Bernie Sanders, Steve Bannon, and ex-Anthropic researcher Jacob Coxon unite for human-controlled AI

Bernie Sanders, Steve Bannon, and former Anthropic researcher Jacob Coxon are set to converge in Washington around an unusually broad political demand: advanced artificial intelligence should remain meaningfully under human control.

The three are expected to participate in the Pro-Human Assembly on September 15, a one-day Washington gathering built around the Pro-Human AI Declaration and a coalition that spans progressives, conservatives, labor groups, religious organizations, AI-safety researchers, artists, parents and civil-liberties advocates.

Their reasons for supporting stricter AI governance differ substantially.

Sanders is pursuing federal legislation that would temporarily pause advanced AI development and permanently prohibit artificial superintelligence.

Bannon has endorsed the Pro-Human AI Declaration, which opposes a race toward uncontrollable superintelligence and calls for limits on concentrated technological power.

Coxon recently left Anthropic after warning that frontier laboratories are racing toward self-improving systems before they know how to keep those systems reliably under human control.

The significance is therefore political rather than ideological.

People who disagree on much of American politics are beginning to converge on a narrower proposition: the decision to create systems potentially more capable than humans should not belong exclusively to the companies building them.

QUESTION	CURRENT POSITION
Event	Pro-Human Assembly 2026
Date	September 15, 2026
Location	Washington, D.C.
Sanders	Temporary advanced-AI pause + proposed permanent superintelligence ban
Bannon	Endorser of the Pro-Human AI Declaration
Coxon	Former Anthropic/OpenAI researcher warning against uncontrolled self-improving AI
Core principle	Humans should retain meaningful control over AI systems
Broader coalition	Labor, conservatives, progressives, faith groups, AI-safety organizations and public advocates
Existing declaration	Pro-Human AI Declaration, released March 2026
Central policy question	Who gets to decide how far frontier AI is allowed to advance?

··········

THE COALITION IS UNITED BY CONTROL, NOT BY A COMMON POLITICAL IDEOLOGY

Sanders and Bannon occupy almost opposite ends of the American political spectrum.

Their presence around the same AI-governance initiative would normally appear improbable.

AI changes the alignment because the disagreement being discussed is upstream of many conventional political disputes.

The question is whether humans should deliberately create systems capable of operating beyond meaningful human supervision.

The Pro-Human AI Declaration answers that question with a strong presumption in favor of continued human authority.

Its framework includes five broad areas:

PRO-HUMAN PRINCIPLE	PRACTICAL DIRECTION
Keeping humans in charge	Human authority, shutdown capability and limits on uncontrolled autonomy
Avoiding concentration of power	Prevent excessive AI power from accumulating in a small number of companies or institutions
Protecting the human experience	Limit AI systems that damage childhood, relationships or human social structures
Human agency and liberty	Protect privacy, freedom and individual control
Corporate responsibility	Liability, independent standards and accountability for AI developers

The declaration goes considerably further than a generic call for “responsible AI.”

It says development of superintelligence should be prohibited until there is broad scientific agreement that it can be created safely and controllably, combined with strong public support.

It opposes architectures designed to self-replicate, autonomously self-improve, resist shutdown or control weapons of mass destruction.

It also calls for independent oversight of highly autonomous systems and rejects the idea that AI developers should regulate themselves without external authority.

Bannon is already an individual endorser of that declaration.

Sanders has reached some of the same conclusions through legislation rather than through the declaration itself.

Coxon has reached them from inside the technical development process.

The paths are different.

The point of convergence is control.

··········

SANDERS IS TRYING TO TURN THE PRINCIPLE INTO FEDERAL LAW

Sanders' position has moved beyond warnings.

On September 3, Sanders and Representative Greg Casar announced plans for the Ban Artificial Superintelligence Act.

The proposal would create two distinct restrictions.

The first is a temporary pause on advanced AI development until a federal regulator establishes safety rules.

The second is considerably stronger: a permanent prohibition on developing and deploying artificial superintelligence.

The proposal would also direct the United States to pursue international agreements intended to prevent superintelligence from being developed elsewhere.

That distinction is technically important.

A temporary frontier pause could end once specified regulatory conditions were satisfied.

A prohibition on superintelligence would instead establish a capability boundary that developers would not be permitted to cross.

Sanders had already asked Anthropic, Meta and OpenAI in August to pause development after a series of incidents involving advanced models behaving outside their intended environments.

His broader argument is that private companies should not independently decide whether the risks of more capable AI are acceptable for society.

The Pro-Human Assembly gives that position access to a coalition extending far beyond the American left.

··········

COXON BRINGS THE ARGUMENT FROM INSIDE THE FRONTIER LABS

Jacob Coxon's participation changes the character of the debate because his warning comes from direct experience building frontier models.

Coxon spent approximately three years working on pretraining research at OpenAI and Anthropic before leaving Anthropic in September.

Pretraining is one of the foundational stages of building large models, where enormous datasets and compute resources are used to create the underlying capabilities later refined through post-training.

His criticism is therefore directed at the development trajectory itself.

Coxon has argued that frontier laboratories are moving toward self-improving AI systems without having established reliable technical methods for controlling what increasingly autonomous systems will do.

Recursive or autonomous self-improvement is particularly important in this debate.

A conventional model remains largely dependent on humans to design its successor, assemble training infrastructure, select experiments and initiate new development cycles.

A sufficiently autonomous system capable of materially improving its own software, research processes or successor models could compress that cycle.

The concern is not simply that the next model would be more intelligent.

It is that humans could lose their current position as the bottleneck controlling the rate of improvement.

That is the capability transition Coxon is warning about.

··········

BANNON’S INVOLVEMENT SHOWS THAT AI CONTROL IS NO LONGER A LEFT-RIGHT ISSUE

Bannon's endorsement of the Pro-Human AI Declaration is politically significant because many arguments for stronger technology regulation are traditionally framed as progressive policy.

The declaration instead combines concerns usually associated with very different constituencies.

Conservatives can focus on family structures, human autonomy, religious liberty and resistance to centralized corporate power.

Labor groups can focus on employment, bargaining power and the distribution of productivity gains.

Progressives can emphasize corporate accountability, inequality and democratic oversight.

AI-safety researchers concentrate on alignment, controllability, autonomous replication and catastrophic-risk thresholds.

Parents and child-safety organizations focus on addictive systems, AI companions and psychological manipulation.

These groups can disagree on almost everything outside AI while still supporting a rule that says humans must retain the capacity to understand, restrict and shut down powerful systems.

The coalition therefore does not require ideological convergence.

It requires overlapping interests around one technological constraint.

··········

DATA STUDIOS MAPS THE MINIMUM POLICY CONSENSUS

The Sanders-Bannon-Coxon combination becomes easier to understand when the positions are separated into individual policy components.

POLICY DIMENSION	SANDERS	BANNON / PRO-HUMAN DECLARATION	COXON
Human control	Strong support	Strong support	Central concern
Frontier slowdown	Supports temporary pause	Supports restrictions where control is uncertain	Supports slowing dangerous development
Superintelligence	Proposed permanent ban	Opposes development until proven safe and controllable	Warns against self-improving superintelligence
Independent oversight	Supports federal regulation	Explicit independent oversight	Argues labs cannot safely manage race dynamics alone
Corporate concentration	Major concern	Explicit concern	Secondary to technical-control problem
International coordination	Included in proposed legislation	Compatible with framework	Necessary if labs remain in competitive race
AI development generally	Does not imply banning ordinary AI	Supports useful human-controlled AI	Criticism focused on dangerous frontier trajectory

This Data Studios mapping shows that the coalition’s viable center is narrower than a shared political platform.

The overlap consists of approximately five ideas: human control, external oversight, a limit on uncontrolled self-improvement, restrictions around superintelligence, and some mechanism for preventing competitive pressure from defeating safety decisions.

Agreement becomes weaker when the discussion moves toward taxation, labor policy, corporate structure, cultural policy or the broader role of government.

That narrowness may actually make the coalition more durable.

It does not need to resolve conventional political disputes to act on a defined frontier-AI issue.

··········

THE PRO-HUMAN DECLARATION HAS ALREADY BUILT AN UNUSUALLY LARGE BASE

The political infrastructure behind the assembly did not appear this week.

The Pro-Human AI Declaration was released in March 2026 following an earlier cross-ideological gathering.

Its current supporter list includes hundreds of organizations and hundreds of individual endorsers, including AI researcher Yoshua Bengio, former Obama national security adviser Susan Rice, Steve Bannon, Glenn Beck, labor organizations, civil-society groups and AI-safety organizations.

The breadth is deliberate.

Rather than defining AI policy through the priorities of one political faction, the declaration attempts to identify positions that remain attractive when the reasons for supporting them differ.

Its own March polling of 1,004 likely U.S. voters reported that 80% preferred keeping humans in charge with strong oversight and limits, compared with 10% who favored fast and lightly regulated development.

It also reported 69% support for prohibiting superintelligence until it is proven safe, 72% support for holding AI companies legally responsible for harms, and 73% support for stronger protection of children from manipulative AI systems.

Those numbers come from polling presented by the coalition itself and should therefore be read as evidence of the organization’s claimed public support rather than as a substitute for broader independent polling.

The direction, however, explains why AI control can attract politicians who otherwise compete for completely different voters.

··········

HUMAN-CONTROLLED AI IS DIFFERENT FROM STOPPING AI

One potential source of confusion is the phrase “pro-human AI.”

The declaration does not argue that machine learning, AI assistants, medical models, scientific tools or productivity software should disappear.

Its proposed dividing line is based more heavily on autonomy and control.

AI SYSTEM	PRO-HUMAN CONTROL QUESTION
Coding assistant	Can humans supervise and override its actions?
Medical decision support	Is responsibility still assigned to accountable humans and institutions?
Autonomous agent	Can its goals, permissions and external actions be reliably constrained?
Self-replicating system	Can replication be prevented or terminated by human operators?
Self-improving frontier AI	Can humans still control the rate and direction of capability growth?
Superintelligence	Can safety and controllability actually be demonstrated before development?

This creates a potential regulatory model based on capability thresholds rather than a blanket definition of AI.

Ordinary controlled systems could continue developing rapidly.

Systems approaching autonomous replication, self-improvement or other dangerous capabilities would encounter progressively stronger requirements.

The technical difficulty is determining the threshold before a model crosses it.

Evaluations can underestimate capabilities.

Post-training can change behavior.

Agent scaffolding can turn a model that appears manageable in isolation into a much more capable operational system.

A human-control requirement therefore needs measurable tests rather than a political slogan.

··········

THE HARDEST PROBLEM IS THE AI RACE ITSELF

Coxon's warning and Sanders' legislation converge on another structural problem: one company slowing down does not necessarily slow the frontier.

If Anthropic pauses while OpenAI continues, Anthropic risks losing capability leadership.

If U.S. companies collectively slow while Chinese laboratories continue, policymakers worry about national-security consequences.

If every major laboratory believes somebody else will continue, each organization has an incentive to keep moving even when its own researchers would prefer a slower pace.

This is a classic coordination problem.

ACTOR DECISION	OTHER LABS CONTINUE	OTHER LABS ALSO SLOW
One laboratory slows	Loses competitive position	Safety benefits with limited competitive penalty
One laboratory continues	Gains relative capability advantage	Breaks collective restraint
Governments coordinate	Competitive incentives can be reduced	Most stable slowdown mechanism
No coordination	Race dynamics remain dominant	Voluntary restraint is fragile

This is a Data Studios reconstruction of the incentive structure, not a description of a formal agreement among AI companies.

It explains why increasingly many safety proposals move from corporate promises toward regulation and international coordination.

A laboratory can control its own training run.

It cannot independently control the incentives of every competitor.

··········

THE ASSEMBLY ARRIVES AS FRONTIER LAB LEADERS ARE THEMSELVES CALLING FOR RESTRAINT

The timing gives tomorrow's Washington meeting considerably more political relevance than the March declaration had at launch.

Anthropic CEO Dario Amodei has now publicly argued that the frontier may need to advance more slowly.

OpenAI CEO Sam Altman has said he agrees with the need to pace frontier development and has committed OpenAI to deeper access for independent evaluators.

Elon Musk has publicly supported Amodei's call for restraint.

At the same time, President Donald Trump has rejected the broader argument that the United States should substantially slow AI development, emphasizing competition with China.

The result is an emerging split that does not map cleanly onto conventional party politics.

Some frontier-lab leaders are asking for mechanisms that make slowing down possible.

Some politicians from both left and right are demanding stronger limits.

Other political leaders view slowing down as a strategic risk.

The central disagreement is becoming who should determine the maximum acceptable speed of AI development.

··········

WASHINGTON NOW HAS TO TURN A CROSS-PARTISAN FEAR INTO A GOVERNABLE THRESHOLD

Bringing Sanders, Bannon and Coxon into the same political conversation is unusual.

It is not yet regulation.

It does not create a technical definition of superintelligence, determine which training runs should require authorization or establish how international verification would work.

Those are the hard implementation questions.

A workable regime would need to determine which capabilities trigger enhanced oversight, how laboratories disclose those capabilities, who performs independent evaluations, what happens when a model fails them, and how regulators distinguish dangerous frontier development from ordinary AI innovation.

It would also need to survive a geopolitical environment in which both the United States and China fear losing technological leadership.

The Pro-Human coalition has therefore solved only the first part of the problem: creating political overlap around the principle that human control should take priority over an unrestricted capability race.

The next step is harder.

That principle has to become a threshold engineers can test, companies cannot evade, regulators can enforce and competing nations have an incentive to respect.

··········

THE NEW AI DIVIDE MAY BE CONTROL VERSUS SPEED

The Sanders-Bannon-Coxon alignment is evidence that the politics of frontier AI may develop along a different axis from conventional left-versus-right politics.

One side will prioritize rapid capability development, economic advantage and geopolitical leadership.

The other will insist that development speed must remain subordinate to demonstrable human control.

Individual politicians and companies may move between those positions depending on the capability involved.

The underlying question will remain the same.

If AI systems eventually become powerful enough to improve themselves, acquire resources and act across critical infrastructure, who has the authority to decide whether that threshold should be crossed at all?

The fact that Sanders, Bannon and a researcher who recently helped build frontier models can now appear in the same coalition suggests that this question is beginning to create its own political constituency.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org