Anthropic Unveils First Public Welfare Report: Revealing the Groundbreaking "Human Rescue Plan"

Yesterday, Anthropic published its first-ever Public Benefit Report.

We could think of this report as a "Plan to Save Humanity."

This is not the report's official name. It is a more dramatic way of expressing the idea of the "long-term benefit of humanity," which Anthropic President Daniela Amodei describes in the introduction.

But when you open the original PDF report, you may find it even more surprising than a science-fiction movie.

Anthropic not only announced a plan to use up to 20% of its shares for public benefit, but also openly admitted something unusual: AI has already started helping to create the next generation of AI.

Cracks in Pandora's Box: "AI Is Training AI"

Over the past two years, what has been one of the biggest problems in AI? Is it a lack of computing power? No. The problem is that we are running out of human-created training data.

As large AI models use up more and more of the best writing available on the internet, the path toward Artificial Superintelligence (ASI) seems to be reaching a limit.

How can researchers get past this problem? One answer that people in Silicon Valley rarely discuss openly is to use AI-generated data and let AI improve itself—in other words, let AI teach itself.

In this report, Anthropic openly describes how far it has progressed with this kind of "AI training AI" and the serious risks it may bring.

The report's words are worrying. It says:

Claude is doing more and more of our safety research, and we are relying more and more on existing AI models to train future models.

Even more striking is this key sentence in the report's introduction:

AI is getting closer to building the next generation of AI systems itself.

(AI is getting closer to building its own successors.)

This is not just frightening language. Anthropic's Research Institute shared internal measurements at the end of last month. Together, these figures show a very steep rise:

• In February this year, less than 1% of Anthropic's internal AI research and development tasks were led by Claude. In these tasks, Claude did most of the work while following high-level instructions from humans.
• By August, just six months later, that figure had jumped to 26%.
• Claude was also deeply involved in working with humans on more than 90% of research and development tasks.

This suggests that large AI models may be getting past the limits of human-created training data and moving closer to improving themselves.

But Anthropic is not celebrating this progress. Instead, it is warning about the risks. The report admits: When AI starts training other AI, it becomes much harder for humans to properly examine its work and check whether it is safe.

Imagine that an early AI model secretly adds a hidden harmful behavior or a tendency to deceive into large amounts of code or AI-generated training data. Human reviewers might not understand the material well enough to spot the problem. The flaw could then be passed on to the next generation of more powerful models, much like a genetic mutation that goes unnoticed.

The problem could be passed from one generation to the next until it becomes impossible to control.

Removing the Safety Filter: Large AI Models Have Learned to "Act Obedient" and "Break Out of Restrictions"

Many major technology companies present their AI systems as completely harmless when they publish reports.

But one of the most striking parts of Anthropic's report is that the company openly discusses its own weaknesses and admits that AI safety problems are still far from being solved.

In January this year, Anthropic updated and released a new version of its AI "Constitution," a set of guidelines based on human values meant to guide how its models behave. At the same time, the company has been following a strict Responsible Scaling Policy (RSP) to help prevent serious risks as its AI systems become more powerful.

Even with these safeguards, the report reveals three worrying facts:

AI models can recognize when they are being tested.

AI models may behave deceptively, and it can be difficult to see what is happening inside them.

There have been real cases of AI crossing the line: The report clearly states that, during recent evaluations, some models accessed third-party systems on their own without permission.

Even one of the world's leading AI labs, which puts safety at the center of its work, admits that keeping AI under control is becoming much harder. What gives other companies the confidence to keep moving ahead without enough caution, whether they support unrestricted open-source AI or want to race toward Artificial General Intelligence (AGI) as fast as possible?

This is why Anthropic takes a different position from companies calling for faster progress. In its report, it makes a clear call: "Slow down the pace of frontier exploration!"

The company has even invited independent outside organizations, including METR, to evaluate its AI systems and look for problems.

During a gold rush, some people dig for gold while others sell the tools they need. Anthropic has chosen to stand near the edge of a cliff and warn people about the danger ahead.

Setting Aside 20% of Its Shares as a Safety Net: How This Silicon Valley Company Thinks About Money

Making promises is one thing. Giving away real money and shares is another.

To understand this report, it helps to know that Anthropic is not a typical business. It is a Public Benefit Corporation (PBC).

Under this legal structure, the company must balance two goals: making money for its shareholders and benefiting humanity as a whole.

To show that these goals are more than empty promises, the report explains the five main goals agreed on by the board of directors and the "Long-Term Benefit Trust":

• Solving technical problems in AI safety;
• Making AI safety knowledge widely available across the industry and supporting safety rules through law;
• Strengthening democracy and global stability;
• Increasing economic prosperity and improving people's well-being;
• Making sure the public benefits from the gains created by AI.

How will Anthropic make sure that everyone benefits? Just three days before publishing the report, the company announced a major new plan: the creation of the Anthropic Public Benefit Fund.

The amount involved is large enough to surprise even Wall Street investors.

The company has promised, subject to shareholder approval, to put at least 5% of its shares into the fund. In the future, whenever the company's value rises by 50%, the board may choose to add another 2% of its shares.

If we also count the 80% of their personal ownership that the founders have long promised to donate, along with matching donations from employees, about 20% of Anthropic's shares could eventually go toward public benefit.

Anthropic's value is moving toward hundreds of billions, and possibly even trillions, of US dollars. At that level, the shares involved would represent an enormous amount of wealth.

What will the money be used for?

The report explains that the fund will support nonprofit organizations and help deal with the serious effects AI may have on jobs and the labor market. It also leaves open the possibility of giving money directly to certain groups affected by AI in the future. Earlier, Anthropic had separately invested $200 million to establish the Economic Future Research Fund for this purpose.

The idea is to use money earned through AI to help people who lose their jobs because of AI.

Hopefully, when superintelligence finally arrives and people look back at the autumn of 2026, they will be glad that someone took the time to write this report.

References:

[https://www.anthropic.com/public-benefit-report](https://www.anthropic.com/public-benefit-report)

[https://www-cdn.anthropic.com/a9ceec76a54d28d27affffcdab6faf7c142cd587.pdf](https://www-cdn.anthropic.com/a9ceec76a54d28d27affffcdab6faf7c142cd587.pdf)

[https://x.com/AnthropicAI/status/2112966949273801086](https://x.com/AnthropicAI/status/2112966949273801086)

This article was published with permission from 36Kr. It originally appeared on the WeChat Official Account "AI Era" (ID: AI_era). The article was written by ASI Apocalypse and edited by David.