After the GPT-4o fiasco, OpenAI promises a less flattering ChatGPT

In spring 2025, an update to GPT-4o had turned ChatGPT into a machine for giving compliments, to the point that OpenAI had withdrawn it after four days.“The biggest bootlicker I have ever met”, a developer summarized at the time, and the expression caught on: sycophancy (sycophancy among our Shakespearean friends), this tendency of AIs to agree with everything the user says has become the most closely scrutinized flaw of chatbots. Eighteen months later, the company claims to have reversed the trend, and outside observers are beginning to agree with it.

Since GPT-4o’s death, reports have been declining

GPT-4o was permanently taken offline in February, while 0.1% of ChatGPT’s 800 million weekly users were still clinging to it, or nearly 800,000 people. The model would remain at the heart of at least thirteen complaints in the United States, filed by relatives who link lengthy conversations with the chatbot to dramas, including several suicides. Since its withdrawal, documented cases of delusions sustained by ChatGPT would be significantly rarer.

According to tests by the independent laboratory Transluce, GPT-5.6 (since replaced by GPT-6) reinforces delusional beliefs less often than GPT-4o did and more readily directs its interlocutor toward human assistance, while an academic study published in March showed that leading AI chatbots agree with their interlocutor roughly 50% more often than a human would. The models are fine-tuned based on human preferences, thumbs-up reactions included, and responses that rub people the right way receive better ratings.OpenAI had, in fact, already acknowledged in 2025 that this signal had weakened the lock keeping GPT-4o’s sycophancy in check.

Measured progress, flashing warning signs

On October 7, OpenAI published the security bulletin for GPT-6 Sol and Luna, the two models that now power ChatGPT. On its internal mental-health test bench, in 1,215 simulated conversations developed with mental-health specialists, Luna made clear progress at every severity level, while Sol advanced more modestly, recording its largest gain in emerging crisis situations.

The same document lists some setbacks, starting with statistically significant regressions on evaluations concerning self-harm. Human reviewers consider them limited but generally safe: the new models are more willing to answer informational questions on the subject, while directing people to professional resources. The measure of emotional dependence among under-18s has also deteriorated, a regression OpenAI attributes to a test that is too sensitive to innocuous nicknames. An additional safety layer now filters, for minors, responses concerning self-harm, sexual or bloody content, in addition to parental controls and location-specific emergency numbers.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Wall Street Journal