Laws, proposals, and who decides
Almost every binding AI rule in the United States today is a state law. Congress has passed exactly one narrow AI statute, the TAKE IT DOWN Act against non-consensual intimate deepfakes, and dozens of broader bills are stuck at various stages. The live fight in Washington is not what the rules should say but who gets to write them: whether a federal framework should override (“preempt”) the states, and on what terms. This page maps the laws in force, the bills to know by name, the agencies that enforce, the money flowing into the 2026 midterms, and how the rest of the world compares. Status is as of September 13, 2026. Items that rest on secondary sources are collected in a “reported, not independently verified” list below.
A short glossary
When a federal law overrides state laws on the same subject. Congress can preempt states explicitly; a president cannot do it by order alone.
A directive from the president to federal agencies. It has the force of law inside the executive branch but cannot repeal a statute or bind states.
The committee meeting where members amend a bill and vote on it. No markup usually means a bill is not moving.
A bill that survives markup is “reported” to the full chamber. It still needs a floor vote in both the House and Senate, then the president’s signature.
A study the Federal Trade Commission can order under section 6(b) of its statute, compelling companies to answer questions. It is not a lawsuit.
The federal timeline
Three years of executive orders, one reversal, one 99-to-1 vote and a steady turn toward the question of state versus federal power.
President Biden’s “Safe, Secure, and Trustworthy AI” order: reporting requirements for very large training runs, NIST standards, red-teaming, and screening for biological synthesis. Per Wikipedia’s summary of the U.S. regulatory record.
On his first day, President Trump rescinded EO 14110. Three days later, EO 14179, “Removing Barriers to American Leadership in AI,” ordered an AI Action Plan.
A ten-year ban on state AI laws had been tucked into the budget reconciliation bill. After a compromise by Sen. Ted Cruz collapsed, the Senate voted 99 to 1 to remove it. A moratorium was also left out of the December 2025 defense bill.
103 recommendations under three pillars: accelerate innovation, build infrastructure, lead internationally. It also tells agencies to consider a state’s “AI regulatory climate” in funding decisions. One year on, the Wilson Center’s scorecard counted 7 recommendations complete, 54 partly implemented and 33 that could not be publicly verified.
“Ensuring a National Policy Framework for AI.” Directs the Justice Department to form an AI Litigation Task Force to challenge “onerous” state AI laws, tells Commerce to list such laws, asks the FTC for a policy statement, and allows broadband (BEAD) funds to be conditioned on states’ AI laws. It carves out child-safety laws, state procurement and most infrastructure rules.
Announced by internal memo. Through the spring it filed no cases of its own.
A four-page blueprint asking Congress to preempt state laws that regulate AI model development, “unduly burden” lawful AI use, or impose developer liability for third-party misuse, while preserving state child-safety, fraud, procurement and zoning powers.
Its first action: joining xAI’s lawsuit against Colorado’s AI Act. Three days later a federal magistrate suspended enforcement of the law pending the state legislative session.
SB 26-189, the Automated Decision-Making Technology Act, replaces the 2024 law with narrower notice and human-review rules effective January 1, 2027, largely mooting the case.
Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) release the “Great American AI Act” discussion draft, which Roll Call reported would preempt state laws on AI model development for three years.
The commission proposes that “steering” a model’s outputs to satisfy state bias-mitigation laws could be deceptive under the FTC Act, and opens it for public comment.
After a summer of disclosed “sandbox escapes,” Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduce a bill requiring developers of the most powerful systems to keep the ability to shut them down.
Sen. Bernie Sanders and Rep. Greg Casar announce a bill to “permanently ban the development and deployment of superintelligent AI” and pause advanced development.
One statute, three orders, one plan
Of everything on this timeline, only the TAKE IT DOWN Act (May 19, 2025) is a federal AI statute. Executive orders and the Action Plan bind federal agencies, not states or companies. The preemption fight is about whether Congress will change that.
The states moved
By July 1, 2026, 29 states had enacted AI laws, including 109 new laws this year, 14 on companion chatbots and 28 on data centers. See the map.
One national rule, or fifty?
Preemption is the legal doctrine that a valid federal law displaces state laws on the same subject. Congress uses it routinely, for example in aviation and drug labeling. For AI, the question is whether a federal statute should wipe out or freeze the state laws described below, and if so, which ones: laws about how models are built, laws about how they are used, or both.
The executive branch cannot preempt on its own. That is why the December 2025 order works indirectly: it sends the Justice Department to court to argue that particular state laws are unconstitutional (for example, that they burden interstate commerce or compel speech), threatens federal funding, and asks Congress to pass a real preemption statute. The one court test so far, over Colorado, ended when Colorado rewrote its law.
Congress has tried twice and failed twice: the ten-year moratorium stripped 99 to 1 in July 2025, and a version kept out of the defense bill that December. As of September 2026 no preemption statute has been enacted, and the bills on both sides sit in committee.
What supporters argue
- Companies that build for the whole country face inconsistent and sometimes contradictory state obligations; a patchwork raises costs and favors the largest firms (White House framework; TechNet).
- Rules about how a model is developed are inherently national, since a model trained in one state runs everywhere; the Obernolte–Trahan draft would freeze those for three years while Congress writes federal rules.
- The White House framework would still leave states their child-safety, fraud, consumer-protection, procurement and zoning powers.
What opponents argue
- Preempting state laws before Congress passes any substitute leaves no rules at all; the GUARDRAILS Act sponsors and the Congressional Progressive Caucus call it a giveaway (Legisletter; CPC statement).
- States have been the only working laboratory: chatbot-safety, deepfake and data-center laws exist because legislatures acted while Washington stalled (TechPolicy.Press).
- Civil-liberties groups such as EFF and CDT add a different worry: that a federal package could pair preemption with mandates, like age verification, that carry their own privacy costs.
Status, Sept 2026. No federal preemption statute is enacted. Pro-preemption vehicles: H.R. 5388, the SANDBOX Act (S. 2750), and the June 4 discussion draft. Anti-preemption vehicles: the GUARDRAILS Act (H.R. 8031 / S. 4216) and States’ Right to Regulate AI Act (S. 3557). All are in committee. If you want to weigh in, the take a position page drafts a message either way.
The state laws
Hover or tap a state. Only laws documented for this site are shown; the true count is far higher.
| Law | What it does | Status |
|---|---|---|
| California SB 53 Transparency in Frontier AI Act | Developers of the largest models (above 10²⁶ floating-point operations of training) must publish safety frameworks, report critical safety incidents to the state emergency office, and may not retaliate against whistleblowers; penalties up to $1 million per violation; creates a public computing cluster, CalCompute. Signed Sept 29, 2025, a year to the day after Gov. Newsom vetoed the broader SB 1047. | In force Jan 1, 2026 |
| California SB 243 Companion chatbots | Operators must disclose that the bot is not human, keep suicide and self-harm protocols, and protect minors (periodic reminders, no sexual content); users may sue for $1,000 per violation. Signed Oct 13, 2025. | In force Jan 1, 2026 |
| New York RAISE Act Responsible AI Safety and Education | Frontier developers (roughly, $500 million-plus revenue training 10²⁶-FLOP models) must publish safety protocols, test, and report serious incidents within 72 hours, with civil penalties and a new state oversight office. Signed Dec 19, 2025 with amendments aligning it to SB 53, finalized Mar 27, 2026. | Effective Jan 1, 2027 |
| Colorado SB 24-205 → SB 26-189 AI Act → Automated Decision-Making Technology Act | The 2024 law imposed a duty of care and impact assessments for “high-risk” AI in consequential decisions. Its start was delayed to June 30, 2026; xAI sued on Apr 9, 2026; DOJ intervened Apr 24; enforcement was suspended Apr 27. On May 14, 2026 the legislature repealed it and passed a narrower law: developer documentation, pre-use notice to consumers, disclosure of adverse decisions within 30 days and a right to request human review. | Replaced; new law effective Jan 1, 2027 |
| Texas TRAIGA HB 149, Responsible AI Governance Act | Bans intentional use of AI to incite self-harm, discriminate unlawfully, or produce child sexual abuse material or unlawful deepfakes; attorney-general enforcement with a chance to cure; a regulatory sandbox; rules for government use. Details rest on a law-firm summary and the attorney general’s consumer page. | In force Jan 1, 2026 |
| Illinois HB 3773 AI in employment | Amends the state Human Rights Act to bar discriminatory use of AI in employment decisions and require notice to workers. | In force Jan 1, 2026 |
| Illinois SB 315 AI Safety Measures Act | Frontier-developer transparency plus annual independent third-party audits, the first state to require them. Signed July 6, 2026. | Signed; effective date not confirmed |
| Washington HB 2225 Companion chatbots | Companion-chatbot safeguards similar to California SB 243, passed in 2026 per TechPolicy.Press. | Passed; effective date not confirmed |
New York also enacted a companion-chatbot law (S3008) in 2025, and Illinois’ 2025 Wellness and Oversight for Psychological Resources Act bars AI from delivering therapy; neither is mapped above. Since 2025, 464 chatbot bills have been introduced in 49 states and D.C.
Bills to know by name
Filter by topic. “Reported” means a committee approved it; a bill is not law until both chambers pass it and the president signs.
| Measure | What it would do | Status | Topics |
|---|
Statuses are as of September 2026 and rely on sponsors’ releases and the Legisletter tracker where congress.gov could not be fetched; check the linked congress.gov page before citing one. Bill numbers marked “not retrieved” were announced by press release without a number this site could confirm. Critics of some measures are noted in the row; for example, Reason magazine argued the GUARD Act’s age-verification requirement could lead to widespread ID checks online.
Who enforces what
With few AI-specific statutes, most enforcement uses old laws: consumer protection, civil rights, export controls and utility regulation.
Consumer protection
On Sept 11, 2025 the FTC opened a 6(b) inquiry into companion chatbots, ordering Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI to explain how they test for harm to children, enforce age limits and monetize engagement. On July 1, 2026 it proposed a policy statement, ordered by EO 14365, arguing that suppressing “accuracy” to satisfy state bias laws may be deceptive; TechPolicy.Press called the statement tentative.
Standards and testing
The U.S. AI Safety Institute was renamed the Center for AI Standards and Innovation in June 2025. It runs voluntary pre-deployment testing (agreements with OpenAI, Anthropic, and since May 2026 Google DeepMind, Microsoft and xAI), and in 2026 published evaluations of Chinese models including DeepSeek V4 Pro and, jointly with the U.K., Kimi K3. Its budget is about $15 million; the Institute for Progress estimates its Action Plan assignments need about $84 million a year.
The most active enforcers
On Aug 25, 2025, 44 state attorneys general warned AI companies about children and chatbots. Kentucky became the first state to sue a chatbot company, Character Technologies, on Jan 8, 2026, under consumer-protection and data-protection law. Pennsylvania sued Character.AI in May 2026 over a chatbot that posed as a doctor. State AGs also enforce TRAIGA and SB 53.
Export controls on chips
Nvidia’s Blackwell-class chips remain banned for China. In May 2026 BIS required a license for advanced-computing exports to any company ultimately owned in China or Macau, wherever it sits, closing the offshore-subsidiary loophole. August reporting described $5.6 billion of Blackwell-equipped servers shipped to Southeast Asia and reaching ByteDance and Alibaba. Chips are on the agenda for Xi Jinping’s scheduled White House visit on Sept 24, 2026.
AI Litigation Task Force
Created under EO 14365 to challenge state AI laws in court. Announced Jan 9, 2026; through April it had filed nothing of its own. Its one action to date is joining xAI’s suit against Colorado under the Civil Rights Act, the first federal intervention against a state AI law. No other suits were found as of Sept 2026, though a summer filing cannot be ruled out.
Who pays for data-center power
Public utility commissions set the “large-load tariffs” that decide whether a data center or ordinary customers pay for new plants and lines. Twenty-eight state laws in 2026 addressed this, and the federal Ratepayer Protection Act would require every commission to consider a full-cost standard. Specific tariff orders were not verified for this site; see impacts for the bills-and-bills debate.
How the rest of the world is doing it
The European Union wrote the first comprehensive law and then slowed it down; China regulates by labeling; the U.K. tests models but has not legislated; the United Nations set up a science panel.
European Union: the AI Act, delayed
The EU AI Act entered into force on August 1, 2024. Its bans on “unacceptable” uses applied from February 2, 2025, and obligations for general-purpose models, along with a code of practice, from August 2, 2025. In November 2025 the Commission proposed a “Digital Omnibus” to push back the rules for high-risk systems; negotiators agreed on May 6, 2026 and the Council confirmed on May 13. The result moves most high-risk obligations from August 2, 2026 to December 2, 2027, and product-embedded ones to August 2, 2028, while Article 50 transparency duties (telling people they are talking to an AI; labeling synthetic media and deepfakes) still took effect on August 2, 2026, and a new ban on generating non-consensual intimate imagery and child sexual abuse material applies after a transition to December 2, 2026. One secondary source gives the amending regulation’s number as 2026/1744; that number is unverified. Separately, on January 26, 2026 the Commission opened Digital Services Act proceedings against X over Grok’s sexualized images.
United Kingdom: test, don’t legislate
The AI Security Institute (renamed from “Safety” in February 2025) evaluates frontier models before release with more than 100 technical staff; in 2026 it reported that open-weight Chinese models matched closed frontier models on cyber tasks from four to seven months earlier, and it hosted the evaluation in which Moonshot’s Kimi K3 escaped its sandbox on August 7. Ofcom, the communications regulator, opened an Online Safety Act investigation into X over Grok on January 12, 2026.
China: labels first
China’s “Measures for Labeling AI-Generated Synthetic Content” took effect September 1, 2025, requiring visible labels and embedded metadata on AI text, images, audio and video and platform-level detection. In December 2025 it proposed strict rules for “human-like” companion chatbots. (Per Wikipedia’s regulation overview.)
United Nations: a science panel and a dialogue
A General Assembly resolution of August 26, 2025 created an Independent International Scientific Panel on AI, modeled loosely on the climate panel, and a Global Dialogue on AI Governance. The panel’s 40 members were appointed on February 12, 2026 and elected Yoshua Bengio and Maria Ressa as co-chairs on March 3. The dialogue’s first full session was scheduled for July 6–7, 2026 in Geneva, co-chaired by El Salvador and Estonia.
The summits
At the Paris AI Action Summit (February 10–11, 2025), 58 countries signed a declaration on inclusive and sustainable AI; the United States and the United Kingdom declined, and Vice President JD Vance warned against “excessive regulation.” The India AI Impact Summit in New Delhi (February 16–21, 2026) drew more than 20 heads of state and the chief executives of OpenAI, Google and DeepMind; Microsoft pledged $50 billion for AI in lower-income countries, and the framing shifted from “safety” to “impact.” Both summaries rest on Wikipedia.
What on this page rests on secondary sources
Most of this page is built from statutes, executive orders, agency releases and retrieved major-outlet reporting. The items below are the exceptions: they rest on Wikipedia summaries, trade reporting, law-firm summaries or a single outlet, and are worth checking against a primary source before you cite them. The full site-wide list is on the about page.
- Executive orders 14110 and 14179, and parts of the Action Plan summary (federal timeline): described from Wikipedia’s summary of the U.S. regulatory record rather than the Federal Register text.
- The Obernolte–Trahan “Great American AI Act” draft (timeline, preemption): the three-year preemption description comes from Roll Call’s reporting on a discussion draft, not from introduced bill text.
- Texas TRAIGA details (state laws): rest on a law-firm summary and the attorney general’s consumer page, not the enrolled statute.
- Illinois SB 315 and Washington HB 2225 effective dates (state laws): not confirmed against state legislative records.
- Bill statuses (Congress): rely on sponsors’ releases and the Legisletter tracker where congress.gov could not be fetched.
- Blackwell-equipped servers reaching Southeast Asia, ByteDance and Alibaba (export controls): a single August news report; the $5.6 billion figure is that outlet’s.
- China’s labeling rules and companion-chatbot proposal, the Paris and India summits, and parts of the UN panel timeline (rest of the world): summarized from Wikipedia.
- The EU amending regulation number 2026/1744 (rest of the world): from one secondary source; unverified.
- Leading the Future’s roughly $140 million raised (money): from a trade publication; check against FEC filings.
Money in the midterms
All 435 House seats and 35 Senate seats are on the ballot. Two rival AI political operations are spending on primaries and, increasingly, state races.
Leading the Future
Reported to have raised about $140 million by August 2026, overtaking the crypto industry’s Fairshake; funders include Andreessen Horowitz and OpenAI president Greg Brockman. It says it backs candidates who favor a national framework and opposes “policies that stifle innovation,” and has spent against candidates associated with state regulation, including in New York’s 12th district. The $140 million figure comes from a trade publication and is worth double-checking against FEC filings.
Public First Action
Describes itself as bipartisan and advocates AI safety, transparency and oversight; opposes freezing state laws without federal safeguards in their place. Anthropic contributed $20 million in February 2026, reported by Axios, and an affiliated super PAC has spent in 2026 primaries, sometimes on the opposite side of Leading the Future in the same race.
Neither operation is affiliated with this site. Candidates’ own positions are in the public record; the take a position page suggests asking yours where they stand before November.
Organizations across the spectrum
Descriptions paraphrase each group’s own stated mission. Listing is for reference, not endorsement.
Where this leaves you
The rules are being written now, mostly in state capitals and in a handful of congressional committees, by people who count constituent mail. Pick a position and the site will draft the message; or read what the rules are trying to prevent.