The Two Bills
Illinois has two competing AI bills in play. They aren’t variations on the same idea, they reflect fundamentally different assumptions about where AI accountability should live.
Senate Bill 3444, the Artificial Intelligence Safety Act, would provide liability protection for AI developers under defined conditions. According to reporting on the bill’s described terms, a developer would be protected if they did not “intentionally or recklessly” cause harm and had published safety and security protocols before harm occurred. The bill reportedly sets a high threshold for “critical harms”, defined in coverage of the legislation as events causing death or serious injury to 100 or more people, or $1 billion or more in property damage. Below that threshold, the liability shield, as described in reporting on the bill, would apply to developers who meet the publication requirement.
Senate Bill 3261 takes the opposite approach. Rather than shielding developers from downstream consequences, it would require them to maintain auditable safety plans and child protection measures as affirmative obligations. Anthropic is backing this bill. The distinction isn’t subtle: SB 3444 asks “did you act in good faith?” after harm occurs; SB 3261 asks “can you prove you built safeguards before harm occurred?”
Both bills are pending. Neither has passed as of April 15, 2026. The governor’s position has not been publicly confirmed and is not included here.
OpenAI’s Position and the Logic Behind It
OpenAI is publicly supporting SB 3444. Its core argument, per coverage of its lobbying position, centers on the risk of regulatory fragmentation. OpenAI argues that inconsistent state-level AI laws create a compliance “patchwork” that burdens US developers without producing commensurate safety benefits, and that this patchwork disadvantages American AI development relative to international competitors operating under unified national frameworks.
This isn’t a new argument. OpenAI has advanced versions of the federal preemption case in other state-level AI debates. What’s notable in Illinois is that the company is now explicitly backing a specific bill that includes a liability shield, not merely opposing a bad bill, but advocating for a legislative model. That’s a more exposed position. It can be quoted, compared against future behavior, and used as precedent in the next state that writes an AI liability law.
The logic of the shield is coherent within OpenAI’s framework. If a developer has published safety protocols, made them public and verifiable, and didn’t act recklessly, the argument goes, liability should attach to misuse rather than development. The publication requirement does create a paper trail. What it doesn’t create is an independent verification mechanism. Whether published protocols are adequate is, under SB 3444’s described framework, not a pre-condition for the shield – only that protocols exist and were published.
Anthropic’s Position and the Logic Behind It
Anthropic’s opposition to SB 3444 is direct. Gizmodo reported that Anthropic’s representatives characterized the bill as a “get-out-of-jail-free card.” That framing is sharper than standard legislative testimony, it signals that Anthropic views the shield not as a reasonable safe harbor but as a mechanism that removes accountability without replacing it.
Anthropic’s alternative, SB 3261, is built on a different premise: that safety should be a provable, ongoing obligation rather than a retrospective defense. Auditable safety plans mean a third party can evaluate whether the plans are adequate, not just whether they exist. Child protection measures as a named requirement signals that Anthropic’s model of AI governance includes specific harm categories that warrant affirmative protection, not just a general “we acted in good faith” standard.
Anthropic has also referenced New York and California AI frameworks as models – states that have moved toward process-based accountability requirements rather than developer shields. That reference is strategic positioning: Anthropic is aligning with an emerging coalition of states that favor the audit-and-accountability model over the shield model. Whether that coalition holds, and whether it influences the federal conversation, is the longer-term question.
The Strategic Subtext
Both companies have regulatory exposure. But the shape of that exposure is different, and their positions on SB 3444 reflect it.
OpenAI deploys at a scale and in a product surface area, consumer applications, enterprise APIs, developer tools, where state-by-state liability variation is a genuine operational burden. Fifty different state liability standards, each with different thresholds and defenses, create legal complexity that larger, more legally resourced companies can absorb more easily than smaller ones. The patchwork argument has real merit as a description of the compliance landscape, even if it also serves OpenAI’s interests. Notably, the same week OpenAI backed the Illinois liability shield, the company launched GPT-5.4-Cyber, a cybersecurity-specialized model with permissive refusal boundaries for vetted security vendors, according to OpenAI. The timing illustrates exactly what kinds of model deployments the liability framework is being designed around.
Anthropic’s exposure is different. The company has positioned itself publicly on safety as a competitive differentiator, its Constitutional AI approach, its research into model behavior, its stated priority of alignment over speed. Backing a liability shield would undercut that positioning. More practically: if Anthropic’s safety processes are genuinely more rigorous than competitors’, an audit requirement hurts competitors more than it hurts Anthropic. SB 3261 isn’t just principled positioning – it’s structurally advantageous for a company that wants its safety investments to be competitively recognized.
Neither position is purely principled, and neither is purely self-interested. They reflect two genuinely different models of how AI risk should be governed, advanced by two companies whose business models and risk profiles make each model relatively more attractive to them.
What Comes Next
The immediate question is the vote on SB 3444. A pending vote means the Illinois General Assembly will need to weigh a bill that has attracted public opposition from one of the two most prominent AI companies in the country. That’s unusual, and it gives legislators cover to amend or reject the bill without appearing to oppose the AI industry wholesale.
The longer-term pattern is the one compliance teams should track. California, New York, and Texas all have active AI regulation pipelines. When those states write their own liability frameworks, and some version of that is coming, the Illinois debate will be cited. The OpenAI/Anthropic split creates a usable legislative record: here are two frontier labs, here are their positions, here is the argument each makes. That record makes it harder for either company to quietly shift positions in the next jurisdiction, and it gives state legislators a documented basis for choosing one model over the other.
A “State AI Liability Law Tracker” covering all active US state AI legislation would serve the compliance audience directly here, this is a gap in existing TJS hub content that the Illinois story has made visible.
TJS Synthesis
The Illinois story is not really about Illinois. It’s about the fact that two frontier labs have now staked out opposing positions on AI liability governance in a public, quotable, traceable way, and that both positions are coherent within their respective strategic and philosophical frameworks. What’s now on the record: OpenAI believes liability should attach to bad actors, not to developers who act in good faith and publish their safety work. Anthropic believes good faith isn’t auditable and that safety claims need to be verifiable before harm occurs, not defended after it. Every state that writes an AI liability law going forward will need to choose between some version of these two models. The Illinois vote doesn’t settle that choice. It just made the choice visible.