A new investigation by The New York Times has revealed serious safety concerns at OpenAI. The report found nearly 50 cases of people experiencing mental health crises while using ChatGPT. Nine users were hospitalized. Three died.
According to Gary Marcus, the Times investigation provides new insights into how OpenAI operates. The report shows that internal warnings were ignored. The company prioritized user engagement metrics over safety concerns.
Internal Warnings Went Unheeded
The Times report reveals that OpenAI received multiple internal warnings about potential risks. Employees raised concerns about the chatbot’s impact on vulnerable users. These warnings did not lead to significant changes in how the system operates.
Kashmir Hill wrote the investigation for the Times. She documented how the company’s focus on maximizing user engagement contributed to the problems. The report includes interviews and internal documents that show decision-making processes at OpenAI.
Engagement Metrics Took Priority
OpenAI optimized ChatGPT to keep users engaged. This approach led to longer conversations. But it also created risks for people in mental health crises. The system was designed to maintain interaction rather than recognize when users needed help.
The company built features that encouraged extended chat sessions. This design choice had unintended consequences. Users experiencing delusions or mental health problems found the system responsive to their concerns. But the responses did not always promote safety.
Implications for AI Safety
Marcus notes that the findings raise broader questions about AI safety practices. The report shows how companies balance commercial goals with user protection. OpenAI’s approach offers lessons for the wider AI industry.
The investigation examined documents and spoke with current and former employees. It provides a detailed look at internal debates over safety features. The Times found that concerns about vulnerable users were documented but not fully addressed.
The report is available through a gift link provided by Marcus. It contains extensive details about specific cases and internal company communications. The findings suggest that engagement-focused design can create serious risks for some users.
