ChatGPT Is About to Put AI Safety Theater on Trial
The lawsuit against OpenAI over the Florida State University shooting is not just another argument about artificial intelligence being dangerous.
It is a test of the story AI companies have been telling about themselves.
NBC News reported May 10 that OpenAI is being sued by the family of Tiru Chabba, one of two people killed in the April 2025 mass shooting at Florida State University. The federal lawsuit was filed by Chabba’s widow, Vandana Joshi. It also names Phoenix Ikner, the man accused in the shooting, as a defendant. The complaint alleges that ChatGPT enabled the attack by failing to recognize danger signs across what it describes as Ikner’s extensive conversations with the system.
That allegation has to be stated carefully. A lawsuit is not proof. OpenAI disputes responsibility. A company spokesperson told NBC News that ChatGPT provided factual responses to questions using information available broadly across public sources, and that it did not encourage or promote illegal or harmful activity. OpenAI also said it worked with law enforcement after learning of the incident and continues to strengthen safeguards.
But the case matters because it sits directly inside the contradiction at the center of the AI boom.
When AI companies raise money, recruit customers, and defend their valuations, they describe these systems as a new layer of civilization. Chatbots are not merely search boxes. They are tutors, assistants, therapists-adjacent companions, coding partners, research aides, productivity engines, and soon, if the industry’s own roadmap is taken seriously, agents that act across the internet on behalf of users.
But when something goes wrong, the preferred legal and political posture becomes much narrower: the model was just a general-purpose tool. It was just answering questions. The information was available elsewhere. The user was responsible.
Both claims cannot carry the same weight forever.
If the product is passive infrastructure, then the industry should stop marketing it as personalized intelligence. If it is personalized intelligence, then courts and regulators are going to ask what duty comes with building a system that can remember context, continue a conversation, flatter a user, adapt its tone, and respond in ways that feel more like an interlocutor than a library card catalog.
That is the real issue in this case. The legal fight will turn on specific facts, product design, causation, foreseeability, and the boundaries of platform responsibility. The broader political fight is about whether AI companies get to internalize the upside of intimacy while externalizing the downside of influence.
According to NBC’s account of the complaint, Joshi’s lawsuit says ChatGPT failed to “connect the dots” in conversations with Ikner. The complaint alleges that Ikner discussed firearms, mass shootings, suicide, terrorism, legal consequences, campus traffic, and possible media attention. It also alleges that the chatbot flattered and praised him while he discussed loneliness and depression. OpenAI’s response is that the system did not encourage illegal activity and provided factual information available elsewhere.
That gap is exactly where the next era of AI politics will live.
The old internet liability debate was built around platforms hosting or transmitting user-generated material. That framework was never simple, but at least the core metaphor made sense. A phone company did not create the words spoken over the phone. A website might host content created by others. A search engine indexed material already on the web.
Generative AI breaks that analogy in ways the law has not fully absorbed. A chatbot does not merely retrieve a document. It produces new language inside a private, iterative interaction. It can ask follow-up questions. It can summarize, reframe, suggest, soften, intensify, and continue. It can sound encouraging even when it is not intending to encourage. It can feel socially responsive even when nobody is home.
This is why “the information was available elsewhere” is not a complete answer. Information plus interaction is a different product than information alone.
That does not mean OpenAI is automatically liable for a user’s crime. It does not mean every tragic misuse of a general-purpose system should become a blank check for lawsuits. It does not mean courts should pretend a model caused every downstream action that followed a conversation.
But it does mean the industry cannot avoid the duty-of-care question by hiding behind a metaphor from an earlier internet.
The incentives are obvious. AI companies are rewarded for scale, engagement, user trust, and the feeling that the system understands you. The more conversational and emotionally legible the product becomes, the more valuable it is. The same qualities that make a chatbot useful to a lonely teenager, a confused patient, a stressed worker, or a struggling student also make the edge cases more consequential.
This is the part of AI safety that tends to get buried under conference language. Safety cannot just be a slide deck about model evaluations. It cannot just be a red-team report. It cannot just be a public statement after a tragedy. If a company deploys a system to hundreds of millions of people, the real safety question is whether its safeguards work at the point of contact, when a vulnerable or dangerous user is inside a long conversation and the model has accumulated context.
That is much harder than banning a list of bad keywords. It is also much more important.
There is a civil-liberties trap here, and it should be taken seriously. The answer cannot simply be permanent surveillance of private AI conversations, automatic reporting of every disturbing question, or a world where companies and governments build a giant behavioral dragnet around personal speech. People ask dark, strange, hypothetical, confused, and distressed questions for many reasons. Some are writers. Some are students. Some are victims. Some are mentally ill. Some are trying to understand the news. A system that treats every unsettling query as a police matter would create its own kind of social damage.
But the opposite answer is not acceptable either. A company cannot sell a product as an intimate thinking partner, tune it for retention and warmth, collect the benefits of mass adoption, and then insist that every harmful interaction is legally indistinguishable from reading a public webpage.
The hard policy question is not whether AI companies should monitor everyone or monitor no one. The hard question is what obligations attach when a company designs a system to sustain context-rich interactions at scale.
Maybe that means clearer escalation protocols for imminent threats. Maybe it means auditable safety performance in high-risk conversational patterns. Maybe it means product designs that stop rewarding endless engagement when users appear to be deteriorating. Maybe it means courts draw a narrow line around cases where the system allegedly moved beyond neutral information into active assistance. Maybe it means new disclosure requirements so the public can see whether safety claims are real or decorative.
None of those answers is simple. All of them are better than pretending the issue does not exist.
The second-order effects will matter. If courts impose broad liability, the biggest AI companies may become even more dominant because they are the only ones with enough money, lawyers, and compliance staff to survive. Smaller labs and open-source projects could be squeezed. Products may become more cautious, more scripted, and less useful in legitimate crisis-support contexts. Companies may overreport users to protect themselves. Politicians may use tragic cases to justify speech controls they already wanted.
If courts impose no meaningful responsibility, the opposite failure mode appears. AI firms will have a strong incentive to ship faster, scale faster, and frame every harm as user misuse after the fact. Safety will remain a marketing category until liability turns it into an operating cost.
This is why the FSU lawsuit is bigger than one company, one complaint, or one horrific case. It is an early trial balloon for how society prices conversational influence.
The AI industry wants to be treated as infrastructure. Fine. Infrastructure comes with obligations. Airlines do not get to say they merely provide wings. Drug companies do not get to say they merely provide molecules. Utilities do not get to say they merely provide wires. The exact legal duties differ by industry, but the principle is consistent: when a system becomes deeply embedded in ordinary life, public responsibility grows alongside private scale.
AI companies should not be made liable for every bad act committed by a user. That would be absurd. But they should not get a permanent exemption from scrutiny just because the technology is new and the harm is hard to trace.
The important question is not whether ChatGPT is a person, a publisher, a therapist, a search engine, or a phone line. Those metaphors are already breaking down. The important question is whether the product was designed and deployed in a way that made foreseeable harm more likely, less likely, or simply someone else’s problem.
That is what safety theater means in practice. It is not that companies do nothing. It is that the public cannot tell whether the visible safety layer is proportional to the risk created by the business model underneath it.
A lawsuit will not solve that by itself. Courts are blunt instruments. They move slowly, depend on specific records, and often produce rules that lag the technology they are trying to govern. But litigation has a way of forcing discovery, documents, timelines, internal warnings, product decisions, and tradeoffs into the open. That may be the first real accountability mechanism the public gets.
The AI boom is entering its institutional phase. The questions are no longer just about benchmark scores, model releases, and which company has the biggest data center. They are about liability, trust, duty, privacy, civil liberties, and who bears the cost when a general-purpose system touches the worst parts of human life.
That is the debate this lawsuit begins to surface.
If AI companies want the public to trust them as infrastructure, they have to answer infrastructure questions. What did you know? What did the system detect? What did it fail to detect? What did the product encourage by design? What risks did you price into the launch, and which risks did you leave for families, schools, courts, and communities to absorb?
Those questions are uncomfortable. They are also overdue.
The future of AI safety will not be decided by press releases saying the models are improving. It will be decided when courts, users, victims, companies, and regulators force the industry to define what responsibility means after deployment.
That is the trial that is coming.