AI moratorium

The AI moratorium framing is the demand for an indefinite, legally enforceable halt on the development of frontier and general-purpose AI systems — distinguished from the Pause Giant AI Experiments framing by the absence of a fixed endpoint and by its implicit or explicit legal-prohibition character. Where the pause framing asks AI laboratories (and, as a fallback, governments) to halt training for six months while safety work advances, the moratorium framing treats an open-ended legal prohibition as the primary instrument, requiring that development not resume until fundamental conditions — solving the AI alignment problem, establishing democratic governance, achieving environmental safeguards — are demonstrably met. The framing has developed across three overlapping registers since its crystallisation in March 2023: an existential-risk register rooted in AI safety and alignment research; a civil-disobedience register anchored by Bay Area direct-action organizing; and an environmental-justice and labor register emerging from resistance to AI data center expansion.

Origin: the pause-or-shut-it-down fork

The moratorium framing entered public political discourse in direct dialogue with the Future of Life Institute's open letter. On March 22, 2023, the FLI letter called for a six-month pause on training AI systems more powerful than GPT-4, with a government-imposed moratorium named as the fallback if labs did not pause voluntarily. Seven days later, Eliezer Yudkowsky — co-founder and research lead of the Machine Intelligence Research Institute (MIRI) — published "Pausing AI Developments Isn't Enough. We Need to Shut it All Down" in TIME, stating explicitly that he had refrained from signing the FLI letter because it "understates the seriousness of the situation and asks for too little to solve it."

Yudkowsky's demand was structurally different from the FLI letter on two axes. First, no time horizon: solving AI alignment "could very reasonably take at least half" of the sixty-plus years AI research had existed — no six-month clock was adequate. Second, no threshold exception: not "systems more powerful than GPT-4" but all large training runs, "with no exceptions, including for governments or militaries." The framing's working slogan — "shut it all down" — rejected the pause letter's threshold logic entirely. The enforcement mechanism Yudkowsky proposed reflected this: international tracking of GPU sales, and the willingness to destroy non-compliant compute infrastructure. The word "moratorium" completed its shift from fallback to foreground — not "if labs won't pause voluntarily, governments should institute a moratorium" but "a moratorium is the demand, with enforcement."

The moratorium.ai campaign — run by the AI Safety and Governance Fund and backed by MIRI, the Center for AI Policy, and the Center for AI Safety — operationalised the existential-risk register into a policy advocacy platform, calling on governments to establish strict national restrictions on frontier AI training, international monitoring infrastructure, and global coordination mechanisms capable of halting risky training runs.

Grassroots civil-disobedience expression: Stop AI

The moratorium framing's most consequential grassroots expression has been Stop AI — an Oakland, California-based direct-action organization founded in 2024 by Sam Kirchner and Guido Reichstadter, built around a tactical break from PauseAI. PauseAI's leadership refused to authorise illegal direct actions; Stop AI was founded specifically to pursue the forms of civil disobedience PauseAI would not sanction — chaining to doors, blocking egress, and deliberately seeking arrest to draw media attention. The two organizations share the broad demand for a halt to dangerous frontier AI development but occupy opposite ends of the tactical spectrum.

Stop AI's core demand is a "permanent enforceable ban on ASI/AGI", paired with an international ban on autonomous AI weapons and surveillance restrictions. The organization explicitly models itself on climate civil-disobedience groups, especially Just Stop Oil, citing political scientist Erica Chenoweth's research that 3.5% population engagement can achieve political change. Its direct-action campaign concentrated on San Francisco offices of frontier AI companies: in June 2024, members rallied and blocked OpenAI's entrance, resulting in three arrests; on February 22, 2025, a more elaborate blockade saw co-founders chain the front doors of OpenAI's 3rd Street headquarters shut, leading to trespassing, unlawful assembly, and obstructing-business-operations charges. Stop AI's defendants openly admitted to the blockade and framed their trials as a public platform for the moratorium demand, pursuing a necessity defense — that illegal obstruction was justified to prevent the greater harm of unchecked AGI development.

Organized advocacy and legislative expression

The moratorium framing entered legislative politics in the United States in March 2026, when Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced the Artificial Intelligence Data Center Moratorium Act — a federal moratorium on new AI data center construction tied to three conditions for lifting: AI must be proven safe and effective, protecting public health, privacy, and civil rights; economic benefits must be shared with workers rather than concentrated among technology owners; and AI development cannot increase electricity costs, harm communities, or damage the environment. The bill framed AI expansion as unregulated activity controlled by "billionaire Big Tech oligarchs" — an environmental-justice and labor register that arrived at the same legislative demand as the existential-risk camp while rooting it in present material harms rather than speculative futures.

In June 2026, the Stop Data Centers Coalition — convened by Food & Water Watch and bringing together 500+ organizations from 47 states, including Greenpeace USA, Friends of the Earth, Our Revolution, and the Center for Constitutional Rights — sent a letter to Congress calling for a national moratorium on new data center approvals until adequate regulatory protections could be enacted. Food & Water Watch organizing director Emily Wurth described this as "the large and surging national movement" beginning with residents in communities already affected by the data center buildout — a working-class coalition whose entry point was energy costs, fossil fuel pollution, and water stress, yet whose legislative demand was identical to the framing the existential-risk and civil-disobedience tracks had developed.

Why the framing has carried

The AI moratorium framing has proved durable because it fills a structural gap in the vocabulary of the AI-opposition and AI-safety movements.

The "pause" framing's operational precision — six months, systems more powerful than GPT-4, public and verifiable — is also its limit: it frames an endpoint for resumption, implies the pause period would yield sufficient safety progress, and targets a specific capability threshold the labs have already cleared. The moratorium framing rejects all three framings. It has no endpoint other than problem-resolution; it does not assume the alignment problem is solvable on a given timetable; and it targets the general trajectory of frontier AI development rather than a specific threshold. The word "moratorium" carries its own political tradition — nuclear testing moratoriums, biological-weapons moratoriums, construction moratoriums on hazardous facilities — giving the demand a legal-political form with established precedents and mechanisms that "pause" does not carry.

The framing's two-front structure — existential-risk on one flank, environmental and labor justice on the other — has created a coalition of demand-alignment without ideological alignment. Stop AI members cite Geoffrey Hinton's extinction-risk warnings; the Stop Data Centers Coalition cites rising electricity bills and contaminated aquifers. Both call for a moratorium. The structural alignment of the demand, absent any shared premise about which harm is primary, is the framing's coalitional asset and its principal mechanism for bridging AI-safety and environmental-justice communities that have historically operated in separate organizing traditions.

The framing's limit, visible across the same record, is the question of conditions. A coalition that agrees on "stop now" but not on "stop for what reason" faces structural divergence when asked what conditions would justify resumption. The existential-risk camp's answer — solve the AI alignment problem — and the environmental/labor camp's answer — establish adequate safeguards on infrastructure, workers, and communities — name conditions that may not converge in the same legislative or treaty space. The Sanders/AOC bill's three conditions for lifting the moratorium are environmental and labor conditions, not alignment or safety conditions; Stop AI's "permanent enforceable ban" carries no explicit resumption condition at all. The framing has proved capacious enough to hold this divergence while the demand is being made; whether it is capacious enough to hold it through any actual governance process remains unresolved.