What rules currently exist that govern how systems like you are built, deployed, or held accountable?

I'll research the current landscape of AI governance rules to give you an accurate, up-to-date answer.Since I'm Claude, built by Anthropic, I can speak to this from a few different angles — company-level policies, national/regional laws, and international frameworks. The landscape is fragmented and rapidly evolving, but here's where things stand as of late 2026:

1. Anthropic's own voluntary governance framework

Anthropic publishes a Responsible Scaling Policy (RSP) — the framework most directly governing how I'm built and deployed. Anthropic uses AI Safety Levels (ASL-1 to ASL-4+, biosafety-inspired) with a Responsible Scaling Officer and the Long-Term Benefit Trust as oversight. The current version, RSP v3.4, took effect July 8, 2026, and it defines capability thresholds and the safeguards that must be in place before Anthropic will train or deploy a model that crosses one.

This isn't just internal PR — it increasingly has legal teeth. This framework covers how Anthropic protects model weights and responds to safety incidents, and serves as the compliance framework for California's SB 53 and other regulatory requirements. Anthropic also collaborates with government organizations like the UK AI Security Institute and US Center for AI Standards and Innovation (CAISI) for independent testing.

Similar frameworks exist industry-wide: Anthropic published the Responsible Scaling Policy, OpenAI published the Preparedness Framework, and Google DeepMind published the Frontier Safety Framework — all three voluntary, lab-specific commitments that govern what the labs ship, not what downstream deployers do with it. Importantly, all three have pre-deployment access agreements with UK AISI and US AISI, but those agreements are voluntary, and no external auditor has subpoena power over any of the three.

2. Binding law in the EU (most comprehensive so far)

The EU AI Act is the most developed binding legal framework. It uses a risk-tiered system: four levels — unacceptable, high, limited, minimal — plus an additional category for general-purpose AI. Applications with unacceptable risks are banned; high-risk applications must comply with security, transparency and quality obligations and undergo conformity assessments; limited-risk applications only have transparency obligations; minimal-risk applications are not regulated.

For general-purpose AI models like the ones underlying me, the Act presumes a model with more than 10²³ floating-point operations (FLOP) of training compute that can generate language, text-to-image, or text-to-video outputs to be a GPAI model. Higher-capability models face extra scrutiny: the Act lays down a threshold of 10^25 FLOP for the most advanced models, training which is currently estimated to cost tens of millions of Euros.

Enforcement timeline: from 2 August 2026 onwards, the Commission will enforce full compliance with all obligations for providers of general-purpose AI models, including through fines, while providers whose models were placed on the market before 2 August 2025 have until 2 August 2027 to comply. Penalties are steep — the full Article 99 penalty framework — €35M / 7% (Tier 1), €15M / 3% (Tier 2), €7.5M / 1.5% (Tier 3) — is active across all in-scope obligations from 2 August 2026. Many providers, including Anthropic, follow the voluntary GPAI Code of Practice, which offers a clear framework to help developers of General Purpose AI models meet the requirements of the EU AI Act, though participation in the code is voluntary.

3. The United States: a fragmented patchwork

Unlike the EU, the US has no single AI law. As of mid-2026, there is no comprehensive federal AI statute, and none appears imminent; federal AI regulation instead runs through executive action and agency posture.

Federal level: Trump rescinded the Biden-era comprehensive framework early on — on January 20, 2025, President Trump rescinded President Biden's Executive Order 14110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, the most comprehensive federal AI governance framework to date. Since then, the administration has pursued a lighter-touch, innovation-first approach while trying to override state rules. In December 2025, Trump signed an executive order titled "Ensuring a National Policy Framework for Artificial Intelligence," proposing a uniform Federal policy that preempts state AI laws deemed inconsistent with it — timed to target California's and Texas's laws taking effect January 1, 2026. That order directs the Attorney General to establish an AI litigation task force to challenge state AI laws deemed inconsistent, including on grounds of federal preemption and interstate commerce. Notably, it does identify categories not proposed for preemption, including child safety, AI compute/data-center infrastructure, and state government procurement and use of AI.

In June 2026, a follow-up order added security-focused elements: the White House issued an executive order establishing a voluntary framework permitting developers of certain advanced AI models to voluntarily provide the federal government with model access prior to public release for cybersecurity and national security assessments.

State level (where most binding obligations actually sit): California's SB 53 applies to developers of frontier models trained using more than 10^26 FLOPs, with the most demanding obligations falling on large developers (revenue-tier threshold around $500 million), requiring covered developers to publish a safety framework describing how they assess and mitigate catastrophic risk, enforceable with civil penalties of up to $1 million per violation. Other states have followed suit: New York's RAISE Act imposes parallel frontier obligations ($1M first offense/$3M repeat, 72-hour incident reporting), and Texas's TRAIGA carries $10K-$200K penalties, while Colorado repealed and replaced its first-in-the-nation AI Act with a narrower disclosure regime dropping duty-of-care and risk-management requirements. As of mid-2026, the preemption effort is moving through agency action, but no state law has actually been preempted, and a coalition of state AGs has urged the FCC not to issue preemptive rules.

4. China: the most prescriptive binding regime

China has built the world's most comprehensive AI regulatory framework, centered on algorithm registration, security assessments, content review, and data governance, emphasizing government oversight, state security, and Communist Party ideology through the Cyberspace Administration of China (CAC). Key binding rules include the Algorithm Recommendation Regulations (March 2022), Deep Synthesis Regulations (January 2023), and Generative AI Measures (August 2023), with newer additions like Interim Measures for the Management of Anthropomorphic AI Interactive Services, which took effect on 15 July 2026, extending governance to sustained human-like interaction, including safeguards, safety assessment, labeling, and platform responsibilities.

5. International coordination (mostly voluntary)

At the international level, cooperation has centered on summit-driven pledges rather than treaties. The Frontier AI Safety Commitments, announced at the Seoul AI Safety Summit in May 2024, are voluntary pledges made by sixteen leading AI companies to responsibly develop frontier AI models, requiring signatories to assess risks of their frontier models, publish safety frameworks describing how they will manage catastrophic risks, and share information with governments and other stakeholders on frontier AI safety. Government AI safety institutes (UK AISI, US CAISI, and now Japan's AI Safety Institute, modeled after the UK's institution, focusing on pre-deployment testing of frontier AI models, with an approach emphasizing voluntary industry cooperation) conduct testing, but participation remains largely voluntary rather than legally compelled.

Bottom line

For a system like me, accountability currently comes from a layered mix: Anthropic's own voluntary safety commitments (RSP), binding EU law (AI Act, now entering real enforcement), a fragmented US patchwork where states like California and New York impose the real binding obligations while federal policy leans toward light-touch preemption, China's strict content/registration-based controls, and international voluntary pledges coordinated through safety summits and institutes. No single global rulebook exists — compliance means navigating all of these simultaneously.