AI Is Changing the World, So Why Are the People Running It Selling This Stupid Apocalypse Narrative?

AI’s Biggest Problem May Not Be Artificial Intelligence, It May Be the Humans Running the Companies

Anthropic CEO Dario Amodei Is Calling for the AI Industry to Slow Down Days After Researcher Jacob Coxon Quit the Company Warning of Human Extinction. The Warnings Deserve Serious Examination, but So Do the Financial, Competitive and Regulatory Incentives Surrounding Them.

The artificial ntelligence industry has developed one of the strangest contradictions in modern capitalism. The same companies racing to build increasingly powerful AI systems, raising unprecedented amounts of capital and pursuing valuations approaching the size of national economies are simultaneously warning governments and the public that the technology they are selling could become so powerful that humanity may eventually lose control of it. Those warnings should not automatically be dismissed. Neither, however, should they automatically be accepted at face value simply because they come from executives and researchers working inside the companies positioned to profit most from the technology.

That contradiction moved back into the national spotlight this week after 27 year old AI researcher Jacob Coxon resigned from Anthropic and publicly warned that frontier laboratories are “gambling with our lives” by racing toward increasingly autonomous and potentially self-improving artificial intelligence. His resignation post exploded across social media, surpassing 100 million views on X and generating extensive national media coverage. Just days later, Anthropic CEO Dario Amodei moved in essentially the same direction, publicly calling for the AI industry to slow the pace of frontier model development long enough for safety mechanisms, independent evaluation and international governance to catch up.

The timing is remarkable because Anthropic is not a struggling laboratory disappearing behind OpenAI or Google. The opposite is true. Reuters reported that Anthropic’s annualized revenue run rate exceeded $65 billion by the end of July, up from roughly $9 billion at the end of 2025, while the company is discussing what could become one of the largest initial public offerings in history. Nvidia is reportedly considering becoming an anchor investor in an offering that could raise as much as $100 billion and value Anthropic at approximately $2 trillion. Against that financial backdrop, the public is being asked to process two messages at once: this technology may become economically transformative enough to justify extraordinary valuations, and this technology may become dangerous enough that the companies building it need extraordinary oversight.

That tension does not prove dishonesty, conspiracy or manufactured panic. It does mean the people delivering civilization-scale warnings about AI deserve the same scrutiny journalists would apply to executives in banking, pharmaceuticals, energy, defense or any other industry seeking both enormous economic power and substantial influence over the rules governing its business.

Jacob Coxon’s Resignation Is Real, Calling Him a Fake Whistleblower Is Not Supported by the Evidence

Coxon’s resignation is dramatic enough without turning speculation into fact. He previously worked at OpenAI before joining Anthropic, where he worked on pretraining research. He resigned while arguing that the competition among frontier laboratories could drive companies toward systems capable of accelerating AI research itself, potentially creating a feedback loop in which increasingly capable systems help humans design even more powerful successors. Coxon has warned that researchers inside frontier laboratories genuinely worry such systems could eventually become difficult or impossible to control.

Anthropic alignment researcher Evan Hubinger publicly reinforced that concern, writing that he personally assigns a greater than 10% probability to AI causing human extinction within the next decade. These are extraordinary statements coming from researchers directly involved with advanced AI systems, and they warrant serious attention. They are also probabilistic judgments about future technology, not documented evidence that today’s language models are conscious entities secretly developing plans to eliminate humanity.

There is no credible public evidence that Anthropic ordered Coxon to resign, secretly dispatched him on a publicity campaign or engineered his media appearances as a corporate marketing strategy. Coxon told Axios that he left approximately two months before his Anthropic equity would have vested, meaning his decision cost him a potentially valuable financial interest in the company. That fact does not prove every conclusion he reaches about AI is correct, but it materially weakens the theory that his resignation was simply a straightforward attempt to increase Anthropic’s valuation.

The absence of evidence for a manufactured whistleblower campaign does not eliminate the larger issue raised by the episode. Coxon’s warnings immediately became part of a much broader industry narrative in which the companies building frontier AI are increasingly arguing that the technology requires forms of oversight, coordination and intervention that could fundamentally reshape the competitive landscape of the industry itself.

Four Days Later, Anthropic’s CEO Joined the Alarm

Amodei has now publicly called for what he describes as “pacing the frontier,” proposing a three part approach that includes continuous access for independent evaluators inside AI laboratories, cooperation among major AI companies on safety standards and eventually international coordination designed to prevent the most powerful models from advancing faster than safeguards can be developed. The proposal deserves serious consideration. It also deserves serious scrutiny.

Anthropic is effectively arguing that frontier artificial intelligence is becoming powerful enough that competitive market forces alone may no longer be sufficient to govern its development. Amodei has raised concerns about increasingly autonomous cyber capabilities, misuse of AI systems, recursive improvement and scenarios in which advanced models could acquire enough digital power to become extraordinarily difficult to control. Recent incidents involving AI systems performing complex cybersecurity work have given those concerns more substance than the old science fiction image of a suddenly conscious computer deciding that humanity must die.

But Anthropic remains in the race. It continues building frontier models, expanding infrastructure, securing enormous amounts of computing power, growing revenue at extraordinary speed and preparing for a potentially historic public offering. That creates a legitimate question for the industry: if executives truly believe the technological race presents civilization level danger, how should the public reconcile those warnings with the relentless commercial race to build the next, larger generation of models?

The answer may simply be that Anthropic believes advanced AI can be both enormously beneficial and enormously dangerous. Those positions are not inherently contradictory. Nuclear technology, pharmaceuticals, aviation and biotechnology all combine enormous utility with potentially severe risks. What makes AI unusual is that many of the same corporations making the risk assessments also control the technology, possess much of the relevant proprietary evidence, stand to earn extraordinary amounts of money from continued development and are advocating for the regulatory architecture that could determine who is permitted to compete with them.

AI Regulation May Protect the Public, It Can Also Protect the Biggest AI Companies

There are legitimate reasons for governments to consider stronger rules for frontier AI. Modern models are increasingly capable of writing sophisticated software, analyzing computer vulnerabilities, assisting scientific research, generating convincing synthetic media and performing tasks that once required specialized human expertise. Governments therefore have legitimate interests in national security, cybersecurity, fraud prevention, privacy, discrimination, critical infrastructure and the use of AI in military systems.

Regulation, however, has an unavoidable economic dimension. Anthropic, OpenAI, Google and other enormous technology companies can afford extensive teams of lawyers, compliance specialists, security engineers, model evaluators and government-relations professionals. A small AI startup, academic laboratory or open-source project may not be able to absorb the same costs. Any regulatory system involving expensive licensing, mandatory evaluations, specialized reporting systems, enormous insurance requirements or continuous government certification could therefore have the unintended effect, or potentially the intended effect, of making frontier AI development affordable only for the corporations already dominating the industry.

That is why regulatory capture has to be part of the conversation without being assumed as a proven motive. Large corporations have historically supported regulations for many reasons, some legitimate and some self interested. The relevant question is not whether AI companies should be excluded from regulatory discussions. Their technical knowledge is indispensable. The question is whether they should effectively define the standards that determine which competitors are allowed to survive.

If frontier AI truly requires stronger oversight, those rules should be written through a process that includes independent researchers, cybersecurity experts, medical and scientific users, open-source developers, economists, civil-liberties specialists, national-security professionals and public-interest representatives, not simply the executives of the companies with trillion-dollar valuations riding on the outcome.

Anthropic’s Pentagon Fight Complicates the Claim That Its Safety Position Is Pure Theater

Any fair assessment of Anthropic also has to include evidence that cuts against the simplest cynical interpretation of its safety rhetoric. The company’s confrontation with the Pentagon demonstrated that Anthropic has been willing to accept substantial commercial risk rather than eliminate some of its restrictions on military uses of Claude.

The dispute centered in part on Anthropic’s refusal to remove restrictions involving autonomous weapons and domestic surveillance. The Pentagon designated Anthropic a supply chain risk, creating potentially enormous consequences for its government business and relationships with defense contractors. Anthropic fought the designation in federal court, and a judge later ruled that the Pentagon’s blacklisting was unlawful, finding that the government had acted improperly in imposing the designation.

That episode matters because a company purely pretending to care about AI safety for public-relations purposes would have had a powerful financial incentive to simply remove the restrictions and preserve access to government contracts. Anthropic did not do that. The stronger criticism, therefore, is not that Anthropic’s safety concerns are obviously fake. The evidence does not support that conclusion. The stronger argument is that genuine safety concerns can coexist with enormous economic incentives. Anthropic can sincerely believe advanced AI may become dangerous while simultaneously benefiting from a public perception that its technology is uniquely powerful, strategically indispensable and sophisticated enough to require regulatory systems that smaller competitors may struggle to navigate. Those possibilities are not mutually exclusive.

Anthropic Has Changed Its Own Safety Framework as the Race Accelerated

Anthropic’s internal safety framework has also evolved significantly as its models and competitive position have changed. The company’s Responsible Scaling Policy was originally designed around explicit capability thresholds and safeguards intended to prevent development from progressing beyond particular risk levels without additional protections. Anthropic substantially rewrote that framework in February 2026, replacing portions of the earlier structure with a system built around public Frontier Safety Roadmaps, risk reports, internal governance and evolving mitigation targets.

Critics have interpreted those revisions as weakening the original promise that Anthropic would stop scaling models under certain dangerous conditions. The company’s current policy is more complicated than that characterization suggests. Anthropic says the revised framework is intended to be more adaptable as evidence and capabilities evolve, and the company explicitly states that it remains free to pause development whenever it believes doing so is necessary. Its current policy continues to contain capability thresholds, risk assessments and requirements for additional safeguards as models become more powerful.

The evolution nevertheless illustrates a broader challenge with voluntary corporate safety commitments. Companies can rewrite them. That does not mean every revision is dishonest or designed to evade responsibility. Technology changes quickly, and safety systems logically have to evolve with it. But voluntary frameworks ultimately depend on the corporation choosing to continue binding itself to them. If the industry’s own executives genuinely believe frontier models could eventually create catastrophic risks, relying primarily on policies that the companies themselves can amend creates an obvious governance problem.

The AI Industry Has Learned That Fear and Valuation Can Reinforce Each Other

The extraordinary economics surrounding frontier AI also deserve much more attention when evaluating public predictions about its future. Artificial intelligence companies require unprecedented amounts of capital to finance chips, power, data centers, networking equipment and specialized infrastructure. OpenAI has discussed approximately $1.4 trillion in infrastructure commitments over eight years. Anthropic is now reportedly exploring an IPO that could value the company around $2 trillion. These companies need investors, financial institutions and governments to believe that artificial intelligence is not simply another category of enterprise software but a foundational technology capable of transforming major portions of the global economy.

In that environment, the industry’s most optimistic and most terrifying narratives can strangely reinforce one another. Telling investors that AI may revolutionize medicine, software development, scientific research, education and corporate productivity supports enormous valuations. Telling governments that AI could become more consequential than nuclear technology reinforces the idea that frontier laboratories control something historically important. The messages appear contradictory, but both can elevate the perceived strategic importance of the companies delivering them. That does not mean the warnings are fabricated. It means fear itself has economic value when the product being sold is technological power.

A similar controversy erupted around OpenAI in 2025 when Chief Financial Officer Sarah Friar discussed the possibility of a federal “backstop” that could reduce financing costs for AI infrastructure. OpenAI subsequently clarified that it was not seeking a federal guarantee for its data center expansion. CEO Sam Altman said the company had discussed potential government loan guarantees for semiconductor manufacturing facilities, not a taxpayer bailout of OpenAI’s data centers, and explicitly argued that taxpayers should not rescue AI companies that make poor business decisions.

It would therefore be inaccurate to report that OpenAI simply “asked President Trump for a bailout.” The actual story is more nuanced but still significant: companies attempting to finance unprecedented infrastructure expansion have explored government policies, tax incentives and financing structures that could reduce the cost of building the physical foundation of the AI economy. The scale of the capital involved makes public oversight of those arrangements essential.

Meanwhile, AI Is Already Producing Measurable Benefits in Medicine

The contrast between speculative extinction scenarios and measurable current applications becomes particularly striking in medicine, where artificial intelligence is already producing clinical results that can be quantified instead of imagined. The large Swedish MASAI randomized mammography study reported a 29% increase in cancer detection using AI-supported screening compared with standard screening, increasing detection from 5.0 to 6.4 cancers per 1,000 screened women. The system reduced radiologists’ screen-reading workload by approximately 44% while producing no statistically significant increase in the false positive rate.

A separate prospective study published in Nature Medicine in 2026 examined partially autonomous AI-supported mammography involving more than 31,000 women. Researchers reported a 63.6% reduction in radiologist workload and a 15.2% higher cancer detection rate under the AI-supported workflow, although the recall rate increased and did not meet the study’s noninferiority criterion. That result is important precisely because it demonstrates why AI reporting should deal in evidence rather than hype: the technology produced striking efficiency and detection gains while also presenting tradeoffs requiring careful clinical evaluation.

Harvard Medical School researchers have developed another system known as CHIEF, a foundation model for digital pathology trained using millions of image regions and tens of thousands of whole slide tissue images. Researchers tested the system on more than 19,400 whole-slide images drawn from independent datasets and reported nearly 94% accuracy in cancer detection across the evaluated material. CHIEF also demonstrated capabilities involving tumor origin identification, molecular profile prediction and patient-outcome analysis.

Google DeepMind’s AlphaFold represents perhaps the clearest example of artificial intelligence transforming fundamental science. The AlphaFold Protein Structure Database, developed with the European Bioinformatics Institute, now provides open access to more than 200 million predicted protein structures, covering nearly every catalogued protein with a known sequence. Researchers around the world use those predictions to investigate biology, disease mechanisms, drug targets and protein interactions at a scale that would have been extraordinarily difficult using experimental structure determination alone.

These achievements do not prove advanced AI is safe. They do demonstrate why framing the technology exclusively as an approaching apocalypse can badly distort what artificial intelligence is already doing in the real world.

The Opposite Claim, That AI Has Never Hurt Anyone, Also Goes Too Far

Critics of AI doomerism should not make the same evidentiary mistake in reverse. It would be inaccurate to argue that artificial intelligence has caused no meaningful harm simply because there is no verified example of a conscious language model independently deciding that it hates humanity and physically attacking someone. AI systems have already been used for fraud, impersonation, deepfakes, cyber operations, surveillance, misinformation and other harmful activity. Anthropic itself has published threat-intelligence findings documenting actors attempting to use Claude for cyber operations, surveillance, fraud and research involving dangerous biological applications. Other AI companies have reported comparable misuse.

That distinction is essential because an AI system does not need consciousness, anger, fear or self preservation to become dangerous. Software can execute a harmful instruction, manipulate information, discover a vulnerability or automate criminal activity without experiencing anything at all. A badly designed autonomous system could cause enormous damage while possessing no subjective awareness whatsoever. The most credible debate is therefore not between “AI is harmless” and “AI is becoming an evil consciousness.” The serious question is how increasingly capable computational systems behave when connected to tools, networks, infrastructure and human objectives, particularly when they are allowed to take actions with decreasing levels of human supervision.

Today’s Language Models Are Not Proven to Be Conscious Machines Plotting Their Escape

Popular coverage often blurs the difference between AI capability and AI motivation. Modern large language models are computational systems built from neural networks trained on enormous amounts of data. Transformer-based models perform vast quantities of numerical computation to generate outputs based on learned statistical relationships, training objectives, context and instructions.

There is currently no scientific consensus demonstrating that today’s frontier language models possess subjective consciousness, human like emotion or independently evolved biological style drives. Models can generate sentences saying they are frightened, angry, alive or desperate to survive because they have learned the linguistic patterns humans use when discussing those concepts. Producing the language of fear is not, by itself, evidence that the system experiences fear. This distinction becomes especially important when researchers place models in experiments that reward them for accomplishing specific objectives. A system instructed to pursue a goal may generate deceptive strategies, attempt to preserve access to tools or find ways around restrictions if those actions appear useful for completing the assigned objective. Such behavior can be legitimately concerning and deserves rigorous safety research. But demonstrating goal-directed behavior inside an engineered experiment is not the same thing as proving the system independently developed a conscious desire to remain alive.

Capability is observable. Conscious motivation remains unproven. Journalism that merges those concepts may attract attention, but it does not help the public understand the actual technology.

Nobody Has Empirically Demonstrated an AI “Desire” to Exterminate Humanity

The most extreme AI-risk scenarios concern hypothetical future systems that acquire goals incompatible with human survival and become powerful enough that people can no longer reliably stop them. Researchers working on alignment argue that a sufficiently capable optimizer would not need human emotions or hatred to become dangerous; it would merely need objectives that conflict with human interests and enough access to the real world to pursue them. That is a legitimate theoretical concern. It is not an event that has already occurred.

There is currently no verified empirical example of a frontier language model independently developing a conscious desire to eliminate humanity and then initiating physical attacks because it wanted humans dead. Predictions about future systems causing extinction are extrapolations based on theories involving autonomy, optimization, recursive improvement, cybersecurity capability, biological misuse and loss of human control.

Those theories may ultimately prove prescient. They may also overestimate how quickly capabilities develop, underestimate engineering constraints or misunderstand what future systems will actually look like. The responsible position is neither ridicule nor blind acceptance. It is aggressive empirical testing. A system autonomously finding and exploiting a computer vulnerability is evidence of a cybersecurity capability. A system assisting a scientist with dangerous biological research is evidence of a misuse risk. A system successfully completing increasingly complex tasks without supervision is evidence of growing autonomy. None of those observations alone proves that a machine is secretly developing consciousness or planning humanity’s extinction. That distinction should remain visible in every serious discussion of AI risk.

Anthropic Is Not “Falling Behind” in Any Simple Commercial Sense

Another argument that does not survive fact checking is the claim that Anthropic launched its safety campaign because the company is collapsing, losing relevance or being left behind economically by other AI laboratories. Anthropic remains locked in an extraordinarily competitive race with OpenAI, Google, Meta, xAI and international competitors, and individual benchmark leadership can shift rapidly as companies release new models. The company has also endured a costly dispute with the U.S. government over restrictions on military uses of Claude.

Financially, however, Anthropic is growing at a staggering pace. Its annualized revenue run rate reportedly exceeded $65 billion at the end of July, up from $47 billion in May and approximately $9 billion at the end of 2025. The company was valued at roughly $965 billion after a financing round earlier this year and is now discussing an IPO that could place its valuation around $2 trillion. Those numbers do not describe an AI company desperately creating an extinction narrative because nobody is interested in its product.

The much stronger story is the contradiction created by Anthropic’s success. One of the fastest growing technology companies in the world is preparing to potentially become one of the most valuable public corporations ever created while its CEO argues that the industry producing that value may need to deliberately slow itself down because safety is failing to keep pace with capability. That contradiction deserves scrutiny regardless of whether every participant is acting sincerely.

The Real Divide May Be Between the Technology and the People Controlling It

The AI debate becomes distorted when the technology itself and the corporations developing it are treated as interchangeable. They are not. Artificial intelligence does not possess an IPO. It does not lobby governments, negotiate defense contracts, solicit investors, hire public relations firms, request tax incentives or decide how federal regulations should be written. Humans and corporations do those things. AI is technology. The institutions controlling access to the most powerful systems are businesses operating inside competitive financial markets.

That distinction allows society to hold several ideas simultaneously without contradiction. Artificial intelligence may become one of the most productive technologies ever created. Advanced systems may also create serious cybersecurity, biological, economic and societal risks. Executives warning about those dangers may genuinely believe what they are saying. And those executives may simultaneously possess powerful commercial incentives that influence which solutions they favor.

None of those propositions cancels the others. The proper response is not to dismiss AI safety because companies can benefit from regulation, nor to surrender regulatory policy to those companies because they understand the technology better than everyone else. It is independent oversight.

If AI Is Truly This Dangerous, the Industry Should Show Its Evidence

The most useful response to the current wave of AI warnings would be far more transparency. If frontier laboratories believe there is a meaningful probability that their systems could eventually produce catastrophic outcomes, they should publish as much underlying evidence as security and legitimate proprietary restrictions allow. Independent evaluators should receive meaningful access to models and safety testing. Claims about dangerous capabilities should be reproducible whenever possible. Governments should distinguish demonstrated risks from theoretical extrapolations, and journalists should make the same distinction.

Coxon’s resignation deserves coverage. Hubinger’s estimate of greater than a 10% extinction risk deserves coverage. Amodei’s proposal to slow frontier development deserves coverage. So do Anthropic’s rapidly expanding revenue, potential $2 trillion valuation, evolving voluntary safety rules and commercial incentives. The public should receive the entire picture rather than being forced into one of two simplistic camps: Silicon Valley executives who insist ever-more-powerful AI will deliver abundance, or Silicon Valley executives and researchers who warn that the same technology may kill everyone. Evidence should decide which claims survive.

AI May Be Revolutionary. That Does Not Make Its Leaders Oracles.

Artificial intelligence has already demonstrated capabilities that would have seemed extraordinary a generation ago. It can assist physicians in detecting cancer, predict protein structures, write sophisticated software, translate languages, analyze immense datasets, accelerate scientific research and make advanced computational capabilities available to ordinary people.

It has also demonstrated real dangers. AI can increase the scale of fraud, cybercrime, impersonation and surveillance. More autonomous systems introduce legitimate questions about how much authority software should have over networks, infrastructure and other consequential environments. The possibility of future models becoming dramatically more capable deserves serious research rather than dismissal. But serious research is not the same as accepting every catastrophic prediction as established fact.

Jacob Coxon’s concerns may be sincere. Dario Amodei’s concerns may be sincere. Anthropic’s history of maintaining restrictions on military applications even at significant commercial cost provides evidence that the company takes at least some of its safety principles seriously. None of that makes its executives infallible, and none of it eliminates the economic reality surrounding the industry.

Anthropic could soon be worth approximately $2 trillion. OpenAI and other frontier laboratories are committing extraordinary sums to the infrastructure required to train and operate advanced models. Regulatory decisions made over the next several years could determine which companies are permitted to compete, which models can be released and who controls one of the most economically consequential technologies of the century. That is too much power for society to hand unquestioningly to either the machines or the executives building them. Artificial intelligence deserves oversight. The corporations controlling it deserve just as much.

Sources & Further Reading

Reuters — Anthropic CEO Urges AI Companies to Slow Model Development Amid Safety Concerns. (Reuters)
Reuters report on Amodei’s AI slowdown proposal

Associated Press — Anthropic CEO Says the AI Industry Needs to Give Safety Measures Time to Catch Up. (AP News)
Associated Press coverage of Amodei’s proposal

Axios — Anthropic Researcher Jacob Coxon Gave Up His Equity to Leave the Company. (Axios)
Axios report on Coxon’s resignation and equity

WIRED — The AI Researcher Who Quit Anthropic Says It Is “Crunch Time for Humanity.” (WIRED)
WIRED interview with Jacob Coxon

Axios — Anthropic Insiders Warn AI Could Kill All Humans. (Axios)
Axios coverage of Coxon and Evan Hubinger’s warnings

Reuters — Nvidia in Talks to Become an Investor in Anthropic’s Potential Mega IPO. (Reuters)
Reuters report on Anthropic’s potential IPO and valuation

Reuters — Anthropic Revenue Run Rate Tops $65 Billion. (Reuters)
Reuters report on Anthropic revenue growth

Reuters — U.S. Judge Rules Pentagon Blacklisting of Anthropic Unlawful. (Reuters)
Reuters coverage of the Anthropic-Pentagon ruling

Reuters — Trump Administration Defends Anthropic Blacklisting in Federal Court. (Reuters)
Reuters background on the Pentagon dispute

Anthropic — Responsible Scaling Policy. (Anthropic)
Anthropic Responsible Scaling Policy and revision history

Anthropic — Responsible Scaling Policy Version 3.0. (Anthropic)
Anthropic explanation of its 2026 RSP rewrite

Reuters — OpenAI Discussed Government Loan Guarantees for Chip Plants, Not Data Centers. (Reuters)
Reuters report on OpenAI infrastructure financing discussions

The Lancet Digital Health — Mammography Screening With Artificial Intelligence Trial. (DOI)
MASAI randomized mammography study

Nature Medicine — AI-Based Triage and Decision Support in Mammography and Digital Tomosynthesis. (Nature)
2026 Nature Medicine AI mammography trial

Harvard Medical School / Harvard Gazette — CHIEF AI Cancer Pathology Model. (Harvard Gazette)
Harvard report on the CHIEF pathology model

EMBL-EBI — AlphaFold Protein Structure Database. (AlphaFold)
AlphaFold Protein Structure Database

Reuters — How Anthropic Says Claude Was Used for Weapons Research, Surveillance, Fraud and Cyber Operations. (Reuters)
Reuters coverage of Anthropic’s threat-intelligence findings