Jack Dorsey has published an extended case for keeping advanced artificial-intelligence systems open to examination, modification and reuse, while allowing temporary restrictions where independently reviewable evidence shows a release would create catastrophic risk. His statement presents openness as a way to broaden technical scrutiny and competition rather than as an argument that every private model must be published.
Dorsey says no company owns the next technological frontier and argues that open alternatives should be able to compete with closed systems. He calls for researchers to have enough access to reproduce evaluations, identify failures, challenge safeguards and develop fixes without first persuading the company that built a model. In his formulation, releasing model weights can help, but reproducible code, documented limitations and published evaluations provide a fuller basis for outside review.
The statement engages directly with proposals to slow frontier development. Dorsey supports independent evaluations and checks aimed at dangerous capabilities, but opposes industry-wide limits designed principally by incumbent laboratories. Rules calibrated to the resources of today’s largest companies, he argues, could create barriers that smaller competitors and independent investigators cannot cross. He similarly rejects a system in which the United States and China effectively determine how much AI capability the rest of the world may develop.
His position is not that release should be unconditional. Dorsey acknowledges that wider access can support dangerous work and that released weights cannot reliably be recalled. He says a temporary hold may be justified by a credible warning of catastrophic harm, provided the reasons are public, the evidence can be independently reviewed, an appeal is possible and the decision is reconsidered on a schedule. Continued withholding, in his view, requires continued justification.
Before restricting publication, Dorsey wants decision-makers to test narrower controls. His examples include patching software vulnerabilities, revoking credentials, limiting an agent’s permissions and stopping a specific unsafe experiment. Access to a model, he notes, need not grant authority to trade, operate equipment or conduct laboratory work. Those operational boundaries would themselves require independent testing before they could be trusted as safeguards.
Dorsey also calls for sustained public funding of shared computing capacity, testing tools and independent research groups. He says those groups should control their investigations and be free to publish unfavorable conclusions without a company or government veto, while handling genuinely sensitive vulnerabilities through responsible disclosure.
The proposal is an intervention in a live policy dispute, not evidence that open releases are always safer. Its central test is proportionality: restrictions should target demonstrated catastrophic dangers that narrower defenses cannot adequately reduce, and should account for the research and defensive work that closed access may prevent. Dorsey accepts that a justified restriction could slow progress, but argues that delay itself should not become the objective.
