In March 2026, over 200 protestors walked to the offices of Anthropic, OpenAI and xAI, demanding that their CEOs commit to a voluntary “pause” on frontier AI research.

Their core concern was—advancements in Artificial Intelligence will one day result in the state of artificial superintelligence–a hypothetical scenario where AI systems will develop capabilities far superior to humans in almost every domain.

A month later, senator Bernie Sanders invited AI safety experts and China based academicians to discuss the “existential threat of AI” in a high profile panel discussion on Capitol Hill. He also called for a ban on superintelligence ahead of the Trump-Xi meet in China in May.

Now, Anthropic AI has proposed a global agreement between AI companies to “slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the (AI) technology.”

In a detailed statement “When AI builds itself” , Anthropic said that internal findings of their organization have revealed ‘self-building capability’ of AI models, with possibility of ‘fully-recursive self improvement’—a feedback loop where an AI system uses its output to make itself better.

The statement by Anthropic comes in the wake of the AI startup company confidentially filing for an IPO with the Securities and Exchange Commission (SEC) last week, as the company heads towards a valuation of $1 Trillion.

Anthropic also had an extremely limited preview launch of its frontier AI model Claude Mythos earlier this year, which as per company’s and independent research findings, puts it ahead of its competitor AI models.

The “pause” proposed by Anthropic comes with several conditions—it has to be unanimous, transparent and duly complied by all companies across the globe–against the backdrop of ongoing  the AI dominance race between the U.S. and China.

In this exclusive article, AI FrontPage speaks to AI risk experts who were part of the ‘Stop the AI Race’ in March 2026, to understand whether the pause proposed by Anthropic is the same as they demanded. We also breakdown the risks related to superintelligence, what Anthropic’s own data and independent data says about fully recursive self-improvement capability of AI models.

Why is Anthropic Calling for a Pause?

In a detailed statement published on June 4, Anthropic claimed that their latest research data shows AI is “already accelerating the development of AI systems”, with future possibilities of a full recursive self-improvement, suggesting their Claude AI model is showing signs of building AI.

Anthropic said that the current trajectory of advancements in AI shows in future a 100-person company could do the work of 10,000-100,000 persons organizations. It also claimed that more than 80% of the code merged into Anthropic’s own production codebase in May 2026 was authored by Claude.

METR (Model Evaluation and Threat Research) has verified that Claude Mythos Preview can sustain autonomous tasks for 16 hours or more (compared to Claude Opus 3’s four minutes in 2024). Its code-optimization benchmark also achieved an astonishing 52x speedup.

“This would revolutionize knowledge work and government services, but could also be turned to harmful ends, from authoritarian surveillance of whole populations to influence operations that tailor manipulation to each individual and run at a scale no human team could match,” read a statement by Anthropic.

In such a scenario, AI building itself might also increase the risks of humans losing control over AI systems, argues Anthropic as it suggests that the world must have an option to “slow or temporarily pause” frontier AI developments to “enable societal structures and alignment research to keep up with the advances of AI.”

In other words, Anthropic has said that human society is simply not ready for an advanced stage of artificial intelligence.

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx

— Anthropic (@AnthropicAI) June 4, 2026

Anthropic says that it will commit to a voluntary pause on frontier AI research only when its competition and other countries’ governments also commit to it.

“A unilateral pause by one lab, by contrast, is achievable immediately, but accomplishes much less: it would change who the front-runner is, but it would not create the wider deliberative process that is currently missing,” argues Anthropic.

An interesting point to note is that initially Anthropic’s policy had the mandatory clause of delaying the development of AI models that exceeded its internal safety controls unless safeguards were in place.

In February 2026, the company introduced a new policy and quietly dropped this particular commitment to pause, amid backlash from AI Safety activists. This was a precursor to the unveiling of Claude Mythos preview model by Anthropic, which is deemed as the most advanced AI model that can pose severe cybersecurity risks.

What is Artificial Superintelligence and Why is it Risky for Humanity?

Superintelligence refers to the hypothetical state where an AI agent/system vastly surpasses human level capabilities in almost every domain including coding, reasoning, complex problem solving and decision making.

Currently, the advancements are at a narrow AI range with the expected trajectory to first reach human level capabilities known as the state of artificial general intelligence (AGI) and maybe, eventually reach superintelligence. There seems to be ambiguity regarding the expected time frame for AI advancements to reach superintelligence levels.

The term ‘superintelligence’ was made popular from the 2014 book ‘Superintelligence–Paths, Dangers, Strategies’ by philosopher Nick Bostrom. The book argues that superintelligence once created would be difficult to control and would achieve its own goals, caring or not caring for the implications on humanity.

To explain his point, Bostrom uses the picture of an owl on his bookcover and begins the book with a fictitious story of Scronkfinkle, a one-eyed sparrow who lives among a group of sparrows on a tree.

The story begins with how sparrows come up with the idea of adopting an owl chick and raising it, in order for them to use the owl to build their nests and protect them from enemies.

While the majority of the sparrows agree with the idea, it was only Scronkfinkle who demanded that the owl chick be trained before inviting it to live among them. The sparrows however shrug off the question, by stating, “why not get the owl first and work out the fine details later?”

There seems to be very little agreement between researchers and scientists when it comes to superintelligence and what the future of humanity would look like in that scenario. However, autonomy of the AI system to the extent of risking humans’ control is a recurring theme.

The International AI Safety Report, prepared under the leadership of Professor Yoshua Bengio of Quebec AI Institute and unveiled at the AI Action Summit 2025 in Paris, France, has mentioned “loss of control” scenarios as one of the major risks, where AI systems come to operate outside of anyone’s control, with no clear path to regaining control.

“Expert opinion on the likelihood of loss of control within the next several years varies greatly: some consider it implausible, some consider it likely to occur, and some see it as a modest-likelihood risk that warrants attention due to its high potential severity,” read the AI Safety Report 2025.

Similarly, Anthropic is also not able to correctly predict the futuristic scenario where AI systems are “completely aligned” and have reached the stage of full recursive self-improvement.

“We do not have good intuitions for what this world would look like, because our economy is currently driven by humans and human-built tools,” said Anthropic in a statement.

“By its nature, a world driven by fast recursive self-improvement could become dominated by the self-improving model as its capabilities fully eclipse those of humans and the model proliferates across the broader economy. It is difficult to predict what the economy looks like if human labor stops being competitive,” it added.

What do AI Safety Activists Say About Anthropic’s Pause Offer?

Back in March 2026, AI FrontPage had conducted an exclusive interview of filmmaker-activist Michael Trazzi, who had organized the ‘Stop the AI Race’ protest in San Francisco. In the interview, Trazzi had claimed that he had better hopes from Dario Amodei, the CEO of Anthropic, after the company had quietly dropped its commitment to pause frontier AI research if the model is deemed too risky.

Also Read: “Just Say You’ll Pause If Everyone Pauses”: Activist Who Led America’s Largest Anti-AI Protest

Now, when asked about his comment on Anthropic’s proposal, Trazzi says, “Anthropic expects it ‘would slow down or temporarily pause’ if other developers at or near the frontier verifiably did too. This is a good first step, but what we really need is a binding commitment from Dario Amodei, and the actual verification regime that would satisfy them.”

Trazzi also points towards the recent IPO filing by Anthropic and says, “they claim to be working on the verification regime, and say they’ll organize conversations around full recursive self-improvement (RSI) and coordination. But when a $ 1 Trillion company is actively working towards RSI files for IPO, I think we need much more than just conversations.

Similarly, Professor David Krueger, renowned AI risk expert and one of the panelists of Sanders’ debate on “Existential Threat of AI” also shared his views on superintelligence and why AI companies need to stop, in a recent conversation with AI FrontPage.

“AI companies are not able to reliably determine whether their systems are safe with any reasonable level of confidence.  Despite this, they have made it clear that they intend to continue racing to build and deploy more and more powerful AI systems, even putting AI itself in charge of that,” said Krueger.

“This process is known as “recursive self-improvement”, and widely acknowledged to be a probable point of no return.  It’s like humanity is driving a car in a straightjacket while blindfolded, and AI companies are pushing the gas instead of the brakes,” he added.

There are many examples of @AnthropicAI declaring one thing and doing something completely different.

If Anthropic truly does support a global pause, it should put these words into action by funding a global campaign. Hire the best strategists and lobbyists to persuade… https://t.co/MzwWTTVUkx pic.twitter.com/hrZzjqWfPD

— Michael Huang ⏸️ (@michhuan) June 7, 2026

Similarly, Michael Huang of Pause AI group opines, “If Anthropic truly does support a global pause, it should put these words into action by funding a global campaign. Hire the best strategists and lobbyists to persuade governments to write and pass an international AI treaty with verifiable red lines.”

Conclusion: Is a Pause on Frontier AI Possible?

In a post-pandemic world, it is possible to imagine governments coming together for a global framework for threats posed against humanity by AI.

Recently, Pope Leo XIV launched his first encyclical Magnifica Humanitas, wherein he called for greater emphasis on humanity and ensuring that AI serves people rather than concentrating power in the hands of few. Incidentally, Christopher Olah, co-founder of Anthropic, was the only senior AI industry representative present with Pope during the launch of Magnifica Humanitas.

Anthropic, in its statement, has cited the example of the 1987 Intermediate Range Nuclear Forces Treaty between the U.S. and Soviet Union (later discontinued in 2019) to show that an international agreement on AI is possible.

However, with the on-going AI dominance race between the U.S. and China, a global pause on AI is easier said than done. Anthropic AI Pause statement also notes that a global voluntary pause is trickier than a nuclear arms treaty as “training runs are far easier to conceal than missile silos as their inputs are general-purpose, and the incentive to defect quietly is enormous.”

As for the future of humanity and artificial intelligence, what we know for sure is that we know nothing.

Coming back to philosopher Nick Bostrom’s story of one-eyed sparrow Scronkfinkle; in his book Superintelligence, the author consciously leaves the story open-ended, with nobody knowing what happened to the group of sparrows who went looking for the owl.

Also Read: “Summoning an Alien Species”: David Krueger on Why Superintelligence Could End Us All