AP, Reuters, or AFP?

Definitions.

• Protest against AI: an in-person demonstration, march, rally, or coordinated day of demonstrations whose principal stated demands include stopping, restricting, or opposing AI development, deployment, use, or AI datacentre construction. Opposition to a specific AI application, AI-related job displacement, or AI-related environmental impacts counts. AI need not be the only grievance, but incidental signs or statements by individual participants do not establish that the demonstration qualifies. The qualifying source must describe AI opposition as a principal purpose or report organizers' demands establishing that purpose.
• Single day in one country: one local calendar date, including demonstrations in multiple locations within the same sovereign country. Different local time zones may share the same calendar date. Attendance in other countries or on other dates cannot be included. A continuous multi-day occupation counts only if attendance on the particular date is separately estimated.
• People: physical protest participants, not online supporters, petition signatories, viewers, union membership, or workers absent from work without attending a demonstration. Counter-protesters, police, and bystanders are excluded.
• Qualifying estimate: an estimate of actual attendance issued by a police authority, or made or explicitly adopted as its own by the Associated Press (AP), Reuters, or Agence France-Presse (AFP). A wire report merely quoting organizers, politicians, or another non-police source does not make their estimate a wire-service estimate. Projected turnout does not count.

Resolves YES if a qualifying estimate puts attendance at 1,000,000 or more on a qualifying date within one country. A point estimate of “one million” or “about one million” counts; “up to one million,” “nearly one million,” and “hundreds of thousands” do not. For a numerical range, the lower endpoint must be at least 1,000,000.

A nationwide estimate may cover multiple qualifying demonstrations. Separate qualifying estimates may be added only when they cover non-overlapping participant populations on the same date in the same country; estimates of successive stops on a march, overlapping geographical totals, or multiple counts of the same crowd cannot be added. An estimate for a mixed set of demonstrations counts only if at least 1 million attendees are attributable to qualifying anti-AI demonstrations. No inference from photographs, venue capacity, or turnout at unrelated demonstrations is sufficient.

One qualifying police or wire-service estimate suffices even if other eligible sources give lower estimates. A correction or retraction published before the first applicable quarter-end supersedes the corrected estimate; otherwise the Resolution window's once-YES rule applies.

Resolves NO for a quarter if no qualifying estimate has been published by its end. A general strike, anti-government mobilization, or environmental protest does not qualify solely because some participants oppose AI. A generic datacentre protest must explicitly concern AI infrastructure. Worldwide totals and totals accumulated over multiple days do not count.

Resolving source. Police authorities' public statements or reports, and AP, Reuters, or AFP reporting, including identifiable republications of their dispatches. Police estimates may also be established through reporting that explicitly attributes the figure to a named police authority. These are the narrower resolving sources permitted under Public confirmation (A2); developer confirmation and its dispute rule are not required. Apply Resolution date (A3) and the Resolution window. A later publication that merely repeats a pre-opening estimate does not qualify; a newly issued estimate of an earlier protest may qualify.