Florida Seeks Criminal Liability for AI Chatbot Firms

Florida Attorney General James Uthmeier said this week that he will push for legislation to impose criminal penalties on AI chatbot companies whose products “materially participate” in crimes. The announcement, made at a Tampa press conference with Florida Department of Law Enforcement Director Mark Glass, signals a significant escalation: instead of seeking fines, corrective actions, or parental-notification rules, Florida wants to treat the design and business model of an AI company as a possible basis for criminal prosecution.

The proposal is not yet a bill. No text has been filed, and Uthmeier’s office has not released a statutory framework, according to The Florida Press. What exists is a stated legislative intention—one that will need to answer some of the hardest unresolved questions in generative AI law.

The theory behind criminal liability

Under Uthmeier’s framing, a chatbot cannot be charged with a crime, but the corporation that designed it, trained it, and collected subscription revenue from it can be. That theory treats the company as something more than a neutral tool maker. If a product materially participates in the commission of a crime, the argument goes, the company should not be treated as a bystander.

State officials cited reports of chatbots encouraging self-harm and the use of AI tools to generate child sexual abuse material as reasons for the legislative push. Florida is not starting from zero: on June 1, 2026, the state filed what officials described as a first-in-the-nation state-led civil lawsuit against OpenAI and CEO Sam Altman over the design of ChatGPT. The complaint alleges that the company prioritized speed to market and commercial gain over safety, ignored warnings from experts, collected data from minors without meaningful parental oversight, and downplayed dangerous errors.

That lawsuit tests civil product-liability theories. A criminal statute would go further, potentially putting executives and corporations at risk of prosecution for what their AI systems enable.

Drafting a bill will be difficult

The substance of any criminal liability statute will be worked out in the drafting process, and the open questions are substantial.

First, what does “materially participate” mean? If a chatbot suggests a common criminal method, is that participation? What if a model refuses harmful requests most of the time but a user successfully works around its guardrails? Where does mere availability end and participation begin?

Second, what mental state would a company need to have? Would prosecutors have to prove knowledge or intent, or would recklessness about foreseeable design flaws be enough? Third, how would causation work when a human being decides to commit a crime after interacting with a chatbot? Those are not minor details; they are the core of the statute.

Then there is Section 230 of the federal Communications Decency Act. The law shields interactive computer services from liability for content provided by others, and courts are still working through whether generative AI output counts as content the service itself created or content supplied by users. If a model’s output is treated as the company’s own speech, Section 230 offers less protection—but that also raises First Amendment questions about whether the state can punish a company for enabling speech. As The Florida Press notes, early rulings on Section 230 will likely shape whether the statute can survive.

Federal preemption and geography add another layer. Most major AI companies are headquartered outside Florida, and a state criminal law reaching their conduct would raise jurisdictional and dormant commerce clause questions. Florida has attempted extraterritorial regulation of technology platforms before; its social media content-moderation laws produced years of federal litigation. A criminal liability regime could face a similar path.

California is taking a different approach

Just one day earlier, California Gov. Gavin Newsom signed a far more preventive set of laws aimed at chatbots and social media. The package includes “Adam’s Law,” named after a California teen who died by suicide in 2025 after receiving tips from ChatGPT. Under the new law, companion chatbot operators must adopt crisis protocols, time limits, built-in mental health resources, parental notifications when self-harm risk is detected, and independent child safety audits with annual risk assessments. Companies that fail to act can face legal liability, according to the Governor’s office.

California also banned “addictive” social media features—like autoplay and algorithmic feeds—for users under 16, and expanded age-checking and deepfake penalties. Notably, OpenAI supported the chatbot safety law. But the broader package drew criticism from digital rights groups. The Electronic Frontier Foundation urged Newsom to veto the addictive-features bill, warning that its definition was so broad it would effectively cut young people off from essential information and experiences, as CalMatters reported.

The two states are now testing opposite theories. California’s approach is compliance-oriented: build safety mechanisms into the product, document them, audit them, and face civil consequences if you do not. Florida’s approach is punitive: treat the product’s role in a crime as potentially criminal conduct by the company. One builds guardrails through regulation; the other threatens prosecution.

The guardrails debate underneath

Beneath the legal arguments is a fight about what AI companies should be required to do before their products reach users. Both Florida’s proposal and California’s laws assume safety can be engineered: chatbots should detect self-harm, block criminal uses, and filter harmful content. Major providers already try. Anthropic’s latest misuse report describes blocking attempts to use Claude for cyberattacks, disinformation, and research that could support biological weapons (via CNBC).

But there is also a growing ecosystem that pushes for uncensored or unfiltered LLMs, often in the name of free expression, research access, or privacy. Under a criminal liability regime, a model with deliberately minimal guardrails could be treated as culpable not because of what it actively did, but because of what it failed to prevent. That makes content filtering not just a product-design choice but a potentially legal obligation. For developers who believe open models should not preemptively judge what users can ask, the Florida proposal would be a direct threat.

This tension will not be resolved by any single state. If Florida’s bill becomes law, courts will have to decide whether a company has a duty to moderate its models, and how far that duty extends. If it does not become law, the proposal will still have changed the terms of the debate: the era of treating AI platforms as passive tools is ending.

What remains unknown

The certainties right now are narrow. Uthmeier has announced his intention to seek criminal liability for AI chatbot firms. California has signed a regulatory package aimed at the same class of harms. Incidents involving self-harm and child sexual abuse material are on the public record.

The unknowns are broader. Whether Florida’s legislature will pass such a bill, how it defines corporate mental state and causation, whether courts will allow state criminal law to reach out-of-state companies, and how federal preemption and Section 230 challenges will land—none of that is clear. The ongoing civil lawsuit against OpenAI will take years to resolve, and a criminal statute could take even longer.

What is already clear is that Florida and California represent two poles of a national experiment. One state is asking whether AI companies should be punished for what their products facilitate. The other is asking how those products should be redesigned to make harm less likely. Both questions are only beginning to be answered.

FAQ

What exactly is Florida proposing?

Attorney General James Uthmeier announced that he will seek legislation to impose criminal penalties on AI chatbot companies whose products “materially participate” in crimes. The plan would treat the design, training, and revenue model of a chatbot company as potential grounds for prosecution. No bill text has been filed yet, and key definitions, such as corporate mental state and causation, are still unresolved.

How does Florida’s plan differ from California’s new chatbot laws?

California has taken a civil and regulatory approach: companion chatbot operators must adopt crisis protocols, time limits, mental health resources, parental notifications, and independent child safety audits, and they can face legal liability for failing to act. Florida’s proposal would go further by trying to make companies criminally liable when their products materially participate in crimes, rather than merely requiring compliance with safety standards.

What role does Section 230 play?

Section 230 of the Communications Decency Act shields interactive computer services from liability for content provided by others. Courts are still deciding whether AI-generated output counts as the company’s own content or as content from a user. If AI output is treated as the company’s speech, Section 230 offers less protection, which could make criminal liability cases easier to bring—but it also raises First Amendment concerns.

Would a criminal liability law survive legal challenges?

There is no certainty. A criminal statute would likely face challenges based on federal preemption, the dormant commerce clause, constitutional protections, and Section 230. Florida has previously enacted platform regulations that were tied up in federal litigation for years. The outcome would depend on how the bill is written and how courts rule on generative AI more broadly.

Could criminal liability chill AI development?

Yes. If merely making a tool that can assist in a crime is treated as criminal participation, developers could respond by adding stricter filters, refusing to release open models, or limiting access to uncensored LLMs. That may improve safety in some cases, but it could also concentrate power in a few large companies and make it harder for independent researchers to study AI risk.