PauseAI Proposal

Notes

Primary doc from advocacy org stating its own policy proposal. High reliability for 'what PauseAI proposes' (self-description); lower for empirical claims it cites. Treated as primary_doc since it is the org's own proposal statement.

Evidence from this source (5)

• CDPauseAI asserts that 64-69% of the U.S. public supports a pause on the training… support
  weight

  0.55

  method: journalistic_report · locator: Opening paragraph

  “the U.S. public also strongly supports a pause (64%-69%).”

• CDPauseAI proposes a global, treaty-based temporary pause on the training of the… support
  weight

  0.95

  method: primary_testimony · locator: Opening / 'Getting to a Treaty' section

  “We do not expect countries or companies to risk their competitive advantage by pausing AI training runs for a long time if other countries or companies do not do the same. This is why we need a global pause.”

• CDPauseAI's proposed treaty would grant signatory countries veto power, exercisab… support
  weight

  0.95

  method: primary_testimony · locator: Treaty Measures, 'Signatory countries have additional veto powers' subsection

  “A sufficiently large coalition of opposing countries can veto any deployment of a safe superhuman AI system. This includes opposing details of a deployment such as the instructions given to a system or the values instilled in it.”

• CDPauseAI's treaty proposal centers on an International AI Safety Agency (AISA) m… support
  weight

  0.95

  method: primary_testimony · locator: Treaty Measures section

  “Set up an international AI Safety Agency (AISA). Inspired by the International Atomic Energy Agency (IAEA) ... tracking GPUs, incentivizing whistleblowers, energy monitoring, data center inspections, financial intelligence, semiconductor manufacturing facility inspections, AI developer inspections, chip location tracking and chip-based reporting.”

• CDPauseAI argues that as training costs fall (potentially to consumer-hardware le… support
  weight

  0.90

  method: primary_testimony · locator: Long term policy section

  “Limit publication of training algorithms / runtime improvements ... Limiting the publication of dangerous research has been done quite effectively in other fields such as biosecurity ... Limit capability advancements of computational resources.”