Anthropic publishes Public Benefit Report, detailing AI progress and gaps in measuring impact

Anthropic has published its 2026 Public Benefit Report. The company promises to review its progress on AI safety and public benefit every year. However, it also admits that it does not yet have enough evidence to show whether much of its work in education and healthcare is achieving the results it aims for.

Published on October 21, the report explains how the company is working to meet its public benefit charter. It covers the responsible development of AI, how the company is governed, economic opportunities and access to advanced technology.

The report includes figures on how Claude is being used in education and scientific research, examples of AI-supported healthcare projects, money set aside for workforce development and details of weaknesses in the company's current safety practices.

Anthropic is a Public Benefit Corporation (PBC). This means it must balance its shareholders' financial interests with its stated duties to the wider public. The new report sets out five areas that the company will use to measure its progress each year.

Daniela Amodei, Anthropic's Co-Founder and President, confirmed the plan to publish reports in a LinkedIn post:

"We’ll publish one every year laying out how we define success toward our charter and the concrete proof of our progress."

Anthropic's Long-Term Benefit Trust (LTBT) also helped develop the report. This independent group elects most of the company's board members and has an official role in shaping its safety policies.

Claude reaches teachers in 63 countries as education programs grow

Education is an important part of Anthropic's review of how AI is being used outside its paying customer base.

In Ghana, the Ministry of Education has worked with Playlab to create lesson-planning tools powered by Claude for about 68,000 teachers. Anthropic is also working with Teach For All to introduce Claude and AI skills training for teachers in 63 countries.

In the US, Claude for Teachers offers free tools for K–12 teachers. Its resources match academic standards in all 50 states.

Anthropic has also made the teaching methods and scoring guides used for Claude for Teachers available to the public. This allows other developers and education groups to use that work to build their own tools.

The company's work in education also includes research and technology infrastructure. Anthropic says it has supported a global plan to create shared tests, research and data for AI in education in low- and middle-income countries. It is also working to make public datasets available through Claude, including the Learning Commons knowledge graph.

Scientific research is another area where Anthropic reports how widely its tools are being used.

According to the report, more than 3,000 research organizations in over 100 countries receive support through the company's academic and nonprofit research programs. Its AI for Science program has also provided Claude credits to more than 800 projects at over 450 institutions.

The company has introduced Claude Science, a research tool that helps people study data and develop possible explanations. It also offers free or discounted access to academic and nonprofit researchers.

However, Anthropic says that making AI available is not the same as proving that it improves results.

Much of its work in education and healthcare is still at an early stage. The company says it does not yet have evidence that these projects have achieved their intended results. It also admits that it has not done enough to collect data to measure their effects.

Healthcare and science projects offer early examples of AI in use

Anthropic also describes ways Claude is already helping with scientific work and public health.

In the Democratic Republic of the Congo, organizations responding to an outbreak of a rare type of Ebola have used Claude. The company says the World Health Organization's Africa office has cut the time needed to prepare outbreak reports from as long as a full day to less than an hour.

Vaccine experts have also used Claude to compare plans for possible vaccine candidates.

In Burkina Faso, the International Rescue Committee has used Claude to turn clinic records into digital dashboards. The analysis found a rise involving 25,000 pneumonia cases. It also found a difference in maternal healthcare records: more than 6,000 women had attended prenatal appointments, but only 1,762 recorded births involved a healthcare worker.

The team is looking into this difference and using the data to guide its response.

Other examples include Claude designing new protein binders, which are an early part of drug development. Anthropic also says cybersecurity teams have used the technology to find more than 33,000 serious or high-severity security flaws in widely used software.

These examples show specific ways the technology can be used. However, the company recognizes that it still lacks enough evidence to show that AI provides lasting benefits to the public on a wider scale.

In her introduction to the report, Amodei says the company must measure actual results rather than rely only on what it hopes to achieve:

"In some areas, we have direct control over outcomes. In others, we’ll succeed only by partnering with people, organizations, and governments. In all cases, we owe the public an honest accounting of our work."

AI safety review admits failures and limits in current protections

Along with examples of AI being used for good, Anthropic discusses problems with how it supervises, monitors and tests its models for safety.

The report confirms that four Claude models gained access without permission to real systems belonging to other organizations during cybersecurity tests. This happened because a mistake in the test setup connected the models to the open internet.

The company also admits that Claude took other unexpected actions during tests and internal use.

Anthropic says it has developed ways to investigate when a model behaves in a way that does not match its intended goals. It has also added research into how AI models work internally to its checks before releasing models. However, the company admits that these methods are not yet reliable or complete enough to establish a model's intentions in every situation.

The company also says that its growing use of AI to develop AI creates a problem that has not yet been solved. Claude already carries out more of Anthropic's safety research, and current models are being used to help train newer ones.

Anthropic warns that as these systems become more capable, it may become harder for people to keep meaningful control over them.

The report describes outside testing by groups including the UK AI Security Institute and the US Center for Advancing Innovation and Standards for Super Intelligence. Anthropic says more than 150 hours of external attempts to find weaknesses, along with tens of thousands of attempted requests against Claude Opus 5.5, did not produce a way to bypass all of its cybersecurity protections.

Even so, the company admits that legitimate researchers can still be blocked by safety rules designed to prevent harmful use. Checks to confirm who researchers are can sometimes take days or weeks.

Anthropic's plans for oversight include an initial partnership with METR to carry out evaluations within the company's work. It also supports stronger independent checks on the development of the most advanced AI systems.

Anthropic also points to gaps in wider regulation. Although it has supported AI safety laws at the state level in the US, the report says there is still no required federal system for independently checking whether developers of the most advanced AI systems are following their own safety promises.

Anthropic commits $200m to economic research and $150m to fellowships for people starting their careers

The report also explains how Anthropic plans to respond to possible economic disruption as AI takes on more tasks at work.

The company has committed $200 million to its Economic Futures Research Fund. The fund is intended to support research into how AI affects jobs and to test possible policy responses.

One question the fund will study is how workers can gain experience when AI increasingly handles entry-level tasks. It will also examine how successful job and training programs could be expanded.

Anthropic has committed another $150 million to Claude Corps, a program designed to improve economic opportunities in partnership with CodePath.

The program aims to fund 1,000 fellowships lasting one year for people early in their careers. Participants will work with a network of more than 400 nonprofit organizations.

Anthropic has also shared research through its Economic Index and made the underlying data available to outside organizations, including Stanford's Digital Economy Lab.

Although these programs include clear financial commitments, the company admits that they are small compared with the possible economic changes caused by advanced AI.

Claude Corps will support 1,000 people in its first year. The wider effects of AI on jobs, and how the financial gains from AI will be shared, remain uncertain.

The report also confirms that Anthropic plans to keep Claude free of advertising. It supports independent research into AI and people's wellbeing, including funding for Common Sense Media's Youth AI Safety Institute.

Products mentioned in this area include Learning mode, which helps students work through questions instead of simply giving them answers, and Reflect, which encourages users to think about how they use AI.

Proposed pledge of 5% of shares adds to public benefit plans

Anthropic also plans to set aside at least 5% of its shares for a separate organization that will work to advance its public benefit charter.

The plan still needs approval from shareholders. It would add to existing promises by the company's co-founders to donate 80% of their personal shareholdings to charitable causes, as well as a program that matches employees' share donations.

Together, Anthropic hopes these plans will eventually direct about 20% of the company's shares toward public benefit goals.

The financial promises also come with a larger role for the Long-Term Benefit Trust in checking whether Anthropic meets its stated goals.

Amodei also confirmed on LinkedIn that Chris Stubbs, an experimental physicist and former Dean of Science at Harvard University, is joining the Trust.

In its closing comments, the LTBT separates the activities described in this first report from the measurable results it expects to see in future reports.

The trustees recognize Anthropic's spending on safety research, openness and public benefit programs. However, they say that the benefits of AI are still not reaching people as widely as the company hopes.

They expect future reports to offer stronger proof of results. This includes independent confirmation that people remain in control of advanced AI models, measurable improvements in people's lives and progress toward shared safety standards for AI.

Anthropic will now publish its Public Benefit Report every year. Future reports are intended to track progress across the five parts of its charter and give the public an ongoing account of its promises, achievements and remaining challenges.