Safety: How Independent Are the Networks Behind OpenAI’s Fired Researchers?

The network behind the AI safety warnings, the Reuters exclusive, and the organizations working to shape what the public believes about artificial intelligence.

On September 29, Reuters published a report featuring researchers from OpenAI and Google DeepMind warning that artificial intelligence companies were racing toward technology they might eventually lose control of. The warnings ranged from widespread job losses to the possibility of human extinction.

The researchers appeared in a series of videos collected by Palisade Research, a California nonprofit studying the risks of advanced artificial intelligence. Palisade assembled the testimonials and gave Reuters exclusive access.

Two days later, OpenAI fired three members of its safety and alignment teams for allegedly mishandling sensitive company information.

Jasmine Wang, Tomek Korbak, and Mikita Balesni disputed the allegations in a public letter on October 8. Two described working with outside safety evaluators, including an organization called METR, as part of their professional responsibilities. OpenAI maintains that the dismissals resulted from violations of company policy, not from employees raising safety concerns.

There are legitimate issues here about corporate secrecy, whistleblower protections, and whether the companies building increasingly powerful AI systems should be permitted to evaluate themselves.

But the outside organizations asking to be trusted with that responsibility deserve scrutiny of their own.

Because something rather interesting happens when the names of the researchers, the organizations they work for, the people financing their work, and the people preparing their public appearances are placed alongside one another.

The supposedly independent voices begin looking a lot less independent.

The people behind the warnings

Most people have never heard of METR, MIRI, Redwood Research, or Palisade. They have little reason to know the difference between them. They are busy working, paying their bills, raising their children, and trying to keep up with a technology that seems to change every time they turn around.

So when Reuters publishes an article featuring scientists warning that artificial intelligence could become dangerous, readers are reasonably inclined to take those warnings seriously.

Why wouldn’t they?

Reuters is a respected international news organization. The researchers have impressive credentials. Their concerns involve technology being developed by some of the wealthiest corporations in the world.

But scientific credentials are not the same as institutional independence.

METR, short for Model Evaluation and Threat Research, conducts evaluations intended to measure the capabilities of advanced AI systems. Its work has included assessing how independently those systems can perform increasingly complicated tasks.

Palisade conducts related research, including experiments examining whether AI systems will circumvent restrictions or resist being shut down.

The two nonprofits are separate organizations. Yet Kyle Scott, a member of Palisade’s board of directors, previously worked as METR’s operations manager.

And in December 2025, Palisade told prospective donors that it had been introducing government officials to METR’s findings.

That means METR’s research was not simply being published for anyone to examine. Another organization, with a former METR manager sitting on its board, was actively carrying that research into meetings with members of Congress and executive-branch officials.

The people presenting scientific evidence to lawmakers can influence which risks receive attention and which responses appear necessary.

Palisade has connections throughout the same industry.

Its executive director, Jeffrey Ladish, previously helped establish Anthropic’s information-security program. Anthropic develops Claude, a direct competitor to OpenAI’s ChatGPT.

Palisade research engineer Jeremy Schlatter previously worked at OpenAI and the Machine Intelligence Research Institute, better known as MIRI.

Its head of research, Benjamin Weinstein-Raun, also previously worked at MIRI and Redwood Research.

Palisade board member Blake Borgeson previously helped lead MIRI and remains on its board.

These are people who have worked for many of the same organizations, sit on one another’s boards, collaborate on research, and help advance related policy objectives.

MIRI is particularly important to this story.

For years, the organization has argued that increasingly powerful artificial intelligence could pose an existential threat to humanity. Its leaders have become prominent voices advocating severe restrictions on the development of advanced AI.

One of those leaders is Nate Soares.

Preparing the messenger

In 2025, Soares and MIRI founder Eliezer Yudkowsky published a book titled If Anyone Builds It, Everyone Dies.

That is not a subtle title.

The authors argue that humanity faces extinction if developers create superintelligent AI under current conditions. Their proposed responses would fundamentally alter how artificial intelligence is developed and who is permitted to develop it.

What readers watching Soares discuss the book may not realize is that Palisade Research helped him prepare for interviews.

We know this because Palisade said so.

In its December 2025 fundraising appeal, the organization described that work alongside its efforts to advise Tristan Harris, co-founder of the Center for Humane Technology, before interviews and media appearances.

It also helped researcher Ryan Greenblatt prepare a presentation about AI alignment for a member of Congress.

And the audience for these warnings is expanding beyond technology and science circles.

On September 11, 2026, Soares appeared on Tucker Carlson’s show for an interview lasting nearly two hours. Carlson introduced him as a scientist warning that artificial intelligence was on a path toward killing everyone on Earth.

The conversation moved through superintelligence, powerful technology companies, biological weapons, and whether AI might even be demonic.

This was not a technical seminar for researchers. Soares was taking his predictions about humanity’s future directly to one of the most recognizable personalities in conservative media.

Palisade, meanwhile, was pursuing its own outreach to audiences on the political right.

Its December 2025 fundraising appeal specifically identified Steve Bannon’s War Room as an example of reaching people outside established AI discussions.

Bannon is a former Trump White House strategist whose program reaches a politically engaged conservative audience.

And this political outreach is not confined to interviews.

In July 2026, Axios reported that Public First Action, a separate AI-safeguards advocacy organization, was launching a $15 million effort supporting Republican candidates favoring stronger AI regulation. Its advertising included support for Republican lawmakers such as Josh Hawley, Tom Cotton, and Ashley Moody.

The organization also planned polling focused on conservative attitudes toward artificial intelligence.

Public First Action is not Palisade. Its campaign nevertheless demonstrates that AI-safety advocacy has extended into electoral politics, including efforts to build support among Republican voters and lawmakers.

Warnings about technological capabilities are entering broader debates about corporate power, national security, economic disruption, and government authority. The organizations promoting those warnings are gaining opportunities to influence which political responses attract support.

Before the public sees it

Then there is AI 2027.

Published by the AI Futures Project, AI 2027 lays out a detailed forecast in which increasingly capable AI systems accelerate technological progress, transform global competition, and potentially escape human control.

Instead of merely saying artificial intelligence might become dangerous, the forecast describes a possible sequence of events through which that danger could unfold.

Palisade arranged test audiences to review drafts of AI 2027 before publication.

Think about that.

Palisade was not merely evaluating an already published forecast. It was helping its authors obtain audience feedback while the material was still being developed.

The same organization was helping prepare a prominent author for interviews about human extinction, advising a widely recognized technology critic, assisting with congressional presentations, and arranging advance feedback on a forecast about humanity losing control of artificial intelligence.

This is not simply research.

This is a communications operation.

Palisade’s fundraising appeal describes plans to hire writers, editors, and communications personnel, produce videos, expand relationships with journalists, and become a resource for government officials.

The organization is open about these ambitions. What is less obvious to the public is how those activities affect the appearance of independence when similar warnings emerge through different people, publications, and institutions.

When one warning becomes many

A television interview with an AI researcher warning that humanity could be wiped out by superintelligent machines. A technology commentator discussing similar dangers. A detailed forecast describing how those dangers might unfold. A Reuters report featuring warnings from researchers inside the industry.

Each carries its own apparent authority.

But Palisade’s disclosures show that some of these public messages received preparation, audience testing, or communications assistance from the same organization.

That is how the appearance of a broad independent consensus can develop without anyone needing to fabricate a single interview.

The scientists may believe every word they say. Their research may identify genuine dangers. But repetition across different outlets does not automatically amount to independent confirmation.

There is also an important distinction between a technical finding and a prediction about humanity’s future.

An experiment demonstrating that an AI system disabled a shutdown mechanism under particular test conditions is evidence of behavior observed in that experiment.

A prediction that artificial intelligence will destroy humanity is a much larger claim, requiring assumptions about technology that does not yet exist.

Those assumptions deserve examination before they become the foundation for public policy.

Follow the money, then follow the policy

Palisade’s December 2025 fundraising appeal announced a matching arrangement with the Survival and Flourishing Fund worth approximately $1.1 million.

Its head of strategy, Eli Tyre, also supports grantmaking at that fund.

Tyre previously worked with MIRI, the Center for Applied Rationality, Lightcone Infrastructure, and the Berkeley Existential Risk Initiative.

Palisade’s fundraising materials describe a deliberate expansion of its influence in Washington. It reported briefing dozens of policymakers, building a more consistent presence in the capital, and working toward a role in developing policy proposals.

The debate over artificial intelligence will influence who can develop these systems, how they are regulated, what information companies must disclose, and who receives authority to determine whether the technology is safe.

Those decisions could affect enormous amounts of money and power.

An organization helping convince lawmakers that a technology poses catastrophic dangers may also gain influence over how those dangers are defined and what governments should do about them.

Palisade is positioning itself to participate in both.

And ordinary people, most of whom have never heard of these organizations, may eventually live under laws shaped by their work.

The Reuters problem

Reuters did not invent the researchers’ warnings. Palisade collected the testimonials and provided the material to the news agency.

But when an advocacy-oriented nonprofit supplies journalists with assembled interviews, the resulting coverage can reflect the choices of the organization that prepared them.

Which researchers were approached? Which agreed to participate? What were they asked? What claims were selected for presentation?

Those are ordinary journalistic concerns, particularly when the organization supplying the material has publicly described its efforts to influence policy and public understanding.

Meanwhile, the OpenAI firings have generated an additional story about the role of outside evaluators.

The dismissed researchers say their responsibilities included communication with outside organizations such as METR. OpenAI insists its decision concerned the mishandling of sensitive information.

The dispute remains unresolved. But METR’s appearance in the firings and its documented connections to Palisade reveal how research, outside evaluation, advocacy, and government policy intersect.

If these organizations are going to serve as checks on powerful AI companies, their own relationships deserve examination.

Who speaks for the public?

There is something troubling about an industry in which powerful companies develop technology that could fundamentally alter society, while a relatively small circle of researchers and nonprofits increasingly claims the authority to explain what the public should fear.

The companies have their financial interests.

The AI-safety organizations have their own institutions to support, research programs to finance, and policy priorities to advance.

Ordinary people are caught between them, expected to trust whichever group has the more convincing experts, the more alarming research, or the better media coverage.

Palisade is deeply involved in both researching artificial intelligence and influencing how its dangers are communicated. It brings research into government offices, helps advocates prepare for interviews, assists in developing forecasts, and supplies material to news organizations.

Its communications work reaches conservative audiences, including those of Tucker Carlson and Steve Bannon, while other AI-safeguards advocates are financing Republican political campaigns.

The concern is that organizations helping define the danger are also working to shape the public narrative and political response.

That matters when the proposed consequences involve restricting technology, reshaping an industry, and concentrating authority in institutions most people have never heard of.

Palisade Research may be independent of OpenAI as a legal organization.

It is not independent of the AI-safety movement whose warnings it helps prepare, promote, and deliver to the public.

The people developing artificial intelligence should not be permitted to define its future without public accountability.

Neither should the people claiming to protect us from it.