By Allum BOKHARI
The company has refused to embrace openness and competition.
Anthropic has put itself at odds with nearly the entire AI industry. In a conspicuous break with its peers, the frontier lab declined last month to sign an open letter supporting a permissive policy toward open-weight AI models, which anyone can download, modify, and run on their own devices.
Anthropic says it acted out of concern for public safety, offering its customary warnings about AI-enabled catastrophe. But its real motives may be considerably more self-serving.
If Anthropic were serious about being the AI company with a conscience, it would have no shortage of public concerns to address. It could devote more attention and money to researching AI-related job losses. Or it could take smaller but symbolically important steps. Recent reports revealed that AI firms have been cutting the spines off books to make them easier to scan for model training. Elon Musk pledged that xAI would scan books "the hard way" instead. Anthropic should do the same.
Anthropic could also do more to answer public concerns about data center growth. Some of those fears drift into conspiracy theory. Theo Von , for example, has suggested that data centers are part of a Silicon Valley plot to achieve immortality, monopolize the relevant technology, and lock down everyone else as governments did during Covid. Environmental complaints about water and energy use are also frequently exaggerated. Still, Anthropic could make a more serious effort to address these concerns.
Instead, it is feeding them.
If Americans worry that Silicon Valley elites intend to monopolize advanced technology, few things will reinforce that suspicion faster than Anthropic's campaign against open-weight AI. Nearly every major company in the field has signed the July letter, including OpenAI, Google, Meta, Microsoft, Palantir, Perplexity, Replit, and Hugging Face. Hundreds of companies have signed in all; Anthropic is the most glaring holdout.
The company again frames its position as a matter of safety, warning that open models could enable catastrophic cyberattacks or bioweapon development. To be clear, Anthropic CEO Dario Amodei has said his firm is not "advocating for a ban on open-weights models as a category." But he is lobbying for measures like mandatory safety testing and a crackdown on large-scale distillation (a process by which existing AI models are used to train newer ones-more on that in a moment), which could significantly restrict the development and profusion of open models.
One X user neatly summarized the argument in a post that went viral: "Guns are dangerous, so only I should be allowed to have them." (Conservatives who have followed the Second Amendment debate will recognize the logic.)
This is precisely the wrong way to address the problem. Anthropic's fixation on top-down guardrails is undermining security rather than strengthening it. Last month, the frontier lab had to reverse a change to its leading model, Fable 5, after AI researchers discovered that their work was being covertly throttled. Anthropic's solution was merely to make the throttling transparent. Security researchers must now turn to open models, the very models Anthropic is lobbying against, to conduct their tests.
How seriously does Anthropic take the threat of AI-enabled cyberattacks if its own guardrails prevent cybersecurity researchers from developing defenses against them?
Look beneath Anthropic's rhetoric about safety and a more familiar corporate motive appears: control of the technology.
That motive is becoming harder to miss in the company's campaign against open-weight models, which Anthropic says are being used to distill its products. Distillation sits at the center of the open-weight debate. Most of the technology industry wants to preserve the ability to learn from competing models. Its leaders correctly point to the role open-source software played in driving innovation during the 1990s and 2000s. Anthropic wants a different arrangement.
The reason is not hard to see. Whatever one thinks of its leadership, Anthropic produces cutting-edge models that often have a clear advantage in coding. Blocking competitors from learning from those models could turn a modest, temporary lead into one that is large and lasting.
Thus, we arrive at a familiar monopoly strategy: wall off the technology from competitors while imposing tight guardrails on everyone outside the company. Anthropic would presumably retain unrestricted access to its own products. The public would not. At that point, fears about digital lords and digital serfs begin to sound less fanciful.
It is convenient for Anthropic that the most prominent beneficiaries of distillation are Chinese labs, particularly Moonshot AI and its Kimi K3 model. Moonshot allegedly used distillation to build an open-weight model that approaches the capabilities of products from OpenAI and Anthropic. The White House has raised concerns, describing the practice as part of the same Chinese playbook of intellectual property theft that fueled the country's other technological advances.
But the narrow focus on China obscures an important fact: American companies also use distillation to compete with one another. OpenAI has warned that combining different models could eventually produce systems that outperform any single competitor. Distillation therefore appears to be an important engine of competition and technological progress. Cutting it off in the United States could hand China an advantage.
Imagine if the American technology sector had forbidden open-source software in the 1990s. Would the United States still enjoy its commanding position in software development?
The dispute over Chinese intellectual property theft is real and difficult. But Anthropic's position also reflects a domestic political agenda. Beyond the obvious desire to pull up the ladder behind itself, the company's fixation on safety and guardrails looks more like the approach favored by Democrats than that favored by the current White House.
As I wrote in The American Conservative in May, Democrats have their own reasons to support top-down control of AI in the name of safety. They want to ensure that the technology cannot be used to undermine the civil rights regime. Democratic lawmakers in several states have proposed bills targeting "algorithmic discrimination" and disparate outcomes for women and minorities. It is the familiar progressive demand for equal outcomes, recast for the age of artificial intelligence.
Colorado legislators went so far as to propose allowing "algorithmic discrimination" when it is intended to "increase diversity" or "redress historic discrimination." In plain English, AI-enabled discrimination against white men.
Like Anthropic, Democrats have an interest in preventing the spread of open-weight models. Such models can operate independently of large companies and may therefore evade state-level DEI mandates.
Anthropic's highly publicized clash with the Trump Pentagon over military uses of its technology adds another political dimension. The company's "guardrails-first" position could help it curry favor with a future Democratic administration. Anthropic can market itself as the AI company that stood up to Trump, while its competitors enthusiastically signed up for the so-called "tech right."
Americans should not fall for the performance. The competitive threat from China is real and must be confronted. But there is little reason to believe that restricting open-weight models would help the United States win that competition. It may do the opposite.
Nor is there a compelling case for giving a small circle of technology executives exclusive control over AI, especially when government bureaucrats could easily use that centralized power to impose their ideology on the future. That's a safety risk all of its own-though one you won't find the loudest "safety" proponents talking much about.
Original article: www.theamericanconservative.com