Policy platform
Our positions.
This page sets out our six commitments, the constitutional argument behind them, and a plain-language guide to the terms. We measure every bill, rule, and court filing against it.
- 01
No licenses for language.
No government may require permission before a model is trained, published, or run.
A language model is text: source code, a file of weights, and the words it produces. Pre-publication review of text is a prior restraint, the one thing the First Amendment forbids most clearly. We oppose any licensing, registration, or approval regime for model developers, whether federal or state.
- 02
No compute thresholds.
Rules that trigger at a number of floating-point operations regulate speech by its length.
A threshold on training compute measures how much work went into a model. It says nothing about what the model does or whom it harms. Laws that attach duties above 10^26 operations place the heaviest burden on the most capable expression, and the First Amendment does not allow speech to lose protection because there is more of it.
- 03
Open weights are protected publication.
Releasing model weights is publishing, and the right to publish does not depend on the publisher’s size.
The government classified encryption as a munition until the courts and the market made that position untenable. Model weights are on the same path. We oppose export controls, distribution bans, and liability regimes aimed at the act of publishing weights.
- 04
Liability follows the human.
People who misuse a model answer for it under the laws that already exist.
Fraud, defamation, harassment, and weapons offenses are already crimes and torts. We support enforcing those laws against the people who commit them. We oppose transferring that responsibility to the author of a general-purpose tool, which would make every publisher a guarantor of every reader.
- 05
If it is a weapon, it is an arm.
A state that calls a model a dual-use weapon has named the amendment that protects it.
Heller extends the Second Amendment to all bearable instruments, including those unknown at the founding. Caetano confirmed that electronic weapons qualify. If a model is a weapon in common use, keeping and bearing it is a right, and the burden to justify restrictions falls on the state.
- 06
One national rule.
A patchwork of state licensing regimes is the most restrictive regime of all.
We support federal action that protects the right to build and publish models nationwide. We oppose federal action that merely replaces fifty state licenses with one. Preemption is worth having only if what it preempts is not reproduced at the national level.
The thesis
Two amendments, one object.
Governments that want to license, cap, or forbid language models must first decide what a model is. There are only two honest answers, and the Bill of Rights has a clause for each.
“Congress shall make no law … abridging the freedom of speech, or of the press.”
If a model is language, it is speech.
Source code, model weights, and model output are all text. Federal courts have held that computer code is protected expression. The Supreme Court has held that the creation and dissemination of information is speech, and that new media get the same protection as old ones.
“… the right of the people to keep and bear Arms, shall not be infringed.”
If a model is a weapon, it is an arm.
The state has repeatedly classified software as a munition: encryption until 1996, gun-design files under the arms-export regime, and frontier model weights under export control in 2025. Heller says the Second Amendment reaches all bearable instruments of offense or defense, even those unknown at the founding.
First Amendment
Code is language.
A language model is written in a programming language, trained on human writing, stored as a file of numbers that can be printed on paper, and used by typing words and reading words back. At no point in that chain does anything exist except language.
The courts have already been here. During the Crypto Wars, the government argued that encryption source code was a regulable device rather than speech. The Sixth Circuit rejected that argument in Junger v. Daley, and the Second Circuit agreed in Universal City Studios v. Corley that code does not lose protection “simply because it is expressed in the language of computer code.”
“Because computer source code is an expressive means for the exchange of information and ideas about computer programming, we hold that it is protected by the First Amendment.”Junger v. Daley, 209 F.3d 481, 484–85 (6th Cir. 2000)
The Supreme Court has added two more planks. In Sorrell v. IMS Health it held that “the creation and dissemination of information are speech within the meaning of the First Amendment.” In Brown v. Entertainment Merchants it held that a new technology, there video games, gets the same protection as books, plays, and films, because “the basic principles of freedom of speech and the press … do not vary” with technology.
Every current proposal to govern models reaches for one of three tools. A compute threshold draws a line at how much computing went into training a model and switches on duties above it. A licensing regime requires the government’s permission before a model may be released. A mandatory safety review requires the model to pass a government-set test, or file a report, first.
Each of the three is a rule that a particular kind of text may not be published until the government is satisfied. The law calls that a prior restraint, and it is the one thing the First Amendment forbids most clearly. The burden to justify it falls on the government, not on the author.
Second Amendment
Language can be weaponized.
Regulators want to license models because a model can write exploit code, plan an attack, or persuade at scale. The premise of every AI safety bill is that a language model is a dual-use weapon.
Take that premise seriously. The United States has already treated software as a munition three times. Encryption sat on Category XIII of the United States Munitions List until Executive Order 13026 moved it in 1996. The State Department treated 3D-printed gun files as arms-export “technical data,” the basis of Defense Distributed v. Department of State. And in January 2025 the Commerce Department created ECCN 4E091 to control the weights of large models as a national-security item.
“The Second Amendment extends, prima facie, to all instruments that constitute bearable arms, even those that were not in existence at the time of the founding.”District of Columbia v. Heller, 554 U.S. 570, 582 (2008)
Heller defined “arms” from the founding-era dictionaries: “weapons of offence, or armour of defence,” and “any thing that a man … takes into his hands … to cast at or strike another.” Caetano applied that to stun guns, with Justice Alito noting that electronic weapons are no more exempt from the Second Amendment “than electronic communications are exempt from the First.” Bruen and Rahimi require the government to find a historical tradition for any burden, and the founding generation never licensed the printing press because it could print sedition.
A model that fits in a pocket, runs on a laptop, and is used by tens of millions of people is an instrument in common use. If the state insists it is a weapon, the state has named the amendment that protects it.
The pincer
Either way, it is protected.
The two arguments work together. A regulator must say either that a model is expression, in which case licensing it is a prior restraint, or that a model is an instrument of force, in which case keeping and bearing it is a right. The only way out is to argue that a language model is neither language nor a weapon, and nobody who wants to regulate it believes that.
The founding generation had already made this choice about the printing press. Dangerous knowledge was allowed to circulate because licensing the press was worse. The First and Second Amendments sit next to each other for that reason.
We do not claim the question is settled. Our claim is that the burden is on the state, and that the state has not met it.
Plain terms
What these words mean.
The argument above uses a handful of legal and technical terms. Here is each one in ordinary language, with the statute or opinion where it actually appears.
- Compute threshold
A line drawn at how much computing was used to train a model. Training is measured in floating-point operations, the basic arithmetic steps a computer performs. Cross the line and extra legal duties switch on; stay under it and they do not.
The closest everyday comparison is a law that applies only to books over 500 pages. It says nothing about what the book contains.
- California’s SB 53 defines a “frontier model” as one trained with more than 10^26 operations (a 1 followed by 26 zeros) and attaches reporting duties above that line. Cal. Bus. & Prof. Code § 22757.11 (SB 53, 2025)
- New York’s RAISE Act uses the same 10^26 line to define the developers it regulates. N.Y. Gen. Bus. Law art. 44-B
- The number comes from a 2023 executive order that required developers to report any training run above 10^26 operations. That order was rescinded in January 2025, but its line survived in the state laws. Executive Order 14110 § 4.2(b)(i) (Oct. 30, 2023)
- Europe draws its line lower, at 10^25 operations, and presumes any model above it carries “systemic risk.” Regulation (EU) 2024/1689, art. 51(2)
Our position. The amount of computing behind a model says nothing about what it does or whom it harms. A rule keyed to it burdens the most capable expression most heavily, and the First Amendment does not work that way.
- Licensing regime
Any rule that requires the government’s permission before you may do something. For a model, that means a license, a registration the state can refuse, or an official sign-off before the model may be trained, published, or run.
- This is what a printing license was. English printers needed government permission for each book until the Licensing of the Press Act lapsed in 1695, and Milton’s Areopagitica (1644) was written against the Licensing Order of 1643. The First Amendment was drafted with that history in view. Lovell v. City of Griffin, 303 U.S. 444, 451–52 (1938)
- In 1931 the Supreme Court struck down a Minnesota law that let a judge shut a newspaper as a “nuisance,” calling the scheme “the essence of censorship.” Near v. Minnesota, 283 U.S. 697, 713 (1931)
- A Senate bill would bar deployment of covered models until they complete a Department of Energy evaluation program. S. 2938 (119th Cong.), Artificial Intelligence Risk Evaluation Act
Our position. Permission-before-publication is the one thing the First Amendment forbids most clearly. We oppose licensing, registration, or approval regimes for model developers, federal or state.
- Mandatory safety review
A rule that a model must pass a government-set test, or file a report with an agency, before it may be released. The review happens first; release is conditioned on it.
- SB 53 requires large developers to publish a “frontier AI framework” and file a transparency report before deploying a new frontier model. Cal. Bus. & Prof. Code § 22757.12 (SB 53, 2025)
- The vetoed SB 1047 would have required safety certifications before training and third-party audits after. Gov. Newsom’s veto message, SB 1047 (Sept. 29, 2024)
- Cars are crash-tested before sale, and the comparison is often made. The difference is that a car is not speech. The Supreme Court has refused to create new categories of unprotected speech by weighing costs against benefits, and it has held that a new medium gets the same protection as the old ones. United States v. Stevens, 559 U.S. 460, 470 (2010); Brown v. Entertainment Merchants Ass’n, 564 U.S. 786, 790 (2011)
Our position. A review that must be passed before publication is a prior restraint, whatever name the statute gives it. We support enforcement against harmful conduct after the fact.
- Prior restraint
A government order that stops speech before it happens, instead of punishing it afterward. The government may sue a newspaper for libel after it prints. It may not stop the presses first.
- The Supreme Court calls prior restraints “the most serious and the least tolerable infringement on First Amendment rights.” Nebraska Press Ass’n v. Stuart, 427 U.S. 539, 559 (1976)
- Any system of prior restraint “comes to this Court bearing a heavy presumption against its constitutional validity.” The Court applied that rule to block the Nixon administration from stopping publication of the Pentagon Papers, even on a claim of national security. New York Times Co. v. United States, 403 U.S. 713, 714 (1971), quoting Bantam Books v. Sullivan, 372 U.S. 58, 70 (1963)
Our position. A compute threshold, a license, and a pre-release review each stop a particular kind of text from being published until the government is satisfied. Each is a prior restraint, and the burden of justifying it sits with the government.
- Model weights and open weights
A trained model is a very large list of numbers, called weights, that training produced. Releasing a model “open weights” means publishing that file so anyone can download, inspect, and run it on their own computer.
- In January 2025 the Commerce Department created an export-control category for the weights of large closed models, treating a file of numbers as a controlled item. The rule was rescinded that May. 90 Fed. Reg. 4544 (Jan. 15, 2025), ECCN 4E091
- The same move was made against encryption software in the 1990s. The Sixth Circuit held that source code is protected speech. Junger v. Daley, 209 F.3d 481, 484–85 (6th Cir. 2000)
Our position. Publishing weights is publishing. The right to publish does not depend on the size of the publisher or the size of the file.
- Frontier model
The label laws give to the largest models. It is not a technical category. Every statute that uses it defines “frontier” by a compute threshold.
- SB 53’s title is the Transparency in Frontier Artificial Intelligence Act, and its definition of a frontier model is the 10^26 line. Cal. Bus. & Prof. Code § 22757.11
Our position. The word suggests danger. The definition in every statute measures size.
- Compelled speech
The government forcing you to say or publish something you would not otherwise say. The First Amendment protects the right not to speak as well as the right to speak.
- The Supreme Court struck down a California law that required licensed clinics to post government-drafted notices, because the state cannot script what a speaker must say. National Institute of Family & Life Advocates v. Becerra, 585 U.S. 755 (2018)
- “No official, high or petty, can prescribe what shall be orthodox in politics, nationalism, religion, or other matters of opinion.” West Virginia State Bd. of Educ. v. Barnette, 319 U.S. 624, 642 (1943)
Our position. A mandatory “safety framework” or transparency report is the state writing part of the publication. We treat those requirements as compelled speech tied to the size of a publication.
- Preemption
When a federal law overrides state laws on the same subject, so there is one national rule instead of fifty. The Constitution makes federal law supreme where the two conflict.
- Article VI makes the Constitution and federal laws “the supreme Law of the Land,” binding on state judges. U.S. Const. art. VI, cl. 2
- A December 2025 executive order directs the Justice Department to challenge state AI laws and asks Congress for a preemption statute. Executive Order 14365, 90 Fed. Reg. 58499 (Dec. 16, 2025)
Our position. We support preemption that protects the right to build and publish models nationwide. We oppose preemption that merely replaces fifty state licenses with one federal license.
Take action
Legislators count letters.
Sign the declaration, write to your representatives with our template, and get one email when a bill moves.