2023 pause letter on giant AI experiments

The Future of Life Institute (FLI) organized an open letter asking AI laboratories to pause training systems more powerful than GPT-4 for at least six months. Dated March 22, 2023, it called for shared safety rules before further increases in capability. It was a public policy appeal, not an enacted moratorium.^[1]

What the pause would cover

The request concerned training systems more powerful than GPT-4, rather than all AI development. The pause would be public, verifiable and include all key actors; governments would impose a moratorium if voluntary coordination failed. Laboratories and independent experts would develop safety protocols, with outside auditing and oversight. The letter also proposed regulatory authorities, model and compute tracking, synthetic-media provenance, liability for harm and public safety-research funding.^[1]

Its rationale combined concerns about misinformation and jobs with possible future loss of human control. Those were the letter’s arguments for intervention, not measured consequences of the proposed pause. The hosted PDF distinguishes the March 22 letter date from its May 5 creation date and labels its roster a sample of prominent signatories, including Stuart Russell and Yoshua Bengio. FLI’s hosting and organizing role should be distinguished from their signatures.^[1]

Reasons and disagreement

In an April 7, 2023 Science Friday interview with host Ira Flatow, Russell explained his signature through the difficulty of assuring the safety of opaque systems. He distinguished immediate risks from possible loss of control over more powerful future systems. He explicitly qualified the latter as not a known risk of the then-current systems. This was a signatory’s rationale, rather than an independent test of those forecasts.^[2]

The letter also prompted disagreement about which harms should guide regulation. A March 31 response hosted by the Distributed AI Research Institute, from the listed authors of Stochastic Parrots, argued that speculation about powerful future minds distracted from present worker exploitation, data practices and concentrated corporate power. They supported provenance measures but called for training-data and architecture disclosure, accountability for builders, and participation by affected people. This was a competing policy argument, not opposition to AI regulation or evidence of a unified research consensus.^[4]

Research and policy

Concrete Problems in AI Safety describes narrower accident mechanisms involving objectives, supervision and learning. Its research agenda supplies neither an extinction-risk estimate nor a finding that six months would make advanced systems safe.^[3]

The later 2023 statement on AI risk makes a different appeal: prioritize mitigation of extinction risk. A signature on one appeal does not establish agreement with the other. The pause letter makes the proposed intervention concrete; assessing its feasibility, adoption or effectiveness requires evidence beyond the appeal itself.