Pause Giant AI Experiments

Pause Giant AI Experiments is the framing of the 22 March 2023 Future of Life Institute open letter calling on "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4." The letter pairs a single concrete operational demand — a six-month, public, verifiable pause, with a government-imposed moratorium as the fallback — with a structural diagnosis ("the dangerous race to ever-larger unpredictable black-box models with emergent capabilities") and a constructive redirect toward making today's systems "more accurate, safe, interpretable, transparent, robust, aligned, trustworthy, and loyal." It is the foundational public-discourse framing of the AI-safety / pause movement: the first widely-circulated artefact that named a specific training-run threshold, a specific time horizon, and a specific addressee, and the framing through which the subsequent grassroots-organising work of PauseAI — founded by Joep Meindertsma in Utrecht two months later — translates the open-letter demand into sustained protest, petition, and treaty-advocacy work.

Origin

The letter was published on 22 March 2023 by the Future of Life Institute, a Cambridge, Massachusetts-based AI-safety-and-existential-risk think tank founded in 2014 under the presidency of Max Tegmark. Its publication came eight days after OpenAI's 14 March 2023 release of GPT-4, and the framing was built around that proximate trigger: "AI systems more powerful than GPT-4" supplies the demand's operational threshold, and the GPT-4 release supplies the moment of public urgency the letter argues from.

The open-letter genre is the framing's load-bearing form. The single-page text is built to be co-signed: a problem statement, a single concrete demand, a list of follow-on policy asks (an independent third-party oversight regime, watermarking and provenance systems, liability, robust public funding for safety research, well-resourced institutions for handling disruption), and a signatory line. The framing's signatory roster — opening with Yoshua Bengio (one of the three Turing Award-winning "godfathers of deep learning"), Stuart Russell (Berkeley, co-author of the field's standard AI textbook), Elon Musk, Steve Wozniak, and Yuval Noah Harari, growing to more than 30,000 further signatories within months of publication and standing at 31,810 on the FLI page at the time of this entry's last check — is the framing's principal mainstream-press penetration mechanism. The letter's authority sits on the named lead figures' direct involvement in the field; its reach sits on the long tail of signatories drawing in researchers, engineers, ethicists, and members of the public who were given a single artefact to put their names to.

The Future of Life Institute itself is structurally a think-tank / policy organisation rather than a grassroots-organising body, and is not represented as an Org entry in this corpus per the mission's lab / think-tank exclusion. The framing's in-corpus anchor is the grassroots-organising actor that propagates it — PauseAI — rather than its authoring institution.

Travel into the AI-safety / pause grassroots movement

The framing's most consequential downstream venue is the AI-safety / pause grassroots movement that took shape inside two months of the letter's publication. PauseAI was founded in Utrecht, Netherlands in May 2023 by Joep Meindertsma, who set aside his role as a software-company CEO to organise a sustained public response to frontier-AI training. The temporal sequence — open letter on 22 March 2023, PauseAI's first public protest at Microsoft's Brussels lobbying office in May 2023 — and the structural overlap between the open letter's six-month pause demand and PauseAI's continuing pause demand make the framing the public-discourse anchor the grassroots movement was built within. The corpus treats this relation as a temporal and structural one rather than a causal one: PauseAI's own published founding narrative credits Meindertsma's intellectual debt to Nick Bostrom's 2014 Superintelligence rather than asserting the FLI letter as the cause of the founding, and the framing's role in the public record is to have supplied the working vocabulary of the demand rather than to have produced the organising vehicle.

What the framing did supply is the operational language the movement's continuing policy proposal carries forward. PauseAI's published policy programme — an international AI safety agency on the IAEA model, an enforceable treaty pausing training runs above a defined threshold of capability or compute, liability for AI developers, and restrictions on training general AI systems on copyrighted material — is the open letter's pause demand translated into a sustained movement platform with a treaty mechanism attached. PauseAI's own learn page names the FLI open letter directly inside the public-discourse history it presents to new supporters ("Future of Life Institute started the open letter, led by Max Tegmark"), treating the framing as part of the working historical record the organising operates within.

The framing has been carried forward in the protest-line vocabulary of PauseAI's subsequent campaign work. The 13 May 2024 international day of action, staged in thirteen countries ahead of the AI Seoul Summit, ran the pause demand directly into the streets in the form the open letter had introduced; the 30 June 2025 Google DeepMind London protest, framed around Google's Seoul Summit Frontier AI Safety Commitments and the March 2025 Gemini 2.5 Pro release, treats the open letter's pause logic as the working baseline of what it means to "honour" voluntary AI-safety commitments. The framing's continued life in 2025 protest organising is one of the corpus's clearest cases of an open-letter framing remaining structurally active as a movement vocabulary years past its publication moment.

Reception and contested status

The framing's reception inside the AI-safety field was immediately fractious, and the public record of disagreement is part of what made the framing carry. Eliezer Yudkowsky argued the letter "doesn't go far enough," advocating an indefinite pause and a willingness to use international force to enforce it. Timnit Gebru and adjacent algorithmic-accountability voices contended that the letter amplified "futuristic, dystopian sci-fi scenario" framings at the expense of present-day AI harms — labour exploitation, surveillance, bias — that the corpus's worker-side and algorithmic-justice tracks anchor on. Bill Gates stated that a pause by one group doesn't solve the fundamental challenges, framing the demand as practically unworkable. Sam Altman, then OpenAI CEO and the named addressee of the letter's GPT-4 threshold, said the proposal was "missing most technical nuance," and Reid Hoffman characterised it as "virtue signalling" with negligible practical impact.

The framing's public-discourse durability has been a function of, not despite, this fractious reception. The pause-too-weak critique (Yudkowsky) and the pause-too-narrow critique (Gebru) have run on opposite flanks of the same open letter, and the framing has remained the reference point both critiques operate against. As of mid-2024 the Wikipedia article records that no pause materialised — AI training continued — but the framing's afterlife is read across the AI-safety field as the moment the pause demand entered the public political discourse in a form that could be co-signed, cited in headlines, and operationalised into a sustained movement vocabulary.

Why it has carried

Three features have made the framing durable.

First, it pairs a structural argument with a single operational demand. The structural argument — that frontier-AI training is racing past humanity's capacity to align, govern, or even understand the resulting systems — is shareable across the AI-safety field's diverse intellectual lineages (existential-risk research, AI-alignment research, governance scholarship). The operational demand — a six-month, public, verifiable pause on training systems more powerful than GPT-4 — gives that structural argument a concrete object to point to, with a specific threshold and a specific time horizon. Most adjacent AI-safety framings supply one or the other; this framing supplied both in a single artefact.

Second, the framing arrived in a form built for co-signature. The open letter as a genre — short, single-page, list of demands plus signatory line — is portable across the AI-safety field's research community, the entertainment and culture industries that signed (Wozniak, Harari), and the general public that the long tail of signatories represents. The form is also reproducible: the framing's downstream propagation has run through PauseAI's own continuing open letters and petitions, through the signatory-letter genre's wider adoption in pause-movement communications, and through the open-letter format becoming a recurring instrument for the AI-safety community's public-discourse interventions in the years after.

Third, the framing names a threshold rather than a category. "AI systems more powerful than GPT-4" is a technical operational benchmark — quantifiable, anchored on a specific public model release, capable of being updated as the field moves — rather than a definitional category that requires consensus on what "general intelligence", "AGI", or "superintelligence" means. The threshold framing has carried through the field's subsequent vocabulary evolution: PauseAI's policy proposal phrases its pause demand around "training runs above a defined threshold of capability or compute," and the threshold logic is the open letter's contribution to the field's settled organising vocabulary even where the literal six-month and GPT-4 framings have been superseded.

The framing's limit, visible across the public-record reception, is that the same threshold logic that gives the demand operational concreteness also gave the labs an off-ramp: GPT-4 became the public benchmark below which training continued without falling under the letter's demand, and successor models were framed by their developers as the legitimate next step from a baseline the open letter had named but not blocked. The framing's afterlife in the grassroots movement — PauseAI's continuing organising around a sustained, treaty-backed pause rather than a one-off six-month moratorium — is the corpus's clearest example of an open-letter framing being carried forward by an organising vehicle precisely because the original artefact's operational demand was not met.