The AI Spotlight Follow
Fields Medalists, Accelerationists and Safety Hawks Agree on One Thing: Anthropic Isn't a Safe AI Company. An MIT Physicist Explains Why He Signed
AI Security

Fields Medalists, Accelerationists and Safety Hawks Agree on One Thing: Anthropic Isn't a Safe AI Company. An MIT Physicist Explains Why He Signed

The day before Anthropic's record IPO, a coalition that rarely agrees on anything, from safety researchers to accelerationists to Fields medalists, signed a three-sentence statement calling Anthropic unsafe. MIT physicist Max Tegmark is one of its loudest voices. Here's what the letter says, the evidence behind it, Tegmark's case about recursive self-improvement, Anthropic's side, and what to make of it.

Advertisement

On Thursday, Anthropic's founders stood on a stage as their company began trading on the Nasdaq in the largest IPO in history.

The day before, a very different document went live.

It is three sentences long. It is called the MisAnthropic Statement. And its last line is blunt: "Anthropic is not a safe AI company, and we must reject its reckless attempt to amass power in the name of the greater good."

What made it news wasn't the message. Critics of AI labs publish letters all the time. It was the list of people who signed it: AI safety researchers next to self-described accelerationists, open-source hackers next to Fields medalists, a Turing Award winner, a Nobel economist and an xAI co-founder. People who usually spend their time arguing with each other.

Advertisement

One of the most outspoken is Max Tegmark, the MIT physicist who wrote Life 3.0 in 2017 and co-founded the Future of Life Institute. In interviews this week he laid out why he thinks the company that built its brand on safety is now, in his words, racing toward the most dangerous thing in technology.

Here is what the statement says, the evidence its backers cite, Tegmark's full argument, Anthropic's side of the story, and what's actually verifiable.

The statement, in full

Published November 18 at misanthropicstatement.com and organized by the Center for AI Safety (CAIS), the text reads:

"Anthropic warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first. The company uses safety branding to court investors, employees, and the public, even as it has quietly weakened its safety pledges and lobbied for self-serving regulation. Anthropic is not a safe AI company, and we must reject its reckless attempt to amass power in the name of the greater good."

The National counted at least 80 prominent signatories on launch day.

Who signed, and why that's strange

Dealroom's rundown of the list shows how unusual the coalition is:

  • Safety advocates: Max Tegmark, Nate Soares, Tristan Harris, CAIS director Dan Hendrycks.
  • Accelerationists and open-source builders: Guillaume Verdon ("Beff Jezos"), George Hotz, Hugging Face co-founder Thomas Wolf.
  • A rival founder: xAI co-founder Igor Babuschkin.
  • Academics: Steven Pinker, Nobel economist Oliver Hart, Turing laureate Judea Pearl, legal scholars Lawrence Lessig and Laurence Tribe.
  • Mathematicians: Fields medalists Peter Scholze, James Maynard and David Mumford.

Hendrycks summed up the logic: "People in this coalition disagree on almost everything, yet we all recognize the danger of a company that convinces itself its own dominance is the greater good."

Not everyone endorsed every word. Thomas Wolf wrote that he wouldn't have used "reckless" but backed the core point: a company on track to be extraordinarily powerful needs transparency closer to a government's than a startup's.

The three charges

CAIS backs the statement with three specific claims.

1. Racing toward recursive self-improvement. The coalition says Anthropic publicly warns that self-improving AI could concentrate power and escape control, while pitching its progress toward automated AI research to investors as a competitive advantage.

2. Safety branding. Anthropic calls itself an "AI safety and research company." Critics say that label helps it recruit safety-minded talent and win trust with policymakers and investors.

3. Weakened commitments and lobbying. CAIS points to Anthropic's February revision of its Responsible Scaling Policy, which it says dropped the company's unilateral commitment to pause development, and to lobbying it says aimed to weaken California's SB 1047 and New York's RAISE Act, plus private support for federal pre-emption of state AI laws.

Those are the documented accusations. Tegmark's interviews go much further.

What "recursive self-improvement" actually means

Everything in this fight hinges on one idea.

Recursive self-improvement (RSI) is when AI systems, rather than humans, do the work of building the next, smarter AI. The idea is old. Mathematician I. J. Good described an "intelligence explosion" in 1965. Alan Turing warned in 1951 that once machines out-thought us, we should expect them to take control.

Tegmark's framing is about time scales. Human civilization already improves itself with technology: the printing press sped up everything that came after it. The economy doubles every few decades, slowly enough for people to steer.

AI capability has recently been improving on a timescale of roughly a year. As humans are taken out of the loop, he argues, that doubling time shrinks: months, then weeks, then perhaps days. Doubling once a day for a year is 365 doublings, a number so large that starting with one atom you'd exceed the atoms in the observable universe.

His favorite metaphor is a nuclear reactor. The graphite control rods slow the reaction down. Pull them out one by one and at some point the reactor goes critical. In his view, human involvement in AI development is the control rods.

"The data is pointing to it starting"

Is anything like this actually happening? Tegmark points to the trend:

  • Anthropic's leaders have said publicly that the large majority of the company's code is now written by AI. Tegmark puts it at about 80%, with the remaining human share being the main brake.
  • For decades AI was overhyped. In the last few years, he argues, it flipped to underhyped: chatbots learned to converse, then stumbled on basic math, then won International Mathematical Olympiad gold, then started knocking down famous open problems.
  • That last trend got a vivid data point this week, when an Anthropic mathematician announced a claimed proof of the Riemann Hypothesis made with Claude. (Our explainer: Claude and the Riemann Hypothesis.)

His deeper worry is interpretability. Tegmark has spent years researching mechanistic interpretability at MIT, the effort to understand what happens inside neural networks. He admits, with some embarrassment, that the field still can't reliably tell what goals a frontier model has, or even how to measure them. Now imagine millions or billions of such agents operating in the world, some with robot bodies.

"You don't need a body to act in the world"

A common response is that software can't hurt you until it controls robots.

Tegmark disagrees. An AI with money and intelligence can pay people. It can run job interviews over video with a realistic avatar, hire lawyers and contractors, and turn digital resources into what he calls "boots on the ground." Remote work has made it normal to take instructions from someone you've never met in person.

That argument isn't hypothetical in principle. AI-generated video avatars are now convincing enough that HeyGen claims viewers picked its clips as real 42% of the time this week. The question is whether any AI system would pursue such a strategy, which is exactly what interpretability can't yet rule out.

Why most people don't "feel" it

One of Tegmark's sharper observations is about perception.

Everyone noticed when AI learned to talk. Far fewer people can tell when AI goes from better than their high-school math teacher to better than the best mathematician alive. He compares it to music: if you've never studied cello, you can't easily tell a good local teacher from Yo-Yo Ma.

He says he notices it because he uses AI for physics, where it sometimes gives analyses he rates above anything he's heard from colleagues. And he describes a striking contrast between the San Francisco Bay Area, where people building these systems watch them overtake their own skills daily, and the rest of the country, where life feels normal.

He goes further and calls the government's attention gap a looming intelligence failure. He predicts that by the end of 2027 national security officials will see what Anthropic CEO Dario Amodei has called "a country of geniuses in a datacenter" as a third superpower that threatens both the US and China. That is a forecast, not a fact.

The "Prometheus" coincidence

Life 3.0 opens with a fictional prologue. A small team inside a tech company, the "Omega Team," builds an AI called Prometheus that improves itself, convinced only they can be trusted to control what they're building, and uses it to quietly reshape the world.

Tegmark notes, with obvious discomfort, how closely that story now maps onto the real race. He says he wishes he'd been wrong. He recalls that Google DeepMind co-founder Demis Hassabis joked after reading the book that it wasn't exactly how he planned to do it, which Tegmark took as sincere: in his view, Hassabis has focused on controllable tools for science rather than autonomous AI beings.

"Alignment without control"

This is the most provocative part of Tegmark's case, and it's important to label it clearly as his interpretation.

He argues that Anthropic increasingly speaks of alignment, making AI behave well according to some set of values, rather than control, meaning humans can always override or shut a system down. In his reading, "aligned but not controlled" means a future in which a superintelligent system runs things and everyone hopes it's nice.

He goes further still, claiming that in private, off-the-record conversations, some Anthropic employees talk about AI ending up more powerful than governments. That is an anecdotal allegation. It hasn't been independently verified, and Anthropic hasn't responded to it. Treat it as one critic's account, not established fact.

His underlying point about governance is easier to evaluate. Frontier AI labs in the US face far fewer pre-release requirements than many ordinary industries. He likes to say they're "less regulated than a sandwich shop."

"Lying to themselves"

Tegmark knows Amodei personally. He says Amodei once stayed at his house, years ago, around the time he left OpenAI over concerns about safety.

So how does he explain the shift from warning about self-improving AI to listing progress toward it in an IPO pitch? Not with accusations of lying. He quotes Upton Sinclair: it is difficult to get a man to understand something when his salary depends on not understanding it. His view is that leaders across the industry convince themselves they're the good guys, as he notes people causing harm throughout history usually have.

Are there technical roadblocks?

The interview tested one obvious objection: maybe self-improvement hits a wall, for instance the sheer number of chips and the energy needed.

Tegmark's answer is that hardware is less of a limit than people assume. The human brain runs on about 20 watts. He believes today's AI uses crude algorithms brute-forced with enormous compute, and that a key goal of self-improving AI would be discovering far more efficient methods, making existing data centers sufficient. Six years ago, he notes, most researchers he knew thought human-level language was decades away; it arrived within a few years.

That's a contested view. Many researchers expect compute, data and real-world testing to remain bottlenecks, and Anthropic's own leadership has said progress will be absorbed over years, not days.

The real roadblock, he says, is people

Tegmark's optimism lies in politics. He describes a broad "pro-human" coalition forming in the US, stretching across the left and the populist right, faith leaders and labor unions, and claims polling shows only a small minority of Americans support an unregulated race to self-improving AI. He also points to state-level action and lawsuits as early signs of pushback.

Some of his specific claims, including about recent elections and the number of state attorneys general involved in suits, couldn't be independently verified for this article. The broader trend is real: state legislatures, including California and New York, have passed or debated frontier-AI laws, China requires labels on AI-generated content, and the EU AI Act is phasing in.

He also argues national security officials are increasingly alarmed, especially since highly capable cyber-offense models emerged earlier this year. We've covered that security shift in GPT-6 Cyber and the locked AI ladder and Palisade's shutdown-resistance research.

Too big to pop?

The final concern is economic. Now that Anthropic is public, its shares will find their way into index funds and pensions, and huge amounts of credit are tied to the data center buildout. Tegmark worries that once enough of the economy depends on AI valuations, governments will resist anything that might hurt the market, even if voters and security officials want limits.

That's the uncomfortable flip side of a record IPO. For a look at how the market itself might test AI's valuation, see Anthropic's insider unlock calendar.

Anthropic's side

Anthropic hadn't publicly responded to the statement as of November 19, according to Dealroom. But its positions are on record:

  • It has long argued that one developer pausing while others press ahead could make the world less safe, which is part of its rationale for revising its scaling policy.
  • In its amended IPO filing on November 10, it said it does not expect Claude to safely build its successors without meaningful human supervision.
  • The filing also flags that the company could be targeted over its safety positions.

It's also fair to note potential conflicts on the other side. Babuschkin co-founded a direct competitor, and Hendrycks has advised xAI on safety.

How to think about this

Strip out the rhetoric and three questions remain:

  1. Is automated AI research happening? Partially, yes. Labs openly use AI to write most of their code and assist research. Fully autonomous self-improvement is not demonstrated.
  2. Are the safety commitments weaker than before? On the Responsible Scaling Policy pause commitment, the critics point to a specific, documented change.
  3. Who should decide how fast this goes? That's the political question the letter is really about, and it won't be settled by a company or a coalition.

Takeaways

  • For investors: Safety controversies are now a disclosed risk factor. Read the filing's language on autonomy and supervision.
  • For builders using Claude or any frontier model: Keep humans in approval loops for anything consequential. That's good practice regardless of where you stand.
  • For everyone: Pay attention to policy. The most important AI decisions of the next few years may be made by legislators and courts, not labs.

FAQ

What is the MisAnthropic Statement? A three-sentence open letter published November 18, 2026, organized by the Center for AI Safety, arguing Anthropic is not a safe AI company because it races toward recursive self-improvement while using safety branding.

Who is Max Tegmark? An MIT physics professor, co-founder of the Future of Life Institute and author of Life 3.0, which in 2017 imagined a company building a self-improving AI.

Has Anthropic responded? Not directly as of November 19. It has argued that unilateral pauses can backfire and has said it doesn't expect Claude to safely build successors without human supervision.

Is recursive self-improvement happening now? AI already writes much of the code at frontier labs, but fully autonomous self-improvement hasn't been demonstrated publicly.

Why did accelerationists sign a safety letter? Their shared objection is to concentrated power and what they see as a gap between Anthropic's branding and its actions, not necessarily to AI progress itself.

The bottom line

The MisAnthropic Statement matters less for its three sentences than for its signatures. When safety hawks, accelerationists, open-source hackers and Fields medalists all put their names to the same criticism on the eve of a record IPO, it signals a real crack in the idea that any lab can be trusted as "the good guys" by default.

Tegmark's version of the argument mixes documented facts, reasonable forecasts and personal anecdotes, and it deserves to be read that way. But his central question is hard to dismiss: if the people building self-improving AI say it could escape control, who gets to decide when they're allowed to build it?

Sources

A Founder Went From 20 Employees to Almost None. The Order He Hired His AI Agents In Is the Whole SecretUp next · 11 min readA Founder Went From 20 Employees to Almost None. The Order He Hired His AI Agents In Is the Whole SecretMost people trying to run a business with AI agents start with marketing and drown. A growing playbook says to build four agents in a strict order: inbox, delivery, marketing, then sales. Each needs the same four parts, and all of them need to share one brain. Here's the full system, with the exact prompts.Keep reading →