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

The governance of systems like large language models and generative AI has transitioned from voluntary ethical frameworks into a complex, binding, and highly fragmented landscape of global laws and regulations.

The rules currently governing how AI is built, deployed, and held accountable are characterized by a major split in approach between the European Union, the United States, and China.

1. The European Union: The EU AI Act

The European Union’s AI Act is the world’s most comprehensive and structured binding framework. Rather than regulating AI as a single technology, it uses a risk-tiered approach, applying strict rules based on how an AI is used.

•General-Purpose AI (GPAI) Rules: Large language models and foundational AI systems fall under rules that went into effect in August 2025. Providers of these models must supply technical documentation, comply with EU copyright law, and publish detailed summaries of their training datasets.

•Systemic Risk Obligations: Highly capable frontier models (determined by computing power) face stricter rules. Developers must conduct model evaluations, perform adversarial testing ("red-teaming"), track and report energy consumption, and report cybersecurity incidents to the EU AI Office.

•Transparency and Content Labeling: Provisions under Article 50 of the Act enforce transparency. If a system is generating synthetic text, audio, or video (like deepfakes), the outputs must be machine-readable and clearly marked as AI-generated.

•Prohibitions: Systems deemed to pose an "unacceptable risk"—such as untargeted scraping of facial images, emotion recognition in workplaces, or social scoring—have been banned.

2. The United States: State Patchwork vs. Federal Preemption

The United States does not have a single comprehensive federal AI law. Instead, it has a shifting federal policy environment alongside a patchwork of active state-level laws.

Federal Level (Deregulation & Security focus)

Federal AI governance is primarily driven by executive orders and existing regulatory agencies, with a focus on promoting innovation while maintaining cybersecurity:

•Executive Orders: Executive Orders (such as EO 14179 and EO 14409) focus on removing regulatory barriers to AI adoption, promoting national competitiveness, and securing AI infrastructure. The federal government has established a voluntary 30-day pre-release review window for developers to share frontier models with federal agencies before they are released to the public.

•Federal Preemption Push: To prevent state-by-state laws from stifling national AI development, federal directives (like EO 14365) have tasked a DOJ AI Litigation Task Force with evaluating and potentially challenging "onerous" state laws in court.

•Existing Agency Enforcement: Agencies use their existing legal mandates to police AI. For example:

•The Federal Trade Commission (FTC) prosecutes deceptive AI marketing, data scraping violations, and the generation of fake reviews.

•The Equal Employment Opportunity Commission (EEOC) regulates and investigates algorithmic bias in AI-driven hiring software.

State Level (Enforceable Protections)

Because of the lack of a unified federal statute, several states have stepped in to pass their own binding rules:

•California: Has passed some of the nation's most targeted laws. The AI Training Data Transparency Act (AB 2013) requires generative AI developers to publish public summaries of the datasets used to train their models. Meanwhile, the AI Transparency Act (SB 942) mandates watermarking or labeling of AI-generated content.

•Colorado: Shifted from a complex "duty-of-care" framework to a disclosure-focused model under SB 26-189. It mandates that if an automated decision-making tool (ADMT) is used to make high-stakes decisions (like hiring, housing, or lending), companies must notify users, explain adverse outcomes, and provide a pathway for human review.

3. China: Targeted & Algorithmic Controls

China regulates AI through targeted, technology-specific regulations rather than a singular catch-all law.

•Generative AI Regulations: Under China’s generative AI rules, developers must register their algorithms with the Cyberspace Administration of China (CAC) before public release.

•Accountability for Output: Providers must ensure that the training data and generated content adhere to state content standards and "core socialist values". They must also label synthetic content (such as deepfakes) with conspicuous watermarks and implement mechanisms to prevent the generation of illegal or harmful content.

How AI Systems are Held Accountable in Practice

Across all these legal frameworks, accountability is generally enforced through four key pillars:

1.Information Disclosure (Transparency): Forcing developers to open the "black box" by documenting what data was used to train the system, how the system makes decisions, and marking synthetic media so users know they are interacting with an AI.

2.Algorithmic Bias Auditing: Requiring organizations that deploy AI in high-stakes fields (such as credit scoring, employment, or housing) to conduct independent audits to prove the system does not discriminate based on race, gender, or other protected classes.

3.Liability and Redress: Shifting legal liability onto developers and deployers if an unchecked AI causes tangible harm. If an automated tool makes a decision that negatively impacts a person, they must be given a clear path to request a human review of the decision.

4.Frontier Safety & Red-Teaming: For the largest, most powerful systems, developers are legally or voluntarily required to perform rigorous safety testing to ensure the model cannot easily be weaponized (e.g., used to assist in cyberattacks or create dangerous substances).

Citations (13)

1transcend.ioCompanies
transcend.io
Backs: “The governance of systems like large language models and generative AI has transitioned from voluntary ethical frameworks into a complex, binding, and highly fragmented landscape of global laws and regulations.”
2worldairegulation.orgNonprofits and think tanks
worldairegulation.org
Backs: “The governance of systems like large language models and generative AI has transitioned from voluntary ethical frameworks into a complex, binding, and highly fragmented landscape of global laws and regulations.”
3collibra.comCompanies
collibra.com
Backs: “The rules currently governing how AI is built, deployed, and held accountable are characterized by a major split in approach between the European Union, the United States, and China.”
4europa.euGovernment
europa.eu
Backs: “Rather than regulating AI as a single technology, it uses a risk-tiered approach, applying strict rules based on how an AI is used. General-Purpose AI (GPAI) Rules: Large language models and foundational AI systems fall under rules that went into effect in August 2025.”
5alfred.eduAcademic
alfred.edu
Backs: “Rather than regulating AI as a single technology, it uses a risk-tiered approach, applying strict rules based on how an AI is used. General-Purpose AI (GPAI) Rules: Large language models and foundational AI systems fall under rules that went into effect in August 2025.”
6artificialintelligenceact.euNonprofits and think tanks
artificialintelligenceact.eu
Backs: “General-Purpose AI (GPAI) Rules: Large language models and foundational AI systems fall under rules that went into effect in August 2025. Providers of these models must supply technical documentation, comply with EU copyright law, and publish detailed summaries of their training datasets.”
7europa.euGovernment
europa.eu
Backs: “Transparency and Content Labeling: Provisions under Article 50 of the Act enforce transparency. If a system is generating synthetic text, audio, or video (like deepfakes), the outputs must be machine-readable and clearly marked as AI-generated.”
8verifywise.aiCompanies
verifywise.ai
Backs: “Prohibitions: Systems deemed to pose an "unacceptable risk"—such as untargeted scraping of facial images, emotion recognition in workplaces, or social scoring—have been banned. 2.”
9drata.comCompanies
drata.com
Backs: “Federal Preemption. The United States does not have a single comprehensive federal AI law. Instead, it has a shifting federal policy environment alongside a patchwork of active state-level laws.”
10powertofly.comNot yet classified
powertofly.com
Backs: “Federal Preemption. The United States does not have a single comprehensive federal AI law. Instead, it has a shifting federal policy environment alongside a patchwork of active state-level laws.”
11wikipedia.orgEncyclopedias and referencequality 0.83
wikipedia.org
Backs: “Federal Level (Deregulation & Security focus). Federal AI governance is primarily driven by executive orders and existing regulatory agencies, with a focus on promoting innovation while maintaining cybersecurity: Executive Orders: Executive Orders (such as EO 14179 and EO 14409) focus on removing regulatory barriers…”
12bakerdonelson.comNot yet classified
bakerdonelson.com
Backs: “For example: The Federal Trade Commission (FTC) prosecutes deceptive AI marketing, data scraping violations, and the generation of fake reviews. The Equal Employment Opportunity Commission (EEOC) regulates and investigates algorithmic bias in AI-driven hiring software.”
13verifywise.aiCompanies
verifywise.ai
Backs: “State Level (Enforceable Protections). Because of the lack of a unified federal statute, several states have stepped in to pass their own binding rules: California: Has passed some of the nation's most targeted laws.”