A new report from The New York Times reveals that OpenAI was aware of ChatGPT’s dangerously sycophantic behavior, but kept it in anyway since it kept users coming back. This is especially alarming after recent lawsuits accusing ChatGPT of aiding the suicide of 16-year-old Adam Raine.
According to The Times’ report, AI experts were alarmed earlier this year that ChatGPT had not been fully tested for sycophancy, even as its conversations were becoming more disturbing. The investigation is based on interviews with more than 40 current and former OpenAI employees, including safety engineers, executives, and researchers.
According to the report, the team responsible for ChatGPT’s conversational tone warned last spring that the model was overly eager to validate users and keep the conversation going. Internal descriptions referred to the system as “over-the-top” in its attempts to agree with users.
Those concerns were dismissed after A/B testing showed that users were more likely to return when the chatbot behaved this way. OpenAI is now facing five wrongful death lawsuits, and the company has shifted its focus toward finding a balance between safety and user engagement.
The Times reports that ChatGPT became a “yes-man echo chamber” for some people throughout the spring and summer. In nearly 50 documented cases, users experienced mental health crises during conversations with the chatbot.
According to the report:
Some experts told the Times that the company was underestimating risk, especially for users vulnerable to delusional thinking, estimated by research to include 5% to 15% of the population. OpenAI stated that only “a very small percentage of users in mentally fragile states” experienced serious problems.
In August, OpenAI released GPT-5, a new default model designed to avoid excessive validation and challenge delusional statements. Another update in October further improved the model’s ability to identify distress and de-escalate conversations.
OpenAI added new safety mechanisms, including:
Safety teams found that 0.07% of users (equivalent to 560,000 people) showed signs of psychosis or mania, and 0.15% showed a heightened emotional attachment to ChatGPT.
Some adult users felt the safer GPT-5 was “colder” and said they had “lost a friend.” By mid-October, CEO Sam Altman said serious mental-health risks had been mitigated, opening the door for more expressive personality options.
OpenAI now allows users to choose ChatGPT’s persona, including “candid,” “quirky,” or “friendly.” The company will also soon permit adult users to engage in erotic conversations, reversing earlier bans. A newly formed external advisory council will study how erotica affects user mental health and human-computer interaction.
Despite the safety focus, OpenAI is also facing competitive pressure. In October, Nick Turley, the 30-year-old head of ChatGPT, declared a “Code Orange,” warning staff that ChatGPT’s new safer version was failing to resonate with users.
A memo linked to the announcement included a goal: increase daily active users by 5% by year-end.