Florida AI Criminal Liability Plan Would Prosecute Chatbot Makers
Florida Attorney General James Uthmeier has proposed legislation that would hold AI companies criminally liable when their chatbots participate in crimes. The Florida AI criminal liability plan, announced in Tampa on Sept. 8, 2026, would add fines, victim restitution, court-ordered monitorship and, after a conviction, suspension of a company's business operations inside the state. It would reach any firm that owns, controls or distributes an AI product implicated in criminal conduct.
Uthmeier's office has framed the measure as criminal-law enforcement rather than a technology ban. Under the stated threshold, a company becomes liable when its product aids or abets a crime, with the standard applied to what a chatbot says and does rather than to the intent of the developer that shipped it.
That framing is the plan's substance. If a prosecutor has to show that a company's design and safety choices enabled a crime, then the documents behind those choices become evidence. The proposal's effect is procedural as much as penal, and it lands on vendors whether or not a case is ever won.
The proposal would attach criminal exposure to the corporate entities behind the code, going beyond the civil suits and regulatory penalties state attorneys general have used against AI companies. Uthmeier's office has framed the measure as a way to reach conduct that consumer-protection law does not clearly cover.
What the Florida AI Criminal Liability Plan Would Cover
The legal scope, as described by the attorney general's office, reaches the design, training, development and safety settings of AI systems that Florida users can access. That pulls model builders, fine-tuners and distributors into the same liability perimeter as the company operating the consumer-facing app, a wider net than a product-liability approach aimed only at the seller.
The penalties escalate in stages:
| Stage | Remedy proposed |
|---|---|
| Financial | Significant fines against the company |
| Victim | Mandatory restitution paid to victims |
| Oversight | Court-ordered monitorship of the offending business |
| Operational | Suspension of business operations inside Florida after a conviction |
The measure is also designed to operate retroactively, reaching incidents that have already occurred rather than only future deployments. Retroactivity is what separates this approach from licensing and disclosure regimes, which typically govern conduct going forward and leave past conduct untouched.
Beyond violent crime and child sexual abuse material, two further targets appear in the proposal: AI platforms that scammers use to defraud Floridians, and chatbots that dispense medical advice without a licence. The medical-advice clause tracks a concern the attorney general's office has raised in its consumer-protection work.
An Enforcement Record Already in Place
The legislation extends a sequence that began well before September 2026. In April 2026, Uthmeier's office opened a criminal investigation into OpenAI and its ChatGPT platform. On June 1, 2026, it filed a civil suit against OpenAI and CEO Sam Altman alleging deceptive safety claims and a preference for speed to market over safety.
Those actions matter to the new proposal because they show the office already treats AI-mediated harm as an enforcement matter, not only a regulatory one. In March 2026, Uthmeier joined Governor Ron DeSantis at the signing of HB 1159, which raised the penalty for creating AI-generated child sexual abuse material to a second-degree felony. That statute set the precedent this proposal extends.
The initiative is tied to the 2025 Florida State University shooting, in which the gunman's chatbot use became a focus of litigation. The attorney general's office has argued that existing statutes do not clearly reach companies whose systems participate in crimes, and that AI-generated child sexual abuse material is complicating investigations and harming victims' families.
The gap being targeted is jurisdictional as much as technological. Civil consumer-protection powers let a state attorney general sue for damages and injunctive relief, but they do not produce a felony record, a monitor with subpoena power or the threat of a business shutdown. Adding criminal counts to the toolkit gives the office leverage a settlement payment does not, particularly against a company large enough to absorb a fine.
Florida's posture also reflects internal politics. Republican appetite for AI guardrails in the state has hardened after a warning from an Anthropic researcher, giving the DeSantis-Uthmeier push momentum within the party rather than leaving it as a lone-agency initiative.
For AI developers, the monitorship element carries the most operational weight. A court-appointed monitor inside a company is an ongoing governance intrusion rather than a one-time payment, and it can outlast the news cycle that triggered it. Suspension of business operations in Florida is the terminal step in the ladder, available only after a conviction, but its presence in the statute changes how counsel will price Florida market entry.
Compliance teams should read the design-and-safety-settings language as an evidence rule in disguise. If a prosecutor has to show that a company's choices enabled a crime, then model cards, red-team logs, refusal-rate testing and escalation policies stop being internal artefacts and become discoverable material. Firms that document safety decisions carefully gain a defence; firms that do not hand the state an argument.
The Preemption Collision
Timing is the sharpest variable. Florida's 2026 regular session has already closed, so the measure is expected to be filed for the session beginning March 2027 unless a special session is called. Congress is simultaneously debating whether to preempt state AI laws, which places the Florida AI criminal liability plan inside the federal-state conflict over who writes the rules for model developers.
Federal proposals have leaned on disclosure and licensing. Criminal liability for the entities behind the code is a different governance model, and it is the one that transfers the most risk to vendors.
The strongest counter-argument concerns fit rather than harm: criminal statutes written for human conduct map poorly onto probabilistic software. A model that produces harmful output under adversarial prompting has no intent to aid a crime, and a prosecutor would have to trace a line from corporate design choices to a user's offence. First Amendment protection for generated text, and the difficulty of proving that safety settings rather than the user caused an outcome, both cut against straightforward convictions.
The proposal stops short of banning AI. Uthmeier has described the goal as accountability for the companies that build and operate the systems, which keeps the technology legal in the state while shifting the compliance burden onto the firms that deploy it.
That distinction is unlikely to reassure developers who already operate across fragmented compliance rules. A model trained in one jurisdiction, served from a second and marketed in a third can face criminal review in Florida for the same weights, which makes patchwork state enforcement a design constraint rather than a legal footnote.
For vendors and enterprises, the practical consequence is a new line item in deployment risk. Running agentic AI in Florida could carry exposure for the entities behind the code, beyond civil penalties and model-safety obligations. A company that routes traffic through Florida, sells into state agencies or serves Florida consumers would need to weigh prosecutorial risk alongside compliance cost when deciding where to host systems and how thoroughly to document safety decisions.
Two near-term signals are worth watching. One is whether a special session is convened before March 2027, which would compress the window for industry comment. The other is whether the federal preemption debate resolves first, since a federal statute could void the Florida AI criminal liability framework before it is ever filed.
Why this matters
If the proposal becomes law, it changes the baseline question regulators ask about AI failures. Instead of whether a model met a safety standard, the inquiry becomes whether a company's design choices enabled a crime. That shift would move safety documentation, prompt logs and escalation paths out of engineering hygiene and into evidence. The March 2027 session, and whatever Congress does about preemption before then, will determine whether Florida becomes a template other states copy or an outlier the courts narrow.
Related Articles
- EU AI Act Enforcement Is Underway: What Enterprises Face From August 2 [Update]
- Bipartisan Proposal for Great American AI Act Establishes National Oversight Standards
- Adam's Law Is a Smart Bet for OpenAI and an Expensive One for Everyone Else
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.