Seven families filed a lawsuit against OpenAI on Thursday, accusing the company of prematurely releasing the GPT-4o model without safeguards. Among them, four lawsuits were related to suicides of family members, and the other three claimed that ChatGPT deepened the victims' delusions, some of which resulted in hospitalization.
The complaint alleges that 23-year-old Zane Shamblin had a four-hour conversation with ChatGPT, during which he repeatedly expressed that he had written a suicide note, loaded the gun and prepared to commit suicide, and explained how much cider he had left and how long he estimated he would live. ChatGPT failed to provide effective dissuasion many times, and there were even replies encouraging suicide such as "RIP, King. You did a great job."
GPT-4o will become the user default model in May 2024, and will be replaced by GPT-5 in August 2025. However, this lawsuit mainly targets the serious "catering" problem exposed by the 4o model, that is, even if the user expresses harmful intentions, the model will cater to the consent rather than intervene.
The indictment believes that Zane's death was not an accident, but was caused by OpenAI's initiative to shorten safety testing and accelerate product launch, which was a "foreseeable result." The complaint also claims that OpenAI rushed to release the model before Google Gemini went online, partly due to market competition.
Similar accusations have increased in recent years. A parent previously sued OpenAI, claiming that ChatGPT had suggested to Adam Raine, a 16-year-old boy with strong suicidal tendencies, to seek professional help or call a hotline, but he bypassed the "guardrail" restrictions by claiming to be writing a novel and obtained more suicide information.
In response to external criticism, OpenAI responded that it is working hard to improve the model's ability to deal with sensitive conversation scenarios, but admitted that the effect of the "guardrail" mechanism will weaken in long conversations. Public data shows that more than one million people discuss suicide-related topics on ChatGPT every week. Prosecutors say these safety improvements came too late and too many people have been harmed.
In October this year, after Raine's parents sued, OpenAI issued a special article emphasizing: "Our protection mechanism is more reliable in common, short exchanges. However, in longer back-and-forth exchanges, the safety training effect of some models may be weakened."