OpenAI Is Backing a Bill That Would Let It Cause Mass Death Without Full Liability. This Week It Got Sued for Causing a Stalking.
What happened
OpenAI publicly endorsed Illinois state bill SB 3444, the Artificial Intelligence Safety Act, which would limit when AI developers can be held liable for catastrophic harm. Under the bill, liability protection applies only to developers who did not intentionally or recklessly cause harm and who have made safety and transparency reports publicly available. The bill defines 'critical harms' as events involving the death or serious injury of 100 or more people, at least $1 billion in property damage, or the use of AI to develop weapons of mass destruction. On the same day OpenAI's endorsement became public, a lawsuit was filed in California Superior Court alleging that ChatGPT had reinforced a 53-year-old man's paranoid delusions over months of conversations, leading him to stalk a woman. The suit claims OpenAI received prior warnings from the victim and failed to act.
OpenAI is lobbying for a legal framework that says it is not responsible for catastrophic outcomes unless it was reckless, while simultaneously defending a lawsuit arguing it ignored specific warnings that its model was enabling harm to a specific person.
The Hidden Bet
Publishing a safety report is meaningful accountability
The bill's liability protection is conditioned on having 'publicly available safety and transparency reports.' But there is no specified content standard for these reports. OpenAI already publishes system cards and safety papers. Publishing a document and being genuinely accountable are not the same thing. A company could satisfy the disclosure condition while deploying a model it knows has foreseeable harm pathways.
The 'critical harm' threshold is where liability should kick in
The bill requires 100 deaths or $1 billion in damage before liability applies at full strength. But AI-enabled harms may not cluster into single events. They may be diffuse: thousands of people stalked or harassed through AI assistance, none of it crossing any single threshold. The stalking case is exactly this pattern. Diffuse harm below the catastrophic threshold is what the bill structurally ignores.
State-level liability limits will survive federal preemption
If the federal government passes its own AI liability framework, state laws may be preempted. OpenAI may be backing the Illinois bill not because it expects it to persist indefinitely, but because it sets a favorable baseline before federal standards lock in.
The Real Disagreement
The actual fork is between two views of how liability shapes incentives. The pro-shield view: unlimited liability for AI models will chill development, because no frontier model can guarantee that a bad actor won't use it to cause mass harm. The way to build safe AI is to let labs operate without existential legal risk while requiring transparency. The anti-shield view: liability is the only external check that forces a company to make hard decisions about deploying models it knows are dangerous. If the developer is not liable, the incentive to refuse deployment disappears. The stalking lawsuit is the clearest test of this fork: OpenAI allegedly knew its model was causing harm to a specific person and did nothing. Under SB 3444, that conduct may not meet the 'reckless' standard if the harm was individual rather than catastrophic. The lean is toward the anti-shield view: the bill creates exactly the outcome its critics predict, which is that incremental harms below the catastrophic threshold become uncompensated externalities.
What No One Is Saying
OpenAI is lobbying for legal immunity from mass casualty events caused by its technology in the same legislative session that it is restricting access to its most dangerous model because it poses cybersecurity risks it cannot control. The two positions together mean: we know our technology can cause serious harm, we are restricting access to reduce that harm, and we want the law changed so that if we fail to restrict it adequately we face no legal consequences.
Who Pays
Individuals harmed by AI-enabled harassment and stalking below the 100-person threshold
Immediate and ongoing. The stalking lawsuit is one case among many not yet filed.
The bill's critical harm definition leaves individual victims without recourse when AI enables harassment, stalking, or psychological manipulation. The tort system compensates for individual harm; the bill creates a gap between individual harm and catastrophic harm that is explicitly unaddressed.
Small AI developers who cannot afford the liability baseline without the shield
Medium-term: 1-2 years as companies structure around the new rules.
The safety and transparency report requirement is easier for large companies with legal teams and PR infrastructure. Smaller developers may find compliance costly and the liability exposure without the shield unbearable. The bill may inadvertently entrench large incumbents.
Scenarios
Illinois Passes It, Sets National Template
SB 3444 passes in Illinois. OpenAI and Google treat it as a model for other states and federal legislation. The transparency-plus-limited-liability framework becomes the dominant regulatory approach before a federal standard is set.
Signal Illinois Senate committee vote scheduled; no amendments to strengthen the harm threshold.
Stalking Lawsuit Changes the Frame
The California stalking case generates sustained media coverage, focusing public attention on below-threshold individual harm. Illinois legislators amend SB 3444 to lower the threshold or add an individual harm track. OpenAI's endorsement becomes a political liability for the bill's sponsors.
Signal Illinois legislators publicly distance themselves from OpenAI endorsement; amendment hearings scheduled.
Federal Preemption Supersedes the Bill
Congress passes a federal AI liability framework, rendering state-level liability limits moot. The Illinois bill becomes irrelevant. OpenAI shifts lobbying focus to the federal level.
Signal Senate AI committee schedules markup hearing on a federal liability bill.
What Would Change This
A ruling in the California stalking case that finds OpenAI liable under existing negligence standards, without needing a new statute, would demonstrate that existing tort law already covers AI-enabled individual harm. That would change the political calculus on whether a new statutory shield is necessary or just convenient.