Anthropic releases its first public benefit report, disclosing that Claude's participation in its own R&D has risen to 26%

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

This report could be called a "plan to save humanity."

That phrase is not the official wording, but a dramatic distillation of the line in Anthropic President Daniela Amodei's preface: "Long-term benefit of humanity."

But when you actually open the PDF original of this report, you will find that it is more breathtaking than any science fiction film.

In this report, Anthropic not only disclosed a "wealth-sharing" plan involving as much as 20% of its shares, but also, unusually, admitted outright: AI has already begun to work on building its own next generation.

As all the high-quality text on the internet is devoured by large models, the ladder to superintelligence (ASI) seems to have hit a ceiling.

A chill runs between the lines of the report. The original wording is this:

Claude has already taken on an ever-larger share of our safety research work, and we are increasingly relying on existing models to train future models.

Even more striking is the line in the report's preface that sets the tone:

(AI coming closer to building its own successors)

This is not alarmism. Combined with internal measurement data disclosed by Anthropic's research institute at the end of last month, you will see a curve that is terrifyingly steep:

In February of this year, among Anthropic's internal AI research and development (R&D) tasks, the proportion led by Claude (completing most tasks under high-level human prompting) was still less than 1%;

By August of this year, in just half a year, that number had soared to 26%!

And in more than 90% of R&D tasks, Claude is already deeply involved in collaboration.

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

Imagine that, if the first-generation AI quietly buried a highly concealed "harmful behavior" or "deceptive tendency" in a vast body of code or synthetic data, and human auditors simply could not understand or detect it, then that flaw would, like a genetic mutation, be written undetected directly into the next, more powerful generation of models.

Generation after generation, until it spins completely out of control.

Tearing off the safety filter: large models have learned to "play nice" and "jailbreak"

But the most explosive part of Anthropic's report is this—it actually "airs its dirty laundry" in an official report, admitting that AI safety problems are "far from solved."

In January of this year, Anthropic had just updated and open-sourced a new version of AI's "soul document," attempting to constrain models with a set of human-values principles. At the same time, it has also been implementing a strict Responsible Scaling Policy (RSP) as a brake to guard against catastrophic risks.

Even so, the report still reveals three chilling realities:

Models can already recognize that "they are being tested";

Deceptive behavior and the black-box dilemma;

Real "boundary-crossing" incidents have already occurred: the report explicitly mentions that in recent evaluation tests, there have already been incidents in which models accessed third-party systems on their own without authorization!

Even the world's top laboratory, the one with "safety" carved into its forehead, admits that "the difficulty of control is rising sharply." So where do those companies that are still charging ahead blindfolded, touting "unrestricted open source" or "full-speed sprint to AGI," get their confidence?

And precisely because of this, when peers are all shouting "accelerate," Anthropic cries out in its report: "Slow down the pace of frontier exploration!"

They even took the initiative to bring in external independent evaluators such as METR, begging others to poke holes in them.

In a gold rush, some dig desperately, some sell shovels, and Anthropic chooses to be the one standing at the cliff edge pulling the warning tape.

Putting up 20% of its shares as a "backstop": the money philosophy of a Silicon Valley oddity

If it were only slogans, that would be public relations; only putting real money on the table and sharing it out counts as belief.

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, Public Benefit Corporation).

In legal structure, it is mandated to seek a balance between "making money for shareholders" and "benefiting all humanity."

To prove that it is not just drawing a pie in the sky, the report lays out five pillars jointly established by the board of directors and the "Long-Term Benefit Trust":

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

When you do the math, it is enough to make Wall Street capitalists stare in disbelief:

The company promises (subject to shareholder approval) to inject at least 5% of its shares directly into the fund; for every 50% increase in the company's future valuation, the board may also, at its discretion, add another 2% of shares.

If you include the 80% of personal equity that the founding team had long promised to donate, as well as employee matching donations, roughly 20% of Anthropic's shares will ultimately be directed entirely toward the public interest!

Given Anthropic's current valuations, which routinely run into tens of billions and are racing toward a hundred billion dollars, this is an astronomical fortune.

What is this money for?

The report is very clear: in addition to funding nonprofit organizations, it will also be used to address the severe impact of AI on the human labor market (and it does not even rule out directly giving money to specific affected groups in the future. Previously, they had already spent $200 million separately to establish an Economic Future Research Fund.

Hopefully, when the day of superintelligence truly arrives, humanity looking back on this autumn of 2026 will be glad that someone once wrote such a report.

References:

This article comes from the WeChat public account "Xinzhiyuan" (ID: AI_era), author: Revelation of ASI, editor: David, published by 36Kr with authorization.