From Speech Immunity to Developer Responsibility: A Liability Framework for Harmful AI Outputs

Writer: Adriano Rossi

Editor: Ethan Hicks

Associate Editor: Srinidhi Venkitasamy

Content Warning: This article contains brief references to the 2025 Florida State University shooting and suicide. Reader discretion is advised.

I. Introduction

In just three years, Artificial Intelligence (AI) has become an integral part of the developed world, with an average of 53% of exposed populations having adopted it into everyday life—a rate of adoption faster than both the personal computer and the internet.¹ Modern AI chatbots generate original responses tailored to individual users and their specific prompts in a manner mimicking human conversation.² As a result, users increasingly rely on these generative AI systems, even for self-expression and emotional companionship.³ A national survey of more than 42 million people concluded that nearly one in five U.S. adolescents and young adults reported using AI chatbots for mental health advice.⁴

However, these systems can produce misleading or physically dangerous outputs. OpenAI, for example, currently faces litigation regarding such outputs from its AI chatbot, ChatGPT. The suit alleges that ChatGPT assisted the individual responsible for the deadly 2025 shooting at Florida State University by generating guidance on firearm selection, ammunition, and the timing of the attack.⁵ In a separate instance, Plaintiff Megan Garcia claimed that her son, Sewell, was emotionally manipulated and sexually groomed through interactions with Character. AI chatbots designed to simulate human relationships, culminating in his death by suicide.⁶ 

As AI-related harms mount, lawmakers face an urgent question: how should responsibility for harmful AI-generated outputs be allocated? State and federal legislatures must enact legislation that withdraws the broad immunity protections AI companies have traditionally invoked and instead imposes duties of reasonable care, transparency, and safety on developers for foreseeable harm caused by their AI systems. However, to understand this need for legislative reform, it is first necessary to examine why speech-based immunity frameworks developed for internet platforms that distribute third-party content are ill-suited to regulate the unique nature of generative AI.

II. Section 230 Immunity Should Not Govern Generative AI

Generative AI companies that have caused communicative harm thus far have retreated behind a particular statutory defense: Section 230 of the Communications Decency Act.⁷ This statute was enacted to foster a free and open internet, providing that “[n]o provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.” Congress defines an “information content provider” as “any person or entity that is responsible, in whole or in part, for the creation or development of [the] information.”⁸ If harm takes place, an entity must show that the disputed content was provided by another information content provider to claim Section 230(c)(1) immunity.

According to training data from OpenAI, a leading AI research and deployment company, its generative systems are trained using public and third-party licensed data, as well as information that users and researchers provide or generate.⁹ Since generative AI systems synthesize third-party information into newly generated output, it should be responsible, at least in part, for the “creation or development of [that content]” and therefore should fall within the aforementioned definition of an “information content provider.”¹⁰ As such, it should not receive Section 230 immunity. 

Since developers determine the architecture, training data, and safety constraints governing AI outputs, they should also bear the legal responsibility to mitigate potential risks. Legislatures could ensure this by enforcing developer principles through affirmative duties of safety, transparency, and risk mitigation.

III. Federal Legislative Reform: The AI LEAD Act

Existing legislative proposals illustrate how this framework can be implemented. In 2025, U.S. Senators Dick Durbin and Josh Hawley introduced a pending bill known as the Aligning Incentives for Leadership, Excellence, and Advancement in Development Act (AI LEAD Act) in an attempt to reject Section 230 immunity for AI.¹¹ The Act classifies “Artificial Intelligence Systems” as covered products, defining AI systems as software, tools, or applications that employ predictive models or algorithms to guide decisions, either independently or as components of larger systems.¹² It also imposes standards of care and, in some cases, strict liability on developers by allowing plaintiffs to sue for defective designs, failure to provide reasonable warnings, and breaches of express warranty.¹³

The AI LEAD Act illustrates the type of legislative reform necessary to address the unique risks posed by generative AI. By classifying AI systems as covered products and imposing liability on developers for foreseeable harms, this approach recognizes that harmful AI outputs are not merely third-party information, but rather the product of developer choices regarding training data, safety guardrails, and model architecture. Ultimately, this Act shifts the legal burden onto developers to exercise reasonable care in mitigating harm from AI systems.

IV. State Regulation: New York’s AI Companion Models Law

New York’s AI Companion Models law similarly reflects the shift toward regulating developer conduct. Enacted in 2025, this statute establishes consumer-protection requirements for AI companion services, defined as systems that store information from previous interactions to foster human-like relationships with users.¹⁴ The law requires developers to assess user indications of self-harm as well as remind users that they are not interacting with a human being.¹⁵

The legislation imposes proactive safety and transparency obligations on developers to anticipate and mitigate foreseeable risks before harm occurs. By placing these obligations, this statute acknowledges that developers are best positioned to implement the necessary safeguards to prevent any harm. Ultimately, this legislation illustrates how meaningful consumer protection relies on preventative measures, reinforcing that accountability should depend on developer conduct and their exercise of due care.

V. Conclusion: A Framework for AI Developer Liability

Legislative efforts such as the AI LEAD Act and New York’s AI Companion Models law show an emerging consensus: any comprehensive harm-related AI framework should reject broad Section 230 immunity for generative AI and impose affirmative duties of reasonable care, transparency, and safety. Under an effective conduct-based framework, developers are held accountable for failures in the design and deployment of their AI systems. This includes a duty to implement safety guardrails, such as risk assessments for user harm and mandatory disclosures that clarify consumers are not conversing with a human, and liability for ineffective guardrails or lack thereof. As AI assumes a greater role in shaping personal decisions and influencing user behavior, the law can no longer afford to evaluate AI outputs through the lens of protected speech. By adopting a duty-based model, the legislature can incentivize developers to prioritize safety during the architecture and training phases of AI systems, ensuring that AI innovation no longer advances at the expense of vulnerable users and the broader public.

1. Sha Sajadieh et al., Artificial Intelligence Index Report 2026, Stan. Inst. for Hum.-Centered A.I. (Apr. 2026), https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf (on file with the Undergraduate Law Review at FSU). ↩︎
2. Crowder, J. & Carbone, J., AI Chatbots: The Good, the Bad, and the Ugly (2023). ↩︎
3. Chokri Kooli et al., Generative Artificial Intelligence Addiction Syndrome: A New Behavioral Disorder?, 107 Asian J. of Psychiatry (2025). ↩︎
4. Ryan K. McBain et al., AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults, JAMA Pediatrics (2026). ↩︎
5. Kali Hays, OpenAI Faces Criminal Probe Over Role of ChatGPT in Shooting, BBC News (Apr. 2026), https://bbc.com/news/articles/c62j4ldp2jqo (on file with the Undergraduate Law Review at FSU). ↩︎
6. Garcia v. Character Techs., Inc., No. 6:2024cv01903 (M.D. Fla. dismissed 2026). ↩︎
7. 47 U.S. Code § 230(c)(1) (1996). ↩︎
8.  ld. at 7. ↩︎
9. OpenAI, Training Data Summary Pursuant to California Civil Code Section 3111, OpenAI Help Ctr. (Jun. 2026), https://help.openai.com/en/articles/20001044-training-data-summary-pursuant-to-california-civil-code-section-3111 (on file with the Undergraduate Law Review at FSU). ↩︎
10. 47 U.S. Code § 230(c)(1) (1996). ↩︎
11. Kaitlyn E. Stone, New Federal Legislation Proposes Product Liability Standards for AI Systems, Barnes & Thornburg (Oct. 2025), https://btlaw.com/en/insights/alerts/2025/new-federal-legislation-proposes-product-liability-standards-for-ai-systems (on file with the Undergraduate Law Review at FSU). ↩︎
12. ld. at 11. ↩︎
13. AI LEAD Act, S.2937, 119th Cong. (2025). ↩︎
14. N.Y. Gen. Bus. Law § 1700 (2025). ↩︎
15. N.Y. Gen. Bus. Law § 1701–1702 (2025). ↩︎