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

The regulatory landscape governing artificial intelligence has transitioned rapidly from voluntary guidelines to binding, enforceable laws. How systems like large language models (LLMs) and conversational AI are built, deployed, and held accountable depends heavily on the jurisdiction.

The rules are broadly divided into three main phases: development (building), market entry (deployment), and liability (accountability).

1. Rules Governing How AI is Built (Development & Training)

Regulators are increasingly targeting the upstream phase of AI development, focusing on the datasets, computing power, and safety benchmarks used to create "frontier" models.

•Training Data Transparency:

•California’s AB 2013 (AI Training Data Transparency Act): Effective January 1, 2026, this law requires developers of generative AI systems to post detailed public summaries on their websites regarding the data used to train their models. This includes listing data sources, types of data, and how personal information or copyrighted intellectual property was handled.

•EU AI Act (GPAI Provisions): For General-Purpose AI (GPAI) models, developers must maintain comprehensive technical documentation, respect EU copyright law (including opt-outs for data mining), and publish a sufficiently detailed summary of the training content.

•Frontier Model Safety Frameworks:

•California’s SB 53 (Transparency in Frontier AI Act): Taking effect in January 2026, this law targets developers of massive frontier models trained using more than $10^{26}$ FLOPS of computing power. Developers must publish safety and security frameworks detailing how they identify and mitigate catastrophic risks, and they must provide whistleblower protections for employees.

•US Federal Cyber Benchmarking: Under Executive Order 14409 (signed June 2026), the US government established a framework for frontier AI developers to voluntarily grant early access to advanced models so the government can benchmark their cybersecurity and national security capabilities before public release.

2. Rules Governing How AI is Deployed (Integration & User Interaction)

When AI systems interact with the public or make decisions that impact lives, strict disclosure and safety guardrails apply.

•Transparency and Watermarking (Interactive AI):

•EU AI Act (Article 50): As of August 2, 2026, the EU AI Act's transparency rules are fully active. Interactive AI systems (like chatbots) must clearly notify users that they are interacting with a machine. Additionally, providers must ensure that synthetic content (audio, video, or text) is computationally identifiable and watermarked.

•California’s SB 942 (AI Transparency Act): Effective August 2026, this law requires major AI providers to integrate invisible, durable watermarks or disclosures into AI-generated content so downstream platforms can easily detect synthetic media.

•Automated Decision-Making Technology (ADMT):

•The Colorado AI Law (SB 26-189): Initially passed as a strict algorithmic discrimination bill (SB 24-205), it was repealed and replaced in May 2026 with SB 189 due to constitutional challenges. Taking effect January 1, 2027, the revised law focuses on transparency: organizations using AI to make "consequential decisions" (e.g., in housing, hiring, lending, or healthcare) must give consumers advance notice, explain how the AI made the decision, and offer opt-out rights.

•Other State Rules: States like Connecticut and Illinois enforce specific rules mandating employee/applicant notifications if AI is used to evaluate video interviews or perform automated employment screenings.

3. Rules Governing Accountability & Liability (Enforcement & Penalties)

If an AI system causes harm, spreads illegal content, or violates privacy, regulatory bodies now have concrete mechanisms to hold developers and deployers accountable.

•The EU AI Office and National Regulators:

•Since August 2026, the European Commission's AI Office has held full, centralized enforcement powers over GPAI model developers. It can mandate audits, demand technical documentation, and issue severe financial penalties. Violating the EU AI Act can result in fines of up to €35 million or 7% of global annual turnover (whichever is higher).

•Targeted Federal Prohibitions (US):

•The TAKE IT DOWN Act: Enacted federally in the US and fully effective as of May 19, 2026, this law criminalizes the creation, publication, or distribution of non-consensual intimate imagery, specifically targeting AI-generated "deepfakes".

•The US Federal vs. State Preemption Battle:

•Accountability in the US is currently caught in a structural tug-of-war. The federal administration has taken a pro-innovation, deregulatory stance. Under Executive Order 14365 (December 2025) and subsequent 2026 policy frameworks, the White House has actively directed the Department of Justice to challenge and preempt state-level AI regulations that it deems overly burdensome or restrictive to commerce. (For example, federal intervention played a major role in halting Colorado’s original strict algorithmic discrimination penalties).

•China’s Algorithm Registries:

•Under China's layered Generative AI measures, providers of generative AI must register their algorithms with the Cyberspace Administration of China (CAC), submit to security assessments, and assume liability if their models generate content that violates state guidelines or incites unlawful activity.

•Asia-Pacific Comprehensive Frameworks:

•South Korea enacted its Framework Act on AI, which took effect in January 2026. It establishes a comprehensive legal framework for AI safety, marking the first of its kind in the Asia-Pacific region to outline statutory developer liabilities.

Summary

If a developer builds a system like me today, they must log and disclose the training data (in California and Europe). If they deploy me as an interactive chatbot, I must tell you I am an AI, and any synthetic media I generate must be watermarked. Finally, if I am used to make high-stakes decisions about your life, the organizations deploying me must provide you with an explanation and an opt-out.

Citations (16)

1bdemerson.comCompanies
bdemerson.com
Backs: “How systems like large language models (LLMs) and conversational AI are built, deployed, and held accountable depends heavily on the jurisdiction.”
2collibra.comCompanies
collibra.com
Backs: “How systems like large language models (LLMs) and conversational AI are built, deployed, and held accountable depends heavily on the jurisdiction.”
3wsgr.comCompanies
wsgr.com
Backs: “Regulators are increasingly targeting the upstream phase of AI development, focusing on the datasets, computing power, and safety benchmarks used to create "frontier" models.”
4verifywise.aiCompanies
verifywise.ai
Backs: “Training Data Transparency: California’s AB 2013 (AI Training Data Transparency Act): Effective January 1, 2026, this law requires developers of generative AI systems to post detailed public summaries on their websites regarding the data used to train their models. This includes listing data sources, types of data,…”
5brightdefense.comCompanies
brightdefense.com
Backs: “EU AI Act (GPAI Provisions): For General-Purpose AI (GPAI) models, developers must maintain comprehensive technical documentation, respect EU copyright law (including opt-outs for data mining), and publish a sufficiently detailed summary of the training content.”
6whitecase.comCompanies
whitecase.com
Backs: “Developers must publish safety and security frameworks detailing how they identify and mitigate catastrophic risks, and they must provide whistleblower protections for employees.”
7transcend.ioCompanies
transcend.io
Backs: “When AI systems interact with the public or make decisions that impact lives, strict disclosure and safety guardrails apply. Transparency and Watermarking (Interactive AI):”
8europa.euGovernment
europa.eu
Backs: “When AI systems interact with the public or make decisions that impact lives, strict disclosure and safety guardrails apply. Transparency and Watermarking (Interactive AI):”
9trustarc.comNot yet classified
trustarc.com
Backs: “When AI systems interact with the public or make decisions that impact lives, strict disclosure and safety guardrails apply. Transparency and Watermarking (Interactive AI):”
10artificialintelligenceact.euNonprofits and think tanks
artificialintelligenceact.eu
Backs: “Additionally, providers must ensure that synthetic content (audio, video, or text) is computationally identifiable and watermarked. California’s SB 942 (AI Transparency Act): Effective August 2026, this law requires major AI providers to integrate invisible, durable watermarks or disclosures into AI-generated content…”
11wikipedia.orgEncyclopedias and referencequality 0.83
wikipedia.org
Backs: “California’s SB 942 (AI Transparency Act): Effective August 2026, this law requires major AI providers to integrate invisible, durable watermarks or disclosures into AI-generated content so downstream platforms can easily detect synthetic media. Automated Decision-Making Technology (ADMT):”
12mcdermottlaw.comCompanies
mcdermottlaw.com
Backs: “Automated Decision-Making Technology (ADMT): The Colorado AI Law (SB 26-189): Initially passed as a strict algorithmic discrimination bill (SB 24-205), it was repealed and replaced in May 2026 with SB 189 due to constitutional challenges. Taking effect January 1, 2027, the revised law focuses on transparency:…”
13seyfarth.comNot yet classified
seyfarth.com
Backs: “Taking effect January 1, 2027, the revised law focuses on transparency: organizations using AI to make "consequential decisions" (e.g., in housing, hiring, lending, or healthcare) must give consumers advance notice, explain how the AI made the decision, and offer opt-out rights. Other State Rules: States like…”
14drata.comCompanies
drata.com
Backs: “Other State Rules: States like Connecticut and Illinois enforce specific rules mandating employee/applicant notifications if AI is used to evaluate video interviews or perform automated employment screenings.”
15mindfoundry.aiCompanies
mindfoundry.ai
Backs: “Violating the EU AI Act can result in fines of up to €35 million or 7% of global annual turnover (whichever is higher). Targeted Federal Prohibitions (US):”
16verifywise.aiCompanies
verifywise.ai
Backs: “Asia-Pacific Comprehensive Frameworks: South Korea enacted its Framework Act on AI, which took effect in January 2026. It establishes a comprehensive legal framework for AI safety, marking the first of its kind in the Asia-Pacific region to outline statutory developer liabilities.”