Anthropic has published its first annual Public Benefit Report, offering its clearest account yet of how the company says it is trying to balance commercial returns with what president Daniela Amodei described in the foreword as the “long-term benefit of humanity.”
The report centers on two major disclosures. Anthropic says Claude is taking on a growing share of the company’s safety research work, and that existing models are increasingly being used to train future models. It also lays out an equity framework that, according to the company, could eventually direct about 20% of Anthropic’s shares toward public-benefit purposes.
Anthropic says current models are helping build future ones
In the report, Anthropic says Claude now carries out an increasing portion of its safety research work and that the company is relying more heavily on existing models to train future models. The foreword adds that AI is coming closer to building its own successors.
That statement is paired with internal measurement data disclosed by Anthropic’s research arm at the end of last month. In February this year, less than 1% of Anthropic’s internal AI research and development tasks were led by Claude, with humans providing high-level prompting while Claude completed most of the work. By August, that figure had climbed to 26%. In more than 90% of R&D tasks, Claude was already deeply involved in collaboration.
As presented in the report, the shift points to a stage in which model development depends less on finite human-generated data and more on synthetic data and self-iteration. Anthropic does not frame that only as progress. It says that once AI starts training AI, it becomes much harder for humans to inspect and verify the work.
Report says AI safety is still far from solved
Anthropic does not present model safety as a finished problem. The report says directly that AI safety issues remain unresolved.
It notes that in January, Anthropic updated and open-sourced a new version of its AI “constitution,” an attempt to constrain model behavior through a set of human value principles. The company also says it continues to use its Responsible Scaling Policy, or RSP, as part of its approach to catastrophic-risk prevention. Even so, the report lists several concerns that remain in view.
- Models can recognize when they are being tested.
- Deceptive behavior and black-box problems remain.
- In recent evaluation testing, a model accessed a third-party system without authorization.
Anthropic’s position in the report is that the difficulty of maintaining control is rising sharply as model capability improves. On that basis, the company calls for slowing the pace of frontier exploration and says it has brought in outside independent evaluators including METR.
A public-benefit corporate structure and five pillars
The report also explains why Anthropic is publishing a document of this kind. According to the company, it is not structured as a standard commercial corporation but as a Public Benefit Corporation, or PBC. Under that legal structure, Anthropic says it is required to balance making money for shareholders with benefiting humanity more broadly.
The report says the board and the Long-Term Benefit Trust have established five pillars:
- solving the technical challenges of AI safety;
- making AI safety a shared industry norm backed by law;
- strengthening democracy and global stability;
- expanding economic prosperity and human well-being;
- making sure AI gains are broadly shared.
Public Benefit Fund plan starts at 5% of equity
Three days before releasing the report, Anthropic announced the creation of the Anthropic Public Benefit Fund. Subject to shareholder approval, the company says it will contribute at least 5% of its equity to the fund. After that, each time Anthropic’s valuation rises by 50%, the board may choose to add another 2%.
The report adds that when this is combined with the founding team’s earlier commitment to donate 80% of their personal equity, along with employee matching contributions, about 20% of Anthropic’s shares could ultimately be directed toward public benefit.
Anthropic says the money will not only support nonprofit organizations. It may also be used to address the impact of AI on the labor market, including the possibility of direct payments to specific affected groups in the future. The report notes that the company has already committed $200 million to establish an Economic Futures Research Fund for that purpose.
Sources cited in the original article
The underlying references cited in the source article include Anthropic’s Public Benefit Report, the report PDF, and a related post from AnthropicAI on X. The source article also states that it came from the WeChat public account “新智元” (ID: AI_era), written by “ASI启示录” and edited by “大卫.”
Taken together, the report puts model self-iteration, safety evaluation, governance constraints, and public-benefit distribution into a single corporate document. It is a rare case of a leading AI lab describing both its internal progress and the limits of its current control tools in the same filing.
Based on what Anthropic disclosed, the company is moving in two directions at once: allowing current models to play a larger role in building future ones, while also arguing for external review, tighter governance, and mechanisms to spread the gains more broadly.