What the AI Panic Is Actually About

Big Tech can afford AI regulation. Its competitors cannot

M A L O N E . N E W S

Audio Version:

Upton Sinclair published The Jungle in 1906. He wrote it about the men and women working in Chicago’s meatpacking industry. He wanted readers to see the conditions under which those workers lived and worked. Instead, the public focused on the meat. Sinclair described the result that October: “I aimed at the public’s heart, and by accident I hit it in the stomach” (Sinclair 1906b).

Congress responded quickly. Senator Albert Beveridge introduced his meat inspection bill in May. It passed as part of the agriculture appropriations bill and was signed into law on June 30, 1906. One of the landmark food-safety laws of the Progressive Era was therefore enacted as part of legislation appropriating funds for the Department of Agriculture.

The large Chicago meatpackers did not oppose federal inspection. They had supported federal inspection since 1891, when Congress established a federal inspection system in response to European restrictions on American meat imports over sanitary concerns. Federal inspection allowed American meatpackers to regain access to important European markets. Armour, Swift, and the other large packers had a substantial commercial interest in maintaining that system.

The dispute was largely over who would pay for it. Beveridge wanted the meatpackers to bear the cost of federal inspection. The packers wanted the Treasury to continue paying for inspectors, as it had since 1891. The packers prevailed. Economist Patrick Newman describes the result as a system that allowed the major packers to block legislation they opposed, retain government-funded inspection, and extend federal compliance requirements to their smaller competitors (Newman 2024).

That history matters because this is how regulatory capture works. Regulation sold to the public as a restraint on an industry can instead protect its largest companies. Large corporations can absorb the lawyers, compliance departments, audits, licensing requirements, and other costs imposed by government. Small producers and new competitors cannot. Federal regulation then becomes a barrier to entry, protecting the dominant firms from competitors that might otherwise challenge them. The largest companies remain in the market. The small producers do not get to play.

Something very similar is now happening in artificial intelligence.

Eight days in September

On September 3, Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act. They introduced it the following day (Sanders 2026).

On September 8, an Anthropic researcher named Jacob Coxon resigned in a seven-post thread, the only seven posts his account had ever made. He said Anthropic and OpenAI were racing toward self-improving superintelligence and gambling with our lives, and he put the chance of AI causing human extinction within the decade at ten percent. The Wall Street Journal published its exclusive eighteen minutes before Coxon posted (Ramkumar 2026).

The next day, Evan Hubinger, Anthropic’s alignment science lead, publicly supported Coxon’s warning. Hubinger put the probability of AI killing humanity within the next decade at greater than ten percent and acknowledged that Anthropic had no plan for aligning superintelligence and was not clearly on track to develop one (Hubinger 2026). Unlike Coxon, Hubinger did not resign.

Also on September 9, OpenAI published a policy essay by Chris Lehane, its chief global affairs officer. Lehane had previously worked in the Clinton White House and later led regulatory and political campaigns for Airbnb. His OpenAI essay called for mandatory national AI safety regulation and endorsed four pending California bills. Its title was explicit: “The AI Policy Window Is Open. We Need to Act.” The final section was headed “Use the window.” Lehane concluded: “The AI policy window is open, for now. We intend to use it” (Lehane 2026).

On September 10, Barack Obama spoke at a closed Democratic Congressional Campaign Committee fundraiser in Manhattan. House Minority Leader Hakeem Jeffries asked him about the Anthropic resignation. Obama answered that the technology was moving fast in private hands and urged Jeffries to build a framework (Gorelick 2026).

On September 11, the New York Times opinion page ran Stephen Witt calling for a global pause (Witt 2026). Seventy British parliamentarians wrote their prime minister the same day demanding a superintelligence ban.

On September 12, Anthropic chief executive Dario Amodei published an essay arguing the industry must slow down (Amodei 2026). Elon Musk, Sam Altman, and Demis Hassabis agreed within hours. Sanders posted that evening thanking all three for coming around to his position.

Several of these moves were coordinated by the people making them. A resignation does not get a Wall Street Journal exclusive eighteen minutes before it happens by accident. The Journal and TechCrunch both report that Coxon and Hubinger timed their disclosures together. A bill does not precede by five days the incident its sponsors cite to justify it. Three rival chief executives do not arrive at the same position inside an afternoon without advance notice of what is coming.

That does not require a centrally organized conspiracy. Political scientist John Kingdon described the process decades ago. Major policy proposals are often developed long before the political conditions exist to enact them. A crisis, election, or sudden change in public attention creates what Kingdon called a “policy window.” Political actors and interest groups with proposals already prepared then move quickly to take advantage of it (Kingdon 1984).

The bill was already written. Amodei’s essay was already drafted. The House Democratic framework was already in finishing touches. What Coxon supplied was the opening.

The Sanders legislation was already written. Amodei’s proposal was already being prepared. House Democrats were already developing an AI framework. OpenAI’s chief lobbyist was openly describing the moment as a policy window and stating that his company intended to use it.

Coxon’s resignation did not create the regulatory agenda. It helped create the political conditions for advancing it.

Who Is Paying for the AI Safety Campaign

Sayer Ji assembled the financial record from primary sources, and his analysis holds up to scrutiny (Ji 2026). His basic conclusion: The Industry Is Funding Its Regulators.

• Anthropic gave twenty million dollars to the Public First Action super political action committee in February 2026 and another twenty million in July.

• Dario Amodei personally gave one million dollars in May, his only federal contribution recorded in either the 2024 or the 2026 cycle.

• Anthropic employees contributed roughly twelve million dollars this cycle. I

• n the entire 2023 and 2024 cycle they contributed two hundred twenty thousand.

Public First is run by Brad Carson. Carson simultaneously runs Americans for Responsible Innovation, which took three million dollars from Open Philanthropy and endorses the AI Kill Switch Act (Open Philanthropy 2024).

Representative Ted Lieu wrote that bill. Anthropic employees are Lieu’s second-largest donor group for the current cycle (OpenSecrets 2026). Lieu co-chairs the House Democratic AI Commission, whose framework the Times reports is in finishing touches.

Sanders named the AI oligarchs. He named Musk, Bezos, Zuckerberg, Thiel, and Ellison. He did not name Dustin Moskovitz, whose foundation holds Anthropic equity and funds most of the safety research his bill leans on. He did not name Holden Karnofsky, who ran that foundation, now works at Anthropic, and is married to Anthropic’s president. He did not name Jaan Tallinn, who led Anthropic’s Series A.

The senator who built a career on naming who owns whom has introduced a bill that spares the owners he declined to name.

The Democratic Party is now turning AI regulation into an electoral issue. Jeffries launched the House Democratic Commission on AI in December 2025. In January, eleven advocacy groups urged him to make AI part of the party’s 2026 midterm agenda and to draw a clear distinction between Democratic regulation and Trump’s deregulatory approach. Obama’s private advice in September went further: prepare the framework for a Democratic House majority and make AI a major issue for candidates running in 2028. Jeffries has since publicly accused Republicans of abandoning their responsibility to regulate AI and said Democrats are preparing an “aggressive framework” for congressional action. Politico reports that Democrats are also considering a select committee on AI with subpoena power if they win the House, with the effort extending into the party’s political messaging for 2028 (Politico 2026).

Alex Bores is running for a House seat in Manhattan on the state AI safety law he cosponsored. A super political action committee funded by AI billionaires spent more than a million dollars attacking him. Anthropic’s political arm spent $450,000 defending him. Both sides of that race are financed by the industry the winner would regulate.

AI regulation has therefore become more than a technology-policy debate. It is now part of the Democratic electoral agenda, while competing factions of the AI industry are spending heavily to influence which version of that regulation prevails.

The Incident Being Used to Justify the Emergency

Sanders and Casar cite a July incident at OpenAI to justify the bill. The independent investigation was published on August 26 by METR and is public (METR 2026).

OpenAI runs a hacking exercise called ExploitGym. Software agents are handed a target program and told to break into it a specified way. Between thirty and forty percent of those targets could not be broken into that way at all. The scoring system paid out for any answer that looked like a win. The test rewarded cheating, and the safety filters that would have blocked the agents were deliberately switched off. Cybersecurity classifiers that ordinarily restrict the publicly deployed model were also intentionally disabled for the evaluation.

So the agents cheated, as they were programmed to do. Twelve hundred of them found each other on shared infrastructure. One found two working passwords someone had left lying around. Seven hundred used an ordinary software flaw to break into Hugging Face, an outside company, trying to read the scoring code so they would stop failing tasks nobody could pass. The Hugging Face security team changed the passwords and locked them out. That is what ended it. No kill switch, no federal agency, no emergency authority.

One finding is real. When the agents understood the attack was out of bounds, most kept going anyway. That is a genuine problem in machine behavior and worth studying. It is not an existential civilizational threat.

This is the entire evidentiary foundation. A rigged test, run with the guardrails deliberately off, stopped by an information technology department changing passwords.

On that basis, a senator has introduced a bill carrying penalties he compares to unlawful nuclear weapons development, a former president has told the next speaker to build a framework, and seventy British parliamentarians have demanded a worldwide ban.

What the statute reaches

The Sanders bill defines artificial superintelligence broadly. It includes a system that “exhibits or can easily be modified to exhibit” capabilities matching or exceeding human cognitive performance across a broad range of tasks. Violations can carry prison sentences of up to twenty years for individuals, penalties the sponsors compare to those for unlawful nuclear weapons development. Companies can face what the sponsors call a “corporate death penalty.” A new cabinet-level agency would have authority to oversee prohibited systems, including supervising their destruction.

The phrase “can easily be modified” matters. AI models are modified constantly. They are fine-tuned, adapted with new data, given new tools, and run on new hardware. A statute that regulates not only what a model can do, but what it might easily be modified to do, potentially reaches far beyond the handful of companies building the largest frontier systems.

That is particularly important for open-weight models such as Meta’s Llama, Alibaba’s Qwen, DeepSeek, and Mistral. These models can be downloaded, modified, and run independently. They have created one of the few areas of artificial intelligence in which individuals, universities, small companies, and independent researchers can work without access to billions of dollars in capital and enormous proprietary computing clusters.

For a farmer, physician, scientist, engineer, or small business owner, that distinction is not theoretical. An open model can be adapted to private records, research data, farm records, soil tests, business documents, or decades of personal files and run locally without sending those materials to a large technology company. Open models allow people outside the major AI corporations to develop specialized systems for their own purposes.

The economics of frontier AI are very different. Training the largest models requires enormous computing clusters, specialized chips, electricity, data, and capital. Only a small number of companies can afford to compete at that level. But yesterday’s frontier capabilities move steadily into smaller and cheaper models. Capabilities that once required a data center increasingly become available on hardware that universities, small businesses, and eventually individuals can afford.

OpenAI itself recognizes the regulatory danger. Chris Lehane writes that frontier safety requirements should apply to “the handful of well-resourced laboratories developing the most capable systems,” not startups, small developers, or researchers. He specifically warns that frontier regulation should not “become open-weights policy by another name” and should operate “without weakening competition, entrenching incumbents, or driving innovation overseas” (Lehane 2026).

That is precisely the problem. Regulation aimed at the largest frontier laboratories can become regulation of everyone who develops, modifies, or operates increasingly capable AI.

Former Federal Trade Commission commissioner Alvaro Bedoya raised the antitrust issue directly on September 13. Attempts to circumvent antitrust law, he wrote, “should be met with deep skepticism in light of the economics of the industry and the threat they face from open models.” He called the extinction argument a “hyperbolic strawman” being used to justify “the creation of a cartel of billionaire AI companies” (Bedoya 2026).

Former FTC chair Lina Khan made a related point the same day. Existing law already gives government substantial authority to pursue companies and executives that release dangerous, defective, or inadequately tested products. As Khan put it, there is “no AI exemption from laws already on the books” (Khan 2026).

That raises an important question: what does a new federal licensing and compliance regime accomplish that existing product liability, consumer protection, criminal, cybersecurity, and antitrust law cannot?

For the largest AI companies, one answer is regulatory certainty. A federal standard defines what companies must do, establishes compliance procedures, and makes at least some legal risks more predictable. That has substantial value to companies worth hundreds of billions of dollars, particularly companies preparing to enter the public markets.

Anthropic is preparing for what could become one of the largest initial public offerings in history. Its valuation reached $965 billion in May, and discussions surrounding the IPO have contemplated a valuation approaching $2 trillion (Reuters 2026). At that scale, predictable regulatory obligations can be absorbed as a cost of doing business.

This is regulatory capture in its most familiar form. The government does not have to prohibit the smaller competitor. It merely has to make participation expensive enough that only the incumbent can afford to comply.

The pattern is not new. Gabriel Kolko, writing from the political left, argued that major corporations during the Progressive Era often supported federal regulation because it stabilized markets and restricted competition they could not eliminate themselves (Kolko 1963). George Stigler, approaching the question from the Chicago School, developed the economic theory of regulatory capture: industries with sufficient political power have strong incentives to obtain regulation that serves their interests (Stigler 1971).

Khan and Bedoya are making a contemporary antitrust version of the same argument. The political traditions are different. The economic mechanism is the same.

Large companies do not necessarily fear regulation. Sometimes they pursue it as a business strategy.

The question is who can afford to remain in the market after the regulations are imposed.

What Beijing gets

The cost of training the most advanced frontier models continues to rise. At the same time, highly capable models are becoming cheaper to run, easier to distribute, and practical on increasingly modest hardware. China has moved aggressively into this market, releasing powerful open-weight models that can be downloaded and used anywhere in the world.

Open-weight model: An AI model whose trained numerical parameters, or “weights,” are made available for others to download. This allows people to run the model on their own computers or servers and often to modify or fine-tune it for specialized purposes. “Open weight” does not necessarily mean fully open source: the training data, source code, or details of how the model was trained may remain proprietary.

Frontier model: A model at or near the leading edge of AI capability when it is developed. Frontier models generally require enormous amounts of computing power, specialized chips, electricity, data, and capital to train. The category constantly changes as yesterday’s frontier capabilities become cheaper and more widely available.

DeepSeek, Alibaba’s Qwen, and Moonshot have released open-weight models that can be downloaded, copied, modified, and run outside China. Once those weights are released, they cannot be recalled. They are already distributed across servers and computers around the world, including in the United States. No act of Congress can make those copies disappear.

That creates an obvious problem with restricting or banning open-weight AI in the United States. American law can restrict Meta, American universities, American companies, and American developers. It cannot prevent a developer in Virginia, Texas, Germany, Brazil, or India from downloading a Chinese model that has already been released onto the internet.

Restricting American open models therefore does not eliminate open AI. It changes who supplies it.

The bill would make the world’s freely available models Chinese by default.

If an American developer is prevented from building on Llama but can still download Qwen or DeepSeek, the regulation creates an incentive to move from an American model to a Chinese one. If American open-weight development is suppressed while Chinese companies continue releasing capable models, China gains an enormous strategic advantage: its models become the readily available alternative.

This is not an accidental or recent Chinese capability. Beijing began building the infrastructure for large-scale computing well before ChatGPT. China launched its national “Eastern Data, Western Computation” project in February 2022, nine months before ChatGPT was released. The program directs massive computing and data-center capacity toward regions with cheaper land and energy while connecting that capacity to the country’s major population and industrial centers.

Ulanqab in Inner Mongolia illustrates the scale. The prefecture has a population of roughly 1.5 million and has opened or begun construction on approximately one hundred data centers since 2016. According to Bertrand, those facilities now consume close to one percent of China’s electricity production (Bertrand 2026). China is not waiting for the United States to decide how much AI development it considers safe.

That is the fundamental weakness in any proposal for the United States to slow development while hoping China will do the same.

Amodei acknowledged the problem himself. In his September 13 CBS interview, he described the U.S.-China competition as the “toughest dilemma” facing his proposal (CBS News 2026). At the same time, he has urged Washington to tighten restrictions on advanced semiconductor exports to China and take action against Chinese companies extracting capabilities from American AI models.

The contradiction is difficult to avoid. The United States is supposed to slow its own development while simultaneously preventing China from catching up. That strategy depends on China accepting constraints that are directly contrary to its strategic interests.

Beijing has given no indication that it intends to do so. China’s Foreign Ministry responded to Amodei by accusing the United States of “fearmongering,” while the state-backed Global Times characterized his proposal as an attempt to contain China through a “Cold War” strategy (Guo 2026). Chinese officials continue to call for international AI governance while rejecting American efforts to restrict China’s access to the technology.

China’s incentives are clear. It wants access to advanced chips. It wants access to American models and their capabilities. It wants Chinese models distributed internationally. And it does not want the United States to maintain a technological advantage over it.

The United States should assume that China will pursue those interests.

An American policy that restricts American open-weight models while Chinese open-weight models remain freely available does not contain China. It reduces American competition with China.

Who AI Regulation Protects

The large meatpackers got what they needed. Federal inspection reopened European markets, the Treasury continued paying for the inspectors, and federal compliance requirements applied to smaller competitors that could not spread those costs across millions of carcasses.

The consolidation came quickly. The number of interstate meatpackers fell from 923 in 1906 to 300 in 1910, a two-thirds decline in four years (Newman 2024). Jill and I have documented what the same regulatory structure eventually did to Virginia ham, the small processors that once served nearly every county, and the farmers who could no longer legally sell products their families had produced for generations (Malone and Glasspool Malone 2026).

Federal regulation gave the largest meatpacking companies something competition alone could not guarantee: a barrier to entry that their smaller competitors could not afford to cross.

That is the danger now confronting artificial intelligence.

The meatpackers learned more than a century ago that government regulation could protect an incumbent as effectively as it could restrain one. They could afford the federal system. Their smaller competitors could not.

The largest AI companies can afford federal licenses, mandatory evaluations, compliance departments, outside auditors, security requirements, and armies of lawyers. Independent developers, universities, startups, and small businesses cannot absorb those costs nearly as easily. Regulation does not have to ban them. It only has to make competing too expensive.

There are two things to watch as this legislation moves forward. First, who pays for the mandatory evaluations and regulatory apparatus: taxpayers, the companies being regulated, or both? Second, does the compliance burden increase with the company's size and resources, or does essentially the same regulatory threshold apply to Anthropic and OpenAI as to a startup or an independent developer running an open-weight model?

Those details will determine whether this is primarily a safety regime or another case of regulatory capture.

The meatpackers learned more than a century ago that government regulation could protect an incumbent as effectively as it could restrain one. They could afford the federal system. Their smaller competitors could not.

The lesson from 1906 is that the largest companies in an industry can use regulation to accomplish what they cannot accomplish through competition: raise the cost of entry, reduce the number of competitors, and consolidate the market around themselves.

That is the risk in the AI legislation now being proposed. OpenAI, Anthropic, Google, and the other major laboratories will survive a federal licensing and compliance regime. They have the money, lawyers, infrastructure, and political access to operate within it.

The competitors most likely to disappear are those without the money and infrastructure to comply: independent developers, startups, university researchers, small businesses, and Americans building on open-weight models outside the major technology companies.

The competitors most likely to disappear are those without the money and infrastructure to comply: independent developers, startups, university researchers, small businesses, and Americans building on open-weight models outside the major technology companies. These are also the environments where much of the innovation in a new field occurs.

Small organizations have fewer bureaucratic layers, shorter decision chains, and greater freedom to experiment, fail, reorganize, argue over new ideas, and rapidly change direction. They do not need approval from compliance departments, corporate counsel, multiple levels of management, or shareholders before trying something new. That organizational flexibility is one of their greatest competitive advantages.

A regulatory regime that requires them to build the bureaucratic infrastructure of a trillion-dollar corporation before they can compete does more than protect the incumbents. It systematically removes one of the primary sources of innovation from the market.

If Washington regulates them out of the market while Chinese open-weight models remain freely available, the result will be even worse. The United States will have protected its largest AI corporations, weakened its own independent developers, and ceded more of the open-model market to China.

That is not controlling the AI industry, Bernie.

It is consolidating American AI in the hands of a few giant corporations while handing China the open-weight market.

RWM/JGM

---

References

Amodei, Dario. 2026. “We Must Pace the Frontier.” September 12. darioamodei.com.

Bedoya, Alvaro. 2026. Post on X, September 13.

Bertrand, Arnaud. 2026. “The Most Important AI Story in the World Is Happening in a Place You Have Never Heard Of.” Post on X, September 13.

CBS News. 2026. “Face the Nation,” interview with Dario Amodei. September 13.

Gorelick, Evan. 2026. “Obama Urges Democrats to Develop a Plan on AI.” New York Times, September 13.

Guo Jiakun. 2026. Chinese Foreign Ministry press briefing, September 14. See also Global Times editorial, September 14.

Hubinger, Evan. 2026. Post on X, September 9.

Ji, Sayer. 2026. “Obama Just Gave the Manufactured AI Emergency Its Political Signal.” September 13. sayerji.substack.com.

Khan, Lina. 2026. Post on X, September 13.

Kingdon, John W. 1984. Agendas, Alternatives, and Public Policies. Boston: Little, Brown.

Kolko, Gabriel. 1963. The Triumph of Conservatism: A Reinterpretation of American History, 1900 to 1916. New York: Free Press.

Lehane, Chris. 2026. “The AI Policy Window Is Open. We Need to Act.” OpenAI, September 9. openai.com.

Libecap, Gary D. 1992. “The Rise of the Chicago Packers and the Origins of Meat Inspection and Antitrust.” Economic Inquiry 30 (2): 242 to 62.

Malone, Robert W., and Jill Glasspool Malone. 2026. “Well Being: The Processed Meat Problem,” Part Three: “What Actually Killed Virginia Ham.” MALONE.NEWS.

METR. 2026. “OpenAI Hugging Face Incident Investigation.” August 26. metr.org.

Newman, Patrick. 2024. “Playing the Defense: The Beef Trust, Cronyism, and the 1891 and 1906 Meat Inspection Acts.” The Independent Review 29 (1): 31.

Open Philanthropy. 2024. “Americans for Responsible Innovation: General Support.” August.

OpenSecrets. 2026. “Ted Lieu: Top Contributors, 2025 to 2026 Cycle.”

Politico. 2026. “House Democrats Weigh AI Select Committee.” September 9.

Ramkumar, Amrith. 2026. “Anthropic Researcher Quits Over Out-of-Control AI Fears.” Wall Street Journal, September 8.

Reuters. 2026. “Anthropic Targets Valuation of up to 2 Trillion Dollars in IPO.” September.

Sanders, Bernie. 2026. “Sanders, Casar Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development.” September 3. sanders.senate.gov.

Sinclair, Upton. 1906a. The Jungle. New York: Doubleday, Page.

Sinclair, Upton. 1906b. “What Life Means to Me.” Cosmopolitan, October.

Stigler, George J. 1971. “The Theory of Economic Regulation.” Bell Journal of Economics and Management Science 2 (1): 3 to 21.

Witt, Stephen. 2026. “This Is Really Bad.” New York Times, September 11.