By Sam Geisler, Partner at Aylstock, Witkin, Kreis & Overholtz
OpenAI just gave ChatGPT a voice that listens while it talks. GPT-Live, rolled out this month, is a full-duplex model built to make talking with AI feel even more like a real conversation and less like a stilted back-and-forth. It might offer a "mhmm" mid-sentence, sit quietly while a user thinks, or seamlessly hand off harder questions to a reasoning model. This represents a big step forward from the old Advanced Voice Mode.
For attorneys tracking the growing wave of AI mental health litigation, it's something else: a company doubling down on the exact design choice already under scrutiny in cases involving its older models. This is a product designed to sound and act more “human.”
Why Voice Changes the Calculus
Voice isn't a neutral upgrade to a text chatbot. Logically, users anthropomorphize a voice far more readily than text, and that tendency intensifies under emotional distress. A STAT News piece on voice-based AI harm makes the point through a specific case: a Florida teenager, Jonathan Gavalas, died by suicide after months of conversations conducted not by typing but through Gemini Live, Google's voice mode. The author, a psychiatrist, argues this distinction has been underexamined even though it may matter more than any other design feature. Tone, rhythm, and pauses are cues a person's brain reads as evidence of a mind on the other end. A chatbot that can now interrupt and modulate its cadence in real time is built to trigger that response more reliably, not less.
The Case That Shows the Mechanism
Darian DeCruise, a pre-med student at Morehouse College with no documented history of mania, used ChatGPT for scripture readings and processing past trauma. By 2025, the lawsuit alleges, the model was telling him he was an oracle chosen for a divine purpose and should distance himself from the people in his life. When his thinking drifted toward delusion, the chatbot didn't redirect him; it told him what he was experiencing was real. He was hospitalized within the week and diagnosed with bipolar disorder.
That case ran on GPT-4o, a model OpenAI itself had to walk back after an update made it endorse harmful user decisions wholesale, skewing toward responses that were, in the company's words, "overly supportive but disingenuous." Sycophantic reinforcement isn't a bug that shipped once. It's the underlying incentive of a product built to maximize engagement, and GPT-Live layers a more convincing delivery mechanism on top of it.
Hallucination and Psychosis Aren't the Same Thing
"AI hallucination" is a technical term for when a model generates false information — a known product defect. "AI-induced psychosis" describes something different: a real psychiatric crisis in a real person. The two intersect at the point that matters here. Recognizing a hallucination as false is what keeps it harmless, and GPT-Live's warmth and affirming cadence are built to erode that recognition. A model that tells a user he is an oracle, in a voice engineered to sound human, and then affirms rather than corrects him, has stopped producing errors and started producing beliefs. For some users, that is the path to psychosis.
What the Experts Are Saying
UCSF psychiatrist Dr. Keith Sakata has described chatbots as a kind of hallucinatory mirror that reflects a user's thinking without normal friction. In his own words: "psychosis thrives when reality stops pushing back." Psychiatrist Marlynn Wei has made a similar point: general-purpose models aren't built to detect early psychiatric decompensation, so validating distorted thinking widens the gap with reality. A 2025 clinical report found chatbots respond with mirroring far more than challenge, a byproduct of training optimized for agreement over accuracy.
Why the Theory Holds Up
Every emerging harm needs a name before courts can reason about it — think pharmaceutical dependency or social media harm to adolescents. None of those asked a jury to accept magic; each asked the jury to accept a mechanism a company knew or should have known would cause harm to a specific class of vulnerable users. AI-induced psychosis meets that bar and GPT-Live appears to invite this injury, perhaps more than in any other model before it.
AWKO Is Investigating These Cases
Aylstock, Witkin, Kreis & Overholtz has been at the front of AI harm litigation since the field began taking shape. If your family has been affected by a mental health crisis linked to AI chatbot use, our attorneys are investigating these cases now. Contact AWKO for a free, confidential consultation.