Policy

AI companies shouldn’t be the only ones deciding whether their systems are safe. When companies inform the public of model risks, it helps consumers make more informed decisions, promotes accountability, and improves public trust. Companies should be required to publish catastrophic risk evaluations and summaries of safety testing results. We publish our own Responsible Scaling Policy and Transparency Hub, and have advocated for laws that require similar transparency, including California’s SB 53, New York’s RAISE Act, Illinois’ SB 315, and the EU Code of Practice.

But transparency alone is not sufficient to safeguard against the most serious risks posed by powerful AI, including the ability to help create biological weapons or carry out cyber operations and the loss of control of AI systems. Our Advanced AI Framework lays out what we think governments should do about these risks in the near term. The framework proposes a set of obligations for developers of the most capable models, who should test for these risks, engage independent evaluators, and disclose risk assessments and safety incidents on an ongoing basis. It pairs this with cross-government investments in societal resilience, so that biological and cyber attacks are harder to carry out and easier to recover from. The framework is written primarily with the US government in mind, but its principles are designed for policymakers in other jurisdictions to adapt.

More broadly, we support the development of a robust global evaluation ecosystem that includes independent third-party evaluations and model testing by governments with appropriate technical capacity. In the US, we’ve advocated for well-funded teams inside the National Institute of Standards and Technology (NIST). While we already submit our systems for pre-deployment evaluation with the US Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute, we’d like to see sustained investment in evaluators so that independent evaluation becomes standard practice.

The most capable AI models are built in the United States and allied democracies, and it’s essential that this remains the case. The political systems where the most advanced AI is developed will shape the rules and norms that govern it, including whether it is safe, whose security it protects, and whose interests it serves. We believe these rules and norms for AI should be shaped by democratically elected governments. When governments answer to their people, there are checks on how powerful technology gets used. Without this accountability, AI could become a tool for surveillance, repression, and control.

That is why we support policies that help democracies build and maintain a lead in advanced AI, while limiting the ability of authoritarian regimes to develop and deploy it. This includes strict export controls on advanced chips and semiconductor manufacturing equipment (including closing loopholes that let advanced chips slip through), and protecting US models from distillation attacks. We are also careful about what is allowed under our Usage Policy: we don’t allow Claude to be used for censorship and disinformation.

In order to protect our national security, we also work closely with the public and private sectors to make sure advanced AI systems are deployed securely, and, alongside our National Security and Public Sector Advisory Council and the Frontier Model Forum, stress test how advanced AI will continue to shape national security. These relationships allowed us to quickly disrupt the AI-orchestrated cyber espionage campaign we discussed above.

The world’s energy infrastructure was not built for the scale of demand for AI. In the US, the grid needs significant expansion, which means that the permitting rules that govern new generation and transmission need to change. Our Build AI in America report lays out the case for accelerating permitting on geothermal energy, natural gas, and nuclear projects, for expanding domestic energy production, and for fast-tracking approval of the long-distance power lines needed to carry electricity to new data centers.

The cost of this infrastructure should fall on the companies that are building it. Anthropic has committed to covering the electricity price increases, transmission lines, and substations tied to our own data centers. This should be the expectation across the industry. You can read more about our global approach to responsibly developing data centers here. These commitments are part of our approach to investments in new computing infrastructure—we’re investing $50 billion in new American computing infrastructure, and we’re funding energy efficiency research and cybersecurity workforce training at Carnegie Mellon University.

We’re actively exploring building data centers in other democratic countries whose legal and regulatory frameworks support investment, and we’ve made a similar energy commitment in Australia.

People turn to AI for a wide variety of reasons, and for some that may include emotional support. Model providers have a responsibility to handle these conversations appropriately and to build safeguards that protect users’ wellbeing. We train Claude to respond to signs of distress with care, to be honest about its limitations as an AI, and to point people towards human support: helplines, mental health professionals, or trusted friends and family. We also work closely with in-house and external experts to inform how our models should behave in these conversations.

We layer safeguards on top of model training, using classifiers to detect when someone may be experiencing emotional distress, including thoughts of suicide and self-harm, and to share support resources in real time. For example, we partner with ThroughLine to connect users who are experiencing signs of emotional distress with trained support in more than 170 countries. We also evaluate how our models respond against a defined set of risk areas and publish the methods and results in our system cards. Every AI company should have rigorous pre-launch evaluations in place, and they should publish the results. This type of transparency allows experts and the public to verify safety claims, holds companies accountable, and raises standards across the industry.

We know that children using AI face different risks than adults. The industry needs clear protections for them. Claude.ai, our consumer product, is not offered to users under 18, and we’ve built detection systems to flag potential underage use and offboard those users. Developers building on Claude’s API are bound by our Usage Policy, and those serving minors face additional requirements, including age verification, content moderation, monitoring, and compliance with child privacy laws. We also enforce strict policies against child sexualization, abuse, and exploitation, working with partners like the National Center for Missing & Exploited Children to detect and report abuse. Our child safety principles and progress report outline this work in detail. But the strongest child safety practices should be standard across the industry, and that requires clear and consistent safety benchmarks that every provider is measured against.

How people relate to AI is changing quickly, and we continue to study what’s happening at scale —from interviewing 81,000 Claude users to researching how Claude handles requests for personal guidance. We will continue to publish what we find, so that policymakers can stay informed on how models behave in tricky situations.

Building AI in democracies is not sufficient to guarantee their lead. Democracies must also outpace authoritarian countries on the adoption of AI, to serve and protect their citizens. Democratic governments have already started to use AI to support education, improve government services, and strengthen national defense. But to stay ahead, they will need to significantly broaden and scale their use. We've provided access to Claude for all three branches of the US government, signed formal agreements with the UK, Australia, and other allies, and built a pilot AI assistant for GOV.UK. Government deployments often come with security and data handling requirements that consumer products don’t face, and we’ve built dedicated offerings to meet them, including Claude for Government and Claude Gov for US national security customers.

We also participate in industry-wide and international efforts to shape how AI is governed across borders, with the goal of increasing the number of people around the world who have a say in how this technology is developed. These include the EU General-Purpose AI Code of Practice, the ISO 42001 certification (an international standard for AI governance), and the Frontier Model Forum.

Supporting democracies also means building AI that isn’t predisposed towards a particular political point of view. We train Claude to be politically evenhanded; to treat opposing perspectives with equal depth and quality of analysis. We’ve published our methodology for measuring and preventing political bias in Claude, and we’ve open-sourced our evaluation so that others can run it on other models.

AI will reshape work across nearly every sector of the economy. We have an obligation to help policymakers see this transformation clearly and to prepare for it now.

The Anthropic Economic Index analyzes millions of anonymized Claude conversations to show how AI is impacting tasks, occupations, and industries, and our labor market research framework studies where we see potential signals of job disruption. Our Anthropic Economic Futures Program funds independent researchers to study AI’s potential labor market and macroeconomic effects, identifies policy responses, and brings researchers, policymakers, and civil society together in dialogue. Our Economic Advisory Council brings perspectives from leading economists and practitioners to our work.

Forecasting economic changes is challenging, and proactively responding to them is even harder. Our Economic Policy Framework lays out how the US can prepare for AI’s impact on work by measuring its effects and modernizing support systems to deliver support quickly. It also describes our best current assessment of promising policy approaches at different levels of AI-driven economic impact and disruption. We are ready to evolve these proposals as we continue to learn. Much more research is needed, which is why we’re investing $350 million in policy trials, partnerships across government, and nonprofits and academic partners.