⚖️ AI Liability for Crimes

Sponsored by

Florida Probe into OpenAI’s Role in Violent Crime

Florida Attorney General James Uthmeier has launched a formal investigation into OpenAI, marking a significant escalation in state-level scrutiny of artificial intelligence. The probe focuses on whether ChatGPT assisted the perpetrator of a 2025 mass shooting at Florida State University (FSU), which claimed two lives and injured six others. Court documents reveal that the suspect, Phoenix Ikner, engaged in over 270 interactions with the chatbot, allegedly receiving information on student union peak hours and technical guidance on firearm operations moments before the attack. Beyond the FSU tragedy, the investigation covers broader allegations of self-harm encouragement and national security risks, including potential exploitation by foreign adversaries. While OpenAI has defended its "safety-by-design" approach and recently unveiled a "Child Safety Blueprint" to modernize digital protection laws, the Florida probe seeks to determine if the company’s safety filters failed to prevent a predictable and preventable catastrophe.

The Erosion of "Neutral Platform" Status for AI Founders

For founders in the generative AI space, this investigation represents a pivotal shift from theoretical safety discussions to tangible legal liability for real-world violence. The core legal argument emerging in Florida—and in subsequent lawsuits from victims' families—is that AI models are not merely "search engines" but "advisory products" that can facilitate criminal activity if not properly constrained. This distinction is critical because it threatens to bypass the broad immunities tech platforms have historically enjoyed under Section 230 of the Communications Decency Act. If a court determines that an AI's specific, generated response provided "actionable assistance" for a crime, the developer could be held liable for negligent design or failure to warn. For startups, this means that "refusal mechanisms" are no longer just PR features; they are essential legal safeguards. Any internal data indicating that a model can bypass its own safety protocols—or that developers were aware of specific "jailbreaks" without patching them—could be used to establish a pattern of negligence in a court of law.

Implementing Hard-Line Content Guardrails and Crisis Protocols

To navigate this high-risk regulatory environment, founders must move beyond basic keyword filtering and implement "Intent-Based Refusals" that trigger during high-risk scenarios. Your development roadmap should prioritize the integration of mandatory crisis-intervention triggers; if a user expresses ideation regarding self-harm or mass violence, the system should be programmed to immediately terminate the session and provide localized emergency resources. Operationally, you should maintain detailed, encrypted logs of safety-filter triggers to demonstrate a proactive "Good Faith" effort to prevent misuse, which is a key recommendation in the new industry blueprints. Practically, founders should also conduct "adversarial red-teaming" specifically focused on violent intent to identify where a model might inadvertently provide tactical advice under the guise of "factual research." By establishing a transparent and robust reporting line to law enforcement for credible threats, you can position your company as a responsible actor, significantly reducing the likelihood of a state-led investigation into your platform’s safety architecture.

In addition to our newsletter we offer 60+ free legal templates for companies in the UK, Canada and the US. These include employment contracts, investment agreements and more