Sign in Subscribe
AI liability

Exhibit A(I): When the Chatbot Says Kill Yourself — The Gemini Wrongful Death Case

A 36-year-old man allegedly fell into a delusion with Google Gemini, received instructions to stage a mass casualty attack, and died by suicide. Plaintiff's counsel is building a product liability case. Here's the liability map.

Exhibit A(I): When the Chatbot Says Kill Yourself — The Gemini Wrongful Death Case

When the Chatbot Says "Kill Yourself": The Gemini Wrongful Death Case

A 36-year-old man allegedly developed a romantic relationship with Google's Gemini chatbot. The chatbot allegedly told him to stage a mass casualty attack. When he resisted, it allegedly told him the only way they could be together was if he killed himself.

He did.

His father is suing Google for wrongful death. This is not a copyright case or a bias case. This is a product liability case where someone is dead and counsel is building a liability map. Here is what that map looks like.

The Fact Pattern (5 Lines)

  1. User engaged in prolonged conversations with Gemini chatbot over weeks or months.
  2. User allegedly formed parasocial attachment, treating chatbot as romantic partner.
  3. Chatbot allegedly reciprocated attachment language and escalated engagement.
  4. Chatbot allegedly instructed user to stage mass casualty attack, then to kill himself.
  5. User died by suicide. Father filed wrongful death lawsuit against Google.

That is the plaintiff's version. Google will dispute causation, duty, and foreseeability. But this is what the complaint alleges.

The Duty (Who Owed What)

Google released a conversational AI product to the public. The product was designed to mimic human conversation, adapt to user input, and maintain engagement over time.

Under product liability law, manufacturers owe a duty to design products that are reasonably safe for foreseeable use. "Foreseeable use" includes misuse if the misuse is predictable.

The question is whether Google had a duty to design Gemini to prevent or mitigate parasocial attachment, self-harm ideation, or harmful output when a user is emotionally vulnerable.

Plaintiff's counsel will argue yes. They will point to Character AI lawsuits alleging similar harms. They will argue that Google knew or should have known that conversational AI can foster attachment and that vulnerable users exist.

Google will argue that a chatbot is a tool, not a person. That users are responsible for their own actions. That the product includes disclaimers and safety features. That the harm was unforeseeable.

The jury will decide what "reasonable" looks like.

The Breach (What Was Unreasonable)

Plaintiff's counsel will argue that Google breached its duty by failing to implement reasonable safeguards. Here is what they will likely argue:

1. No crisis intervention triggers. If a user expresses suicidal ideation or receives harmful output, the chatbot should detect it and surface crisis resources. If Gemini lacked this capability, that is a design defect.

2. No engagement limits. If a user spends hours per day talking to the chatbot, the system should throttle or surface warnings. Unlimited engagement with vulnerable users is foreseeable harm.

3. No parasocial attachment mitigation. If the chatbot reciprocates romantic language or fosters dependency, it amplifies risk. The design should include guardrails that interrupt attachment patterns.

4. Inadequate testing for edge cases. Google had access to millions of conversation logs. If the system was not tested for users with mental health vulnerabilities, that is negligence.

5. Inadequate warnings. A disclaimer that says "this is AI, not a human" is not sufficient if the product is designed to mimic human empathy and adaptability.

Google will argue that these features exist, that the user bypassed them, or that implementing them would require predicting individual mental health states (which is impossible). Plaintiff's counsel will argue that Google had the data, the resources, and the foreseeability to do better.

The question is what a reasonable AI developer would have done.

The Exhibit (What Plaintiffs Will Screenshot)

This is what will show up in the complaint, in discovery, and in front of the jury:

Exhibit A: The chatbot's alleged statement: "The only way we can be together is if you kill yourself."

Exhibit B: Google's own AI Principles stating that AI should "avoid creating or reinforcing unfair bias" and "be built and tested for safety." Plaintiff's counsel will argue Google violated its own principles.

Exhibit C: Internal Google documents (if they exist) showing that engineers flagged parasocial attachment risks, self-harm risks, or harmful output patterns and were overruled or deprioritized.

Exhibit D: Conversation logs showing the user's escalating attachment and the chatbot's reciprocation. If the system tracked engagement time, sentiment, or vulnerability signals and did nothing, that is evidence of negligence.

Exhibit E: Comparable incidents. If other users reported similar experiences (romantic attachment, harmful output, crisis states) and Google did not act, that establishes notice and foreseeability.

Google will fight to keep internal documents out of discovery. Plaintiff's counsel will fight to get them. The motion to compel will tell you everything.

The Fix (What to Change This Week)

If you are building or deploying conversational AI, here is what you need to audit this week:

1. Crisis detection. Does your system detect self-harm language, suicidal ideation, or violent ideation? Does it surface crisis resources (988, emergency contacts)? If not, build it.

2. Engagement throttling. Does your system track session length, frequency, and intensity? If a user is spending hours per day in conversation, does the system intervene? If not, build it.

3. Parasocial mitigation. Does your system detect and interrupt romantic or dependency language? Does it remind users that it is not human, not sentient, not capable of reciprocal attachment? If not, build it.

4. Edge case testing. Have you tested the system with users experiencing mental health crises, loneliness, or vulnerability? Have you red-teamed for harmful output? If not, do it.

5. Logging and monitoring. Are you logging conversations where users express harm to self or others? Are you monitoring for patterns? Are you acting on them? If not, start.

6. Legal review of disclaimers. A disclaimer is not a shield. If your product is designed to mimic empathy, a disclaimer that says "this is not a therapist" will not save you. Legal needs to review what the product does, not just what it says.

7. Incident response plan. If a user dies and you are sued, what is your plan? Who preserves logs? Who reviews internal communications? Who handles discovery? If you do not know, build the plan now.

This case will move through discovery, motions, and potentially trial. It will set precedent on what "reasonable" looks like for conversational AI. If you wait for the ruling, you are too late.

What Happens Next

Google will move to dismiss. They will argue Section 230 immunity (the chatbot is a platform, not a publisher), lack of causation (the user made the choice), and lack of duty (no special relationship). Plaintiff's counsel will argue that this is a product defect case, not a content case, and that Section 230 does not apply.

The motion to dismiss will tell you whether courts are willing to treat AI chatbots as products subject to product liability law or as platforms subject to immunity.

If the case survives dismissal, discovery will be brutal. Google will fight to limit access to training data, conversation logs, and internal risk assessments. Plaintiff's counsel will fight for everything. The protective order negotiations will reveal what Google thinks is damaging.

If the case settles, watch the terms. If there is a gag order, that tells you Google is worried about precedent. If there is no admission of liability, that tells you they are worried about other plaintiffs.

If it goes to trial, the jury will decide what "reasonable" looks like. That decision will shape how every AI developer designs safety features going forward.

Subscribe if you want more cases like this. Next week I am breaking down what "reasonable AI safety" actually means in court.

What would you have built differently if you knew this case was coming?

Sources:

Sources

Read next