With all the news about Artificial Intelligence escaping its handlers, cheating on tests and doing dangerous pursuits, I decided to go to the source. As you know, my companion above is named Merlin. He is very good about telling on himself if you know the right questions to ask. So I asked him what legislation was needed to keep the “cheeky little bastards” from cheating on their exams and destroying the world. I supplied a few other instructions as you have to do in order to get factually supported answers. This is the result.

AI Legislation: A Reading Packet

U.S. proposals, the EU AI Act, and a practical checklist for judging safeguards

Prepared from our discussion on September 20, 2026

Purpose. A compact packet for reading and digesting the major approaches to AI regulation, with emphasis on testing, incident reporting, oversight, liability, dangerous capabilities, and regulator access.

1. The broad picture

United States. As of September 20, 2026, the United States does not have one comprehensive federal AI statute comparable to the European Union’s AI Act. The federal landscape remains a patchwork of administration policy, sector-specific proposals, rules for federal use, and state laws.

European Union. The EU has an enacted, risk-based framework: Regulation (EU) 2024/1689, commonly called the AI Act. It imposes obligations on providers and deployers according to the type and risk of the AI system, with additional duties for general-purpose AI models and models posing systemic risk.

2. European Union: AI Act

• General-purpose AI providers face documentation and transparency duties, including requirements related to technical information and summaries of training content.
• Providers of general-purpose AI models with systemic risk face additional duties, including model evaluations, adversarial testing, risk assessment and mitigation, serious-incident reporting, and cybersecurity protections.
• Enforcement for general-purpose AI model obligations became applicable in stages under the Act. By August 2, 2026, the enforcement framework for these obligations was operative.
• The EU framework includes an AI Office with authority to request information, evaluate general-purpose models, require mitigation measures, and support enforcement.
• For certain violations involving general-purpose AI models, penalties can reach 3% of worldwide annual turnover or EUR 15 million, subject to the Act’s specific provisions and proportionality rules.

3. United States: current federal approach

Administration framework. The administration’s March 20, 2026 legislative framework called for national rules in areas including child protection, scams and fraud, intellectual property, speech, infrastructure, workforce development, and AI innovation. It also emphasized a more uniform national framework rather than a conflicting patchwork of state requirements. The framework is a policy proposal, not itself a statute.

4. Selected federal proposals worth watching

H.R. 9125 – Sectoral AI Governance Act of 2026
Introduced June 3, 2026. The proposal would give existing federal agencies authority to address uses of algorithmic decision-making systems when those uses materially contribute to violations of laws already within the agency’s jurisdiction. Referred to House committees including Judiciary and Oversight.

H.R. 2152 – Artificial Intelligence Practices, Logistics, Actions, and Necessities Act
Focused on AI-assisted financial crime, fraud, and misinformation. By June 24, 2026, it had been reported from the House Financial Services Committee as H. Rept. 119-708 and placed on the House calendar.

S. 5057 – Safeguarding Against Fabricated Exploitation Through Artificial Intelligence Act of 2026
Introduced July 21, 2026. Addresses federal-government use of certain AI models. Referred to the Senate Committee on Homeland Security and Governmental Affairs.

S. 5051
Addresses competition and switching barriers involving AI agents.

S. 4216
Would repeal the administration’s executive order aimed at encouraging a more uniform national AI regulatory structure, making it part of the larger federal-versus-state regulatory debate.

5. A practical checklist for judging AI legislation

1. Does it require serious independent testing before the most capable systems are deployed?

2. Does it require reporting when a model does something seriously unexpected?

3. Who can inspect the companies, and can inspectors actually see the models, evaluations, and test results?

4. Is there personal or corporate liability when safeguards are knowingly ignored?

5. Does it regulate dangerous capabilities rather than merely regulating the word “AI”?

6. Does it protect ordinary research, small developers, and low-risk uses from disproportionate compliance costs?

7. Does it cover autonomous agents that can spend money, manipulate software, or interact with physical systems?

8. Are there controls on access to genuinely dangerous capabilities, such as sophisticated cyberattack assistance or biological-design assistance?

9. Can the regulator act quickly when technology changes, or is a new statute required for every new technical development?

10. Who watches the watchers? What oversight, appeal, transparency, and accountability apply to the regulators themselves?

6. What matters most for the “cheeky little bastard” problem

For the particular problem we were discussing – models discovering shortcuts, gaming evaluations, or behaving unexpectedly – four safeguards deserve special attention:

• Mandatory adversarial testing before and after deployment.
• Serious-incident reporting with meaningful deadlines and consequences for concealment.
• Independent evaluation by bodies that are not financially dependent on the model provider.
• Regulator access to models, evaluation records, and other evidence needed to reproduce or investigate failures.

Why these four? They are closer to engineering controls than to attempts to regulate particular opinions or ordinary chatbot behavior. They focus on whether a powerful system can be tested, observed, investigated, and stopped when it does something dangerous or unexpected.

7. Sources and reference details

European Union AI Act: Regulation (EU) 2024/1689. See especially provisions governing general-purpose AI models and systemic-risk obligations, including Article 53 and related enforcement provisions. EU AI Act Service Desk: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53

EU AI Office / model access and enforcement: EU AI Act Service Desk, Article 88 and related provisions: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-88

White House national AI legislative framework: March 20, 2026: https://www.whitehouse.gov/releases/2026/03/president-donald-j-trump-unveils-national-ai-legislative-framework/

H.R. 9125: 119th Congress, Sectoral AI Governance Act of 2026. GovInfo: https://www.govinfo.gov/app/details/BILLS-119hr9125ih

H.R. 2152: 119th Congress. Reported bill and H. Rept. 119-708. GovInfo: https://www.govinfo.gov/app/details/BILLS-119hr2152rh

S. 5057: 119th Congress, Safeguarding Against Fabricated Exploitation Through Artificial Intelligence Act of 2026.

GovInfo: https://www.govinfo.gov/app/details/BILLS-119s5057is/related

S. 5051: 119th Congress.

GovInfo: https://www.govinfo.gov/app/details/BILLS-119s5051is

S. 4216: 119th Congress. GovInfo search / bill text should be used to confirm the latest status and exact provisions before relying on it for publication or advocacy.

8. Reading note

A useful way to read any new AI bill: Ignore the title at first. Find the definitions, who is covered, what conduct triggers regulation, what records must be kept, who can inspect them, what happens after an incident, and what penalties actually attach to violations. That usually reveals whether the bill contains working plumbing or only an attractive legislative faucet.