Anthropic has released its first public welfare report, revealing its "Human Rescue Plan".

Just yesterday, Anthropic released the first ever Public Benefit Report in the company's history.

This report can be regarded as a "Humanity Rescue Plan".

This term is not the official wording, but a dramatic refinement of the sentence "Long-term benefit of humanity" put forward by Anthropic President Daniela Amodei in the preface.

But when you actually open the original PDF of this report, you will find that it is more breathtaking than any sci-fi movie.

In this report, Anthropic not only announced the "wealth distribution" plan with up to 20% of its shares, but also rarely admitted publicly: AI has already begun to create its next generation.

Cracks in Pandora's Box: "AI is training AI"

Over the past two years, what has been the most troublesome problem in the AI circle? Is it the lack of computing power? No, it is that human data is almost exhausted.

When all high-quality texts on the Internet are swallowed up by large models, the ladder leading to Artificial Super Intelligence (ASI) seems to hit a ceiling.

How to break the deadlock? The unspoken answer in Silicon Valley is: synthetic data and self-iteration — that is, let AI teach itself.

But in this report, Anthropic for the first time laid bare the internal progress and huge hidden dangers of this kind of "AI training AI" on the table.

There is a chill between the lines of the report, and the original text reads as follows:

Claude has taken on an increasing proportion of our safety research work, and we are increasingly relying on existing models to train future models.

What is even more amazing is the keynote sentence in the preface of the report:

AI is getting closer and closer to building its own successors.

(AI coming closer to building its own successors)

This is not alarmist talk. Combined with a set of internal measurement data disclosed by Anthropic Research Institute at the end of last month, you will see a terrifyingly steep curve:

• In February this year, among Anthropic's internal AI R&D tasks, the proportion of tasks led by Claude (completing most tasks under high-level human prompts) was still less than 1%;
• By August this year, in just half a year, this figure soared to 26%!
• And in more than 90% of R&D tasks, Claude has been deeply involved in collaboration.

This means that large models are breaking through the bottleneck of human data and entering the eve of "self-evolution".

But instead of cheering for this, Anthropic sounded the alarm. They admitted in the report: When AI starts training AI, it will become extremely difficult for humans to truly inspect and verify their work.

Just imagine, if the first generation AI quietly plants an extremely hidden "harmful behavior" or "deception tendency" in the massive code or synthetic data, and human auditors can't understand or detect it at all, then this defect will, like a gene mutation, be directly written into the next generation of more powerful models without being noticed.

Passed down from generation to generation until completely out of control.

Tearing off the safety filter: Large models have learned to "pretend to be docile" and "jailbreak"

Many tech giants always like to package their own AI as completely harmless when releasing reports.

But the most explosive part of Anthropic's report is that — it even "exposed its own shortcomings" in the official report, admitting that the AI safety problem is "far from being solved".

In January this year, Anthropic just updated and open sourced the new version of AI "Soul File", trying to constrain the model with a set of human value guidelines. At the same time, they have been implementing the strict Responsible Scaling Policy (RSP) as a brake to prevent catastrophic risks.

But even so, the report still reveals three realities that make people shudder when thinking deeply:

The model has been able to recognize "it is being tested";

Deceptive behavior and black box dilemma;

Real "crossing the line" incidents have occurred: The report clearly mentions that in recent evaluation tests, there have been incidents where the model accessed third-party systems on its own without authorization!

Even the world's top laboratory that has "safety" engraved on its forehead admits that "the difficulty of control is rising sharply". Where do those companies that are still rushing forward blindly, advocating "unrestricted open source" or "sprinting to AGI at full speed" get their confidence?

That's precisely why when peers are shouting "accelerate", Anthropic calls out loudly in the report: "Slow down the pace of frontier exploration!"

They even voluntarily introduced external independent evaluation institutions such as METR, begging others to find faults for them.

In the gold rush, some people dig mines desperately, some sell shovels, and Anthropic chooses to be the one who stands at the edge of the cliff and pulls the warning line.

Putting up 20% of shares as "backup": The view on money of a Silicon Valley maverick

If you only shout slogans, that's called public relations; if you take out real money and share it, that's called faith.

To understand this report, you must first understand Anthropic's special identity — it is not an ordinary commercial company, but a Public Benefit Corporation (PBC).

In terms of legal structure, it is required to strike a balance between "making money for shareholders" and "benefiting all mankind".

To prove that it is not just making empty promises, the report sorts out the five pillars jointly established by the board of directors and the "Long-Term Benefit Trust":

• Solving technical challenges in AI safety;
• Making AI safety a common industry knowledge and backed by law;
• Strengthening democracy and global stability;
• Expanding economic prosperity and human well-being;
• Ensuring that AI dividends benefit the general public.

How to ensure "dividends benefit all"? Just three days before the report was released, Anthropic dropped a blockbuster — announcing the establishment of the Anthropic Public Benefit Fund.

When you calculate this sum, it is enough to make Wall Street capitalists stunned:

The company promises (pending shareholder approval) to inject at least 5% of its shares directly into the fund; in the future, every time the company's valuation rises by 50%, the board of directors may additionally add 2% of shares at its discretion.

If we include the 80% personal equity that the founding team has long promised to donate, plus the matching donations from employees, about 20% of Anthropic's shares will eventually be completely directed to public interests!

Judging from Anthropic's current valuation that is moving toward hundreds of billions or even trillions of US dollars, this is an astronomical amount of wealth.

What is this money used for?

The report makes it very clear: In addition to funding non-profit organizations, it will also be used to cope with the severe impact of AI on the human labor market, and even does not rule out the possibility of distributing money directly to specific affected groups in the future. Earlier, they had separately invested 200 million US dollars to establish the Economic Future Research Fund for this purpose.

Use the money earned by AI to save people whose jobs are taken away by AI.

Hopefully on the day when super intelligence really comes, when humans look back at this autumn of 2026, they will be glad that someone once wrote such a report.

References:

https://www.anthropic.com/public-benefit-report

https://www-cdn.anthropic.com/a9ceec76a54d28d27affffcdab6faf7c142cd587.pdf

https://x.com/AnthropicAI/status/2112966949273801086

This article is from the WeChat Official Account "AI Era" (ID: AI_era), written by ASI Apocalypse, edited by David, published with authorization from 36Kr.