OpenAI has testified in support of an Illinois bill, SB 3444, that would limit when AI labs can be held liable for harms tied to their models—even when those harms are described as “critical harms,” including mass death or large-scale financial or property damage. The move, reported by WIRED, marks a shift in OpenAI’s legislative strategy from largely opposing AI liability legislation toward advocating for a specific legal framework for advanced systems.
What SB 3444 would change in AI liability
According to WIRED, the Illinois bill would shield AI labs from liability in cases where AI models are used to cause serious societal harms: death or serious injury of 100 or more people, or at least $1 billion in property damage. The proposal centers on the idea that the most advanced systems should have a clearer liability boundary.
The bill would protect “frontier” AI developers from liability for “critical harms” caused by their frontier models, provided developers did not intentionally or recklessly cause such an incident. It also requires that developers have published safety, security, and transparency reports on their websites.
SB 3444 defines a “frontier model” as any AI model trained using more than $100 million in computational costs. WIRED notes that this definition “likely could apply” to major U.S. labs, naming OpenAI, Google, xAI, Anthropic, and Meta as examples of organizations that could fall under the threshold.
OpenAI’s legislative strategy shift
WIRED describes the effort as a potential change in OpenAI’s legislative strategy. Until now, the company “has largely played defense,” opposing bills that could have made AI labs liable for harms tied to their technology. In that context, SB 3444 stands out because it proposes a liability limitation rather than simply resisting liability.
WIRED reports that “several AI policy experts” told the outlet that SB 3444 is “a more extreme measure than bills OpenAI has supported in the past.” This characterization suggests the bill could be interpreted as moving further toward broad liability exemptions than prior approaches.
OpenAI’s position, as quoted by WIRED, is that the model is designed to target risk reduction while maintaining adoption. In an emailed statement, OpenAI spokesperson Jamie Radice said: “We support approaches like this because they focus on what matters most: Reducing the risk of serious harm from the most advanced AI systems while still allowing this technology to get into the hands of the people and businesses—small and big—of Illinois.” Radice also said the approach “help[s] avoid a patchwork of state-by-state rules and move toward clearer, more consistent national standards.”
How “critical harms” are defined
SB 3444’s definition of “critical harms” includes several categories WIRED describes as common areas of concern for the AI industry. One example is a bad actor using AI to create a chemical, biological, radiological, or nuclear weapon. The bill also covers a scenario where an AI model engages in conduct on its own that—if committed by a human—would constitute a criminal offense and leads to the extreme outcomes listed under the bill.
The core technical-policy linkage is clear: liability is tied to the model being “frontier” by compute threshold, to whether the incident meets the severity thresholds, and to whether the developers satisfied conditions like not intentionally or recklessly causing the incident and publishing safety, security, and transparency reports.
Technical implications for model developers
From a technology standpoint, SB 3444 would push advanced model developers toward a more standardized public posture around safety and security documentation. WIRED reports that the bill requires published safety, security, and transparency reports on a developer’s website as part of the liability shield. The structure suggests that compliance would depend on having a repeatable reporting process aligned with the legal definition of a “frontier model.”
The compute threshold—more than $100 million in training computational costs—has a direct technical implication: it ties legal risk categorization to measurable training expense. This could influence how labs document training runs and model development budgets, since the definition is based on computational cost rather than other metrics such as model size or capability benchmarks.
Additionally, the bill’s carveout depends on intent and recklessness. This introduces a legal standard that may be difficult to map directly onto technical development practices, particularly for complex systems where causality can be hard to establish. WIRED’s framing indicates that liability would be limited when developers did not intentionally or recklessly cause the incident, which suggests courts and regulators would need to interpret development and deployment actions in relation to “critical harms.” Observers may watch for how developers document decision-making and risk controls, since the exemption is conditioned on both behavior and reporting.
OpenAI’s quoted rationale emphasizes avoiding “a patchwork of state-by-state rules” and moving toward “clearer, more consistent national standards.” If other states adopt similar frameworks, the legal environment for frontier model deployment could become more uniform. If not, the industry could still face uneven compliance burdens.
Implications for advanced AI systems
SB 3444, as described by WIRED, is not a general AI regulation; it is a liability framework aimed at “frontier” systems defined by compute cost and tied to specific thresholds for harm severity. For technology teams building advanced models, the bill’s design connects legal exposure to engineering governance: which models qualify, what developers publish, and how intent-based standards are interpreted when extreme outcomes occur.
The reported shift in OpenAI’s legislative strategy—from opposing liability bills to supporting a specific exemption—could signal how major labs want to shape the rules that govern deployment of advanced AI. As WIRED notes, policy experts characterized SB 3444 as potentially more extreme than prior bills supported by OpenAI, which suggests the industry may be entering a phase where liability policy becomes closely coupled with how frontier models are trained and documented.
Source: WIRED