(Live working draft. Last updated on September 11th, 2026.
Under construction with prospective partners)

One-Pager

Where: To-be-determined location in Washington DC.

When: To-be-determined dates in Q4 2026.

Aims: Foster convergence among a critical mass of "humans first" US AI policy powerbrokers on the activation of an ongoing process to create a new, cohesive US executive-branch AI policy. Such a policy will be centered on an extraordinarily bold, timely and well-designed US-China-led treaty-making process that can reliably and durably:

• Prevent the immense risks of permanent loss of human control over AI, durable global or national concentration of unaccountable power and wealth, and other catastrophic AI harms, including the enormous risks of an improper treaty or treaty-making process.

• Foster a cautious realization AI's astounding potential for Americans, all humans, and other sentient beings.

Why now: Four months ago a US-China AI treaty was a fringe proposition. Since then, the leading US labs have requested an international effort to pace the frontier, President Xi has called for measures to forestall loss of control, the White House has suspended top models on safety grounds, and US-China safety dialogues have formally begun. The US pro-human movement has won the domestic argument for keeping humans in charge. What is missing is the lane that argument now requires: a process by which the US actors who decide such things converge on whether and how to pursue a treaty.

Who & What: About 20–30 delegates (senior staff or advisors) of the executive-branch officials, AI labs, and political and media leaders who most shape US AI policy — plus a few relevant experts, meeting in DC in closed-door sessions every three weeks, supported by professional facilitators, under the Chatham House Rule. This is the missing lane: a one-sided Track 1.5, in which the officials and non-officials of one party align positions before engaging the other side — as the US did in the Acheson–Lilienthal consultations of early 1946, months before the Baruch Plan reached the UN. Given the scope of ideas discussed, this kind of dialogue can only happen at the delegate level and not directly among powerbrokers.

Expected outcomes: A shared, documented map of where those powerbrokers genuinely converge and diverge on treaty-making design and safeguards; short-term plans to scale up engagement toward a desired treaty-making process. No joint statement, declaration, or signatory list.

Conveners: A steering board of US AI treaty-advocacy organizations, now forming, holding full decision rights over the program, invitations, and outputs, with the Coalition for a Baruch Plan for AI as secretariat and a Washington institutional co-host. CBPAI's own advocacy positions are set out in its Deal of the Century initiative; they are not the positions of the Roundtables, their conveners, co-hosts, or participants. The Roundtables are the diplomatic complement to the bipartisan Pro-Human Movement: they take up the question that movement has left open — what stops the race across borders — and serve the Pro-Human AI Declaration's most demanding commitment, no superintelligence absent scientific consensus and public buy-in, by examining the only instrument that can make it durable.

Humans-First & AI Consciousness: The roundtables will include an intense open dialogue of the prospect of the creation of immense happiness or suffering following the possible future emergence of digital beings or human mind uploads with certain levels of consciousness. The August 2026 cover and special issue of the Economist and the recent post by CAIS director Dan Hendrycks are just some signs of the fact that estimates by the controller of lading AI labs about the likelihood of such prospects (along safeguards of a treaty-making and the prospect of human extinctions) are key to their decisive support of a treaty-making process.

Our Key AI Challenges

Due to the volume of AI news and the scarcity of expertise among commentators, the signal is drowned out by the noise. Key news and challenges are overshadowed in public discourse by lesser ones. Perhaps the main challenge is: 

Given plausible timelines to immense safety and power concentration risks of AI, should the US Administration engage in a properly designed, US-China-led, and multi-stakeholder treaty-making process of radically unprecedented bandwidth, scope, urgency and effectiveness to have a fair chance at tackling those risks and realize, in the process, AI's astounding potential for humanity? In other words, does the current US-China AI dialogue need to be intensified by two orders of magnitude into a "global Apollo Program for an AI treaty," co-led by the US and China? 

If so, further key questions or sub-challenges arise:

• What does the evidence on recursive self-improvement and loss of control now say, and what does it imply for timelines and for the minimum a treaty must do?

• Can loss of control be prevented by deterrence or domestic regulation alone — or only by an agreement with verification?

• How can a consensus emerge among a critical mass of key formal and informal US AI policy powerbrokers — both conservative and techno-optimist humanists — that can successfully, decisively, and properly spur and support the US Administration in such a process? 

• What features of such a treaty-making process that will sufficiently reduce the enormous risks of an improper treaty-making and treaty? And how can those features best be conceived, agreed upon and entrenched early-on? 

• Which treaty-making models and methods, in what sequence, and with whose participation to maximize competence, accountability, and fairness?

• Does an effective treaty need citizen co-ownership of large AI firms and other accountability measures — or are these separable?

• How can a treaty make US-China economic competition, via Pax Silica and WAICO, more constructive, benefiting everyone?

• What are the treaty enforcement mechanisms and diplomatic-communication infrastructure needed, and how are they built to be sufficiently mutually trusted and on time?

• What does the 1946 Baruch Plan for nuclear technology — the only comparable attempt, presented by Truman on the day Donald Trump was born — teach? Perhaps, if the September 1945 attempt by Secretary of War Stimson suggestion to initially frame it with the Soviets rather than western allies would have saved it?

Our AI Predicament: a Strategic Outlook 

Since June, we have witnessed an astounding acceleration in global concern about AI safety and the risks of concentration of power. This shift spans scientists, voters, US lab leaders, and even Presidents Xi and Trump, sparking calls for bold new international institutions to tackle these challenges head-on.

How the White House Spurred an AI Policy Snowballing

After 20 months of a "hands-off" AI policy that consolidated US dominance while avoiding major harms, the Trump Administration has pivoted radically in response to mounting evidence of risk and widespread concerns. Key actions include a June 2nd Executive Order on safety pre-testing, public acknowledgment of AI safety dialogues with China, and sudden safety suspensions of top US AI models.

Meanwhile, over a few weeks in June, a previously unimaginable, wide bipartisan consensus emerged in favor of extensive citizen co-ownership of large AI firms. Pioneered by Rosenstein, by Berggruen, and others, it was proposed in terms of 50% co-ownership by Bannon on May 29th (accruing to individual citizens) and, two days later, by Sanders (accruing to a sovereign fund) — and then supported in principle by Trump, Vance, Altman, and Amodei.

Meanwhile, US citizens have long been extremely concerned. While 63% of US voters believe it's likely that humanity will lose control of AI, 53% believe it's likely that AI will destroy humanity, and 77% call for a strong global AI treaty.

Although President Xi advocated for a UN-led framework, Trump maintained that AI governance must primarily be established by the two leading AI powers. OSTP Director Kratsios explicitly rejected the UN as a viable platform for AI governance during his address to the UN Security Council, urging last September direct "cooperation of statesmen" instead. This echoes Secretary of War Stimson's 1945 advice to Truman — a recommendation that, had it been followed, might have saved the Baruch Plan for international nuclear control.

Xi and AI Labs Are Loudly Calling for a Treaty to Prevent Loss of Control

On July 18th, while launching a 29-nation AI alliance, President Xi called for shared benefits and issued stark warnings about the risk of (permanent) loss of human control over AI. His statement, "We must make its oversight and governance precise and effective, and constantly refine measures to forestall loss of control", unequivocally signals the need for highly empowered global institutions with "radical optionality," a concept advocated by Caleb Watney and others.

Then, on July 29th, the two leading AI labs, Anthropic and OpenAI, and most AI lab leaders issued a stark statement: "We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” — notably in terms of a request and not a mere appeal.

The Need for a Moonshot Diplomatic Effort

These breakthroughs offer a profound sense of hope for safe and fair AI governance; more progress has been made in the past three months than in the previous three years combined.

However, the US Administration is still far from a cohesive vision for a treaty and a treaty-making process capable of uniting key powerbrokers and engaging effectively with China and other middle powers. Much work remains to be done:

• Not Just Any Treaty. Several signers of the July 29th statement have for months expressed deep concerns about an AI treaty. 

  • Some fear it will turn into a global authoritarian dystopia (e.g., Amodei, Altman, Thiel, Karnofsky). Some are skeptical that it can prevent loss of control, as we have "entered the singularity" (Altman, Musk), i.e., a point of no return towards a (permanent) loss of control and AI takeover. 

  • Many fear a bad treaty can lock away astounding future human benefits, including forestalling other human extinction risks (Bostrom, Amodei, Thiel), or inadvertently create immense digital suffering. 

  • A powerful minority appears to be outright opposed to preventing AI loss of control (e.g., Page, Brin, Thiel) or dismisses it as "nonsense" (e.g., Zuckerberg, Huang).

• Treaty-making Process. Given the past failure of treaties and treaty-making, the extremely short timelines, the immense scope of the treaty, and the right fears for global authoritarianism, it must be truly extraordinary in scale, speed, and effectiveness while being rooted as much as possible in past successes and evidence:

  • Scale and speed: Does the timeline require something closer to an "Apollo Program for an AI treaty" — a heavily resourced process (including specialized private philanthropies) at wartime pace — than to a sequence of periodic summits? What's the role of ultra-high-bandwidth diplomatic infrastructure? How best turn those unprecedented commitments into suitable results?

  • Model: Should a bilateral US-China deal precede multilateralization, and if so, what treaty-making model follows it? CBPAI has argued for a GDP-adjusted constitutional convention model. Are there any other treaty-making models capable of preventing vetoes, catalyzing sufficient global consensus within months for a deal of such scope, and preventing excessive concentration of power by ensuring federalism and subsidiarity? If not, how should such a model be deployed in detail?

  • Enforcement and communication infrastructure: How can we best leverage modern state-of-the-art treaty enforcement and diplomatic communication socio-technical systems, and jointly create more suitable, trustworthy, and mutually trusted ones at wartime speed — ensuring "sufficiently" reliable verifications (rather than "perfect"/"truly" ones, as some claim, impossible in complex real-world systems)? 

• Citizen co-ownership models. Citizen co-ownership of AI labs could be implemented in many forms, most of which are ineffective, repealable, or worse. An informed debate about them is in its infancy, including why they can work only if integrated into a global AI treaty.

Once again, as in 1946, a dozen US executive-branch officials and individuals have the unique opportunity and responsibility of persuading another pragmatic US president to pursue a bold treaty to manage a catastrophic new technology — closing what would be by far history's boldest, most impactful and positive deal: the Deal of the Century.

Conveners and Steering Board

The Roundtables are convened through a steering board of US AI treaty-advocacy organizations, now forming, each holding a seat with full decision rights over program, invitations, outputs, and the selection of co-hosts and partners. The board operates by consensus under a Convening Protocol that binds the conveners, protects every partner's right to withdraw, and renews association edition by edition. Board members fundraise jointly for the series.

The Coalition for a Baruch Plan for AI — an international coalition of NGOs, advisors and contributors operating in the US as a 501(c)(3) fiscally sponsored by Anti Entropy Inc." — conceived the Roundtables and serves as secretariat: research base, protocol, mediation design, and logistics. A Washington institutional co-host anchors the event; supporting partners contribute participants, expertise, or resources.

The research base is CBPAI's Deal of the Century initiative: a series of strategic memos with over 356 pages of feasibility analysis, scenario planning, and executive-branch policy recommendations, 150 pages profiling US AI policy powerbrokers, and over 90 briefings and consultations with their advisors and officials in Washington and the Bay Area — including, for four of those powerbrokers, over 30 hours with senior staff who report directly to them. That research is an input to the Roundtables, not their premise.

Those powerbrokers are certain executive-branch officials (Kratsios, Rubio, Vance, Bessent, NSA director Rudd, Chief of Staff Wiles); certain political, religious, and media figures (Bannon, Pope Leo XIV, Carlson, Beck, DeSantis, Rogan); and certain humanist AI lab leaders (Amodei/Clark, Altman, Hassabis, Musk, Sutskever).

CBPAI was launched in December 2024 with a Strategic Memo and Open Call that stated: "We call on all heads of state, President-Elect Trump and President Xi, and their advisors and security agencies, to engage in an open global treaty-making process for safe and fair AI of radically unprecedented scope, urgency and effectiveness." Its director, Rufo Guerreschi, brings rare convening and mediation experience from the 2015–2021 Free and Safe in Cyberspace series.

Expected Outcomes 

The Roundtables produce no joint statement, declaration, communiqué, or signatory list, and take no position on behalf of participants, co-hosts, or partners.

Their intended product is a single document: a shared, documented map of where US executive-branch AI policy powerbrokers genuinely converge and diverge on treaty-making design and safeguards — precise enough for participants to carry into their own work, and honest about what remains unresolved. It is circulated to participants under the Chatham House Rule, attributes no position to any named person or institution, and is not published.

CBPAI intends the Roundtables to be the first in a series held at short intervals. Co-host and partner association applies to this edition only and is renewed, or not, for each subsequent one.

Participants

Primary participants will be advisors, staff, and envoys of key formal and informal US executive-branch AI policy powerbrokers, as well as individuals best placed to brief them.

Secondary participants will be selected among:

• Relevant DC-based or Valley-based experts, NGOs and think tanks that are fit to contribute, validate and/or introduce — especially those close, influential or respected by the IC and/or the Administration. 

• Relevant experts within our extensive CBPAI network and beyond.

Participants are selected by the steering board together with co-hosts. Co-hosts are notified of the invitation list and the program in advance, and may withdraw at any time.

Join the Steering Board, Co-Host, Partner, or Participate

The conveners commit to this Convening Protocol. Every co-host and partner is protected by it and may withdraw at any time.

• Steering-board member (organizations with a treaty-related mission; full decision rights, joint fundraising): email rufo@trustlesscomputing.org with a note on your treaty work and times for a 30-minute Zoom.

• Co-host or supporting partner: email rufo@trustlesscomputing.org, attaching an as-is or amended Convening Protocol, with times for a 20-minute Zoom.

• Individual participant: email cbpai@trustlesscomputing.org.

If you like what we do more generally, pleasejoin ordonate.

Guidelines

The Missing Lane

Since March 2026, the Pro-Human Movement has assembled the broadest US coalition ever built for AI regulation — faith leaders, unions, parents, and political figures from Bannon to Sanders. Its declarations, state campaigns, and job-loss trackers are winning the domestic argument. None of them answers the question that scientists, lab leaders, and President Xi are now asking aloud: what stops the race to superintelligence across borders?

Treaty design work exists — drafted statutes, verification research, red-line calls with an end-2026 deadline. What does not exist is a process by which the US actors who decide such things converge on whether and how to pursue one. Convenings in DC and the Bay Area have concentrated on guardrails, standards, and narrow technical or geopolitical questions; they have rarely addressed a treaty proportionate to the risks, and have treated civilian and military governance separately.

The Roundtables are built for that gap. Given the dismal track record of treaty-making in recent decades, the process must be largely designed rather than improvised through a sequence of summits — "realistic first steps" are unfit for the timelines and scope of the problem. The Roundtables are the treaty lane the pro-human movement has not yet built, run by the organizations that have made a treaty their central ask.

A Scientific Approach to Convergence

The roundtables will utilize the most effective negotiation and co-decision mechanisms to, in part, (1)negotiate, converge, and find win-win among different positions, worldviews, and interests; and, in part, to (2) jointly explore and seek the best means to achieve shared broader objectives, while transparently exposing and tackling the most fundamental disagreements.

The deliberations will not primarily be framed as negotiation between adversaries. As with the Acheson-Lilienthal process of 1946, they assume participants share substantially the same ultimate goals and diverge mainly in predictions, priorities, and maximization functions. They will therefore deploy several best-of-breed, scientifically validated convergence methods, each serving a function no single one can cover.

Before the event, anonymized expert elicitation via the Delphi method and opinion-cluster mapping via Polis will produce a shared “map of divergence.” During sessions, double crux will isolate the empirical beliefs that actually drive each disagreement, while adversarial collaborations will pair skeptics and proponents to co-draft agreed-upon facts and residual disagreements. Convergence will be measured, not asserted — via before-and-after Deliberative Polling and, potentially, AI-assisted synthesis like DeepMind’s Habermas Machine, shown in Science to measurably increase group agreement. 

Lobbying and Advocacy Disclaimer

The Roundtables are convened by a steering board of organizations, several of which advocate — separately and in their own terms — for an international AI treaty, and are served by the Coalition for a Baruch Plan for AI, which advocates for a US-China-led treaty exclusively toward the US executive branch. The Roundtables themselves are not an instrument of that advocacy. They are a closed-door deliberation on open and contested questions and produce no joint statement, declaration, or signatory list. Positions taken by any convener elsewhere are that convener's alone. Co-hosts, partners, and participants are bound by nothing beyond their own remarks. The conveners are bound by the Convening Protocol.

Roundtables' Assumptions

The Roundtables proceed from a stated premise: that the loss-of-control and power-concentration risks of advanced AI may require an international instrument, and that the design of the process for reaching one is under-examined. The premise is stated, not assumed proven — participants are not asked to accept it, and several will not.

Within that scope, the questions are genuinely open and the program is built to keep them so. Sessions on loss of control include those who consider the risk remote or manageable domestically; sessions on citizen co-ownership include those who consider ownership mandates unnecessary or counterproductive; sessions on process include those who doubt any treaty-making can work on the relevant timelines; the predictions session includes transhumanist and other positions argued by their proponents rather than characterized by others.

More information