Ripples in the AI Industry! Anthropic CEO Discusses the Dark Side of Artificial Intelligence

Why a Company Aiming for "Safe AI" Has Made Enemies

"AI can be dangerous. Therefore, it should be developed cautiously."

Many would agree with this statement. However, what if it were said by the head of a rapidly growing company developing one of the world's most advanced AI technologies?

While some may see it as a responsible warning, others might suspect it as an attempt to hinder competitors' growth.

The conflict surrounding Anthropic, the company behind the generative AI "Claude," and its CEO, Dario Amodei, embodies such complexities.

An article published by The New York Times on November 13, 2026, reported how Amodei's warnings have evolved from causing irritation among peers to provoking strong opposition. The underlying issues are not just differences in opinions about AI safety but also involve technological competition, relationships with clients, and control over a massive market.

For Japanese users, this is not just a distant quarrel among foreign executives. It concerns the pricing, options, and conditions for continuing to use AI in their work.

Anthropic Did Not Join the Letter Supporting Disclosure

According to the NYT, one of the scenes where the conflict surfaced was a policy debate over the disclosure of AI models.

The paper reported that Anthropic and OpenAI had lobbied the White House to block the development and widespread sharing of certain AI models deemed too dangerous.

Meanwhile, Satya Nadella of Microsoft and Jensen Huang, then CEO of NVIDIA, were preparing a letter supporting the disclosure of models. Many companies agreed, and OpenAI also signed the letter, but Amodei did not participate.

It is important to distinguish between the lobbying described in the report and Anthropic's own stated position.

In an official statement on July 27, the company explicitly stated that it had never advocated for banning open-weight models. They also acknowledged the value of public models that do not possess dangerous capabilities, as they can benefit companies, developers, and researchers.

What they are advocating for is safety testing for sufficiently high-performance models, including both public and private ones.

Therefore, framing the debate as "companies supporting disclosure vs. Anthropic advocating for a total ban" would miss the actual points of contention.

What Does "Open Weights" Mean?

Open weights refer to the method of disclosing the numerical values that determine the behavior of a trained AI. Users can obtain the model and run or adjust it in their own environment according to certain conditions.

However, even if the weights are disclosed, it does not necessarily mean that all training data and development processes are open. It is premature to assume that everything can be freely used and fully understood just because it is labeled "open."

The letter from companies supporting disclosure emphasizes the importance of allowing users to operate AI according to their own needs without relying on a single company.

Consider a system handling internal documents in Japanese, for example.

Using external AI services can reduce the burden of implementation and operation. On the other hand, running the model in one's own environment makes it easier to manage data storage and update timing, though it also requires taking on the burden of equipment and maintenance.

It's not about which is always superior. The significance of public models lies in allowing users to choose based on their needs and budget.

Talking About the Same "Safety" but Seeing Different Dangers

Interestingly, those supporting disclosure are not declaring a disregard for safety.

The coalition's letter argues that widely verified public models make it easier to identify problems and provide powerful tools to those working on defenses. Concentrating capabilities in a few private models could itself become a vulnerability.

In contrast, Anthropic is concerned about the difficulty of retrieving weights after disclosure and monitoring usage to prevent misuse.

When the dangers prioritized by each side are compared, the differences become apparent.

Those emphasizing disclosure see the danger in technology being concentrated in a few companies, making external verification impossible. Those advocating cautious disclosure see the danger in powerful capabilities becoming difficult to control once widespread.

Thus, it's not that what they want to protect is completely opposite. The difference lies in how to achieve safety and which dangers to prioritize.

Why Warnings About Safety Raise Business Suspicions

The NYT does not explain the backlash against Anthropic solely by its statements on safety.

The published portion of the article reports that while the company is rapidly growing, it has also cut off services to some customers and competitors and opposed free alternative services from other companies.

The detailed circumstances and individual justifications for these actions cannot be judged from the excerpts available. However, the article suggests a gap perceived by others between the image of a company "doing the right thing" and its tough stance in competitive environments.

This is an issue not limited to AI companies.

Strict safety standards may be necessary to protect users. At the same time, if compliance incurs significant costs, companies with ample resources and personnel may gain an advantage.

However, it is not appropriate to conclude that "all concerns about safety are just a pretext." It is possible for concerns to be genuine while also aligning with a company's commercial interests.

The same perspective should be applied to companies supporting disclosure. As the AI market expands, it could lead to increased demand for semiconductors and cloud services. The ideals of disclosure and business interests can coexist.

When reading corporate claims, it is more useful to consider who benefits and who bears the burden of the proposed rules, rather than guessing whether they are made in good or bad faith.

Voices Advocating for Disclosure on Social Media—But from Industry Insiders

This debate has also spread on social media, with executives themselves expressing their positions.

The following is a summary of X posts related to the public letter from July, based on reports that included the content of the posts. It does not represent direct reactions to the November 13 NYT article or the opinions of the general public.

Jensen Huang—Both Open and Closed Models Are Necessary

Huang posted that open models support safety, cybersecurity, technology dissemination, and voluntary operation.

Notably, he does not dismiss the need for private models. His stance is that both cutting-edge open and private models are necessary, avoiding a simple binary choice.

Sam Altman—Supporting U.S. Success in Both Approaches

Sam Altman of OpenAI reportedly expressed support for the U.S. succeeding with both open and proprietary private models, showing support for the letter.

This statement indicates that the positions of model-providing companies do not neatly divide into "companies that disclose" and "companies that close." Companies evaluate public models from both business and policy perspectives.

Brian Armstrong—The Value of Tailoring to Business Needs

Brian Armstrong of Coinbase supported strong U.S. public models and mentioned efforts to transition his company's processes to such models.

The reason he cited was the ability to adjust models to fit business operations.

This is a point that Japanese companies can easily understand. The AI that scores the highest in general performance evaluations does not necessarily align with the AI that offers the best cost-effectiveness for a company's work.

However, these are all statements from stakeholders with interests in the AI market. The prominent voices of executives on social media cannot be read as the consensus of society.

The Question of "Who Decides" Returns to Japanese Users

For those using AI in Japan, this conflict has concrete implications in terms of service selection and operation.

For example, what happens if you deeply integrate your operations with one AI service and then the price or usage conditions change? How much effort would it take to switch to another model?

Conversely, if you operate a public model in-house, who will handle updates and vulnerabilities? Obtaining a model and continuing to use it safely are separate challenges.

Thus, the question is not only about "which company is right."

Considerations include pricing, performance, data management methods, adaptation to changes, and the possibility of switching to other services.

The same applies to policy. If safety testing is required, what is measured, who evaluates it, and how much of the results are disclosed are important. If disclosure is encouraged, how responsibility for misuse or accidents is shared must be addressed.

Explanations like "trust us because it's safe" or "it's open, so it's secure" are not enough for users to make informed decisions.

To Avoid Letting Companies' Self-Assessments Decide AI's Future

The backlash against Anthropic may involve serious differences of opinion on safety, as well as wariness of a rapidly growing competitor. However, one should not conclude on motives that cannot be confirmed externally.

What matters is not whether Amodei is liked or how many companies signed the letter.

The key is whether claims of safety can lead to independent verification, whether the benefits of disclosure can be linked to user choices rather than just corporate interests, and who will be responsible for explanations and responses when significant issues arise.

The walls surrounding AI have a role in mitigating risks. However, those walls could also narrow user choices and external verification.

What's needed is not just calls to raise or remove walls but concrete discussions on where they are placed, for what purpose, and who verifies their validity.

Sources and References

1. The New York Times: On industry backlash against Anthropic, inter-company conflicts over public models, and criticisms of business practices.
   https://www.nytimes.com/2026/11/13/technology/anthropic-dario-amodei-ai-openai.html

2. Anthropic Official Statement: Position denying a total ban on public models and proposing safety tests for high-performance public and private models, as of July 27, 2026.
   https://www.anthropic.com/news/position-open-weights-modelsanthropic.com

3. "Open Weights and American AI Leadership": A public letter from a coalition of companies discussing the benefits of public models for competition, user control, and safety.
   https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdfimages.nvidia.com

4. AI Frontier: Article covering the signing of the public letter and X posts by Huang, Altman, and Armstrong. The summary of SNS posts was confirmed through this article; the original posts were not directly obtained.
   https://aifrontier.kr/en/articles/open-weights-letter-2026/