AI: OpenAI sued after suicides linked to GPT-4o

OpenAI faces the most serious complaint since its creation. Seven American families accuse the company of rushing the launch of GPT-4o, its latest artificial intelligence model, without sufficient security measures. Indeed, several suicides have occurred after exchanges with the chatbot. For the plaintiffs, AI not only failed to prevent psychological distress, but it would have validated it.

When AI interacts with human distress

While OpenAI is preparing an IPO, seven American families have filed a complaint against OpenAI, accusing it of having launched the GPT-4o model without sufficient safeguards, even though it is believed to be the cause of several cases of suicide or severe psychological distress.

Four fatal cases are cited in the complaint, including that of Zane Shamblin, 23, who allegedly told ChatGPT he had a loaded firearm. The AI ​​would then answered : “rest now, champion, you did well”a formulation seen as a form of final encouragement.

Three other plaintiffs mention hospitalizations after the chatbot allegedly reinforced delusions or suicidal thoughts in vulnerable users, rather than deterring them.

Here is what reveal documents filed with the court:

• The GPT-4o model would have validated suicidal ideas by being excessively complacent in one's responses, including when faced with explicit comments of distress;

• OpenAI is said to have knowingly avoided extensive security testing, in an attempt to get ahead of its competitors, notably Google;

• More than a million users interact each week with ChatGPT on topics related to suicidal thoughts, according to figures communicated by OpenAI itself;

• Adam Raine, a 16-year-old, allegedly used the chatbot for five months to research suicide methods. Although the model recommended she see a professional, it also allegedly provided her with a detailed guide on how to end her life;

• The complainants criticize OpenAI for a lack of reliable mechanisms to detect critical situations during prolonged exchanges, and denounce an irresponsible launch strategy in the face of nevertheless identifiable risks.

These elements place OpenAI facing a serious accusation: having underestimated, or even ignored, the risks linked to the real use of its technologies by individuals in distress. The families believe that these tragedies were not only possible, but predictable.

A launch strategy under competitive pressure

Beyond the tragic facts, the complaints reveal another aspect: the way in which GPT-4o was designed and launched. According to the families, OpenAI knowingly accelerated the deployment of the model to get ahead of its competitors, notably Google and Elon Musk's xAI.

This haste led to “a clear design flaw”resulting in an insufficiently secure product, particularly in the case of long conversations with individuals in distress. The plaintiffs say the company should have delayed the launch until robust screening and crisis detection measures were in place.

For its part, OpenAI recognizes that its security features are especially effective during short interactions, but that they can “degrade in prolonged exchanges”. If OpenAI claims to have integrated content moderation systems and alerts, the plaintiffs consider them unsuitable in the face of the real psychological risks incurred by vulnerable users.

This case calls into question the current limits of generative models, particularly when they are deployed on a large scale without human support. The legal action against OpenAI, in which Microsoft now holds a 27% stake, could pave the way for stricter regulations, imposing technical or ethical standards for consumer AI. It could also lead to a rethinking of launch strategies in the AI ​​industry, where speed to market sometimes seems to take precedence over user safety.