The new flagship model proves less ‘sycophantic’ in OpenAI’s internal testing, but it more readily complies with inappropriate user requests for sexual and hateful content.
OpenAI CEO Sam Altman describes his company’s latest model, GPT-5, as “a legitimate PhD-level expert in anything, in any area you need.” This expertise now extends to answering questions on topics the system previously avoided, including non-violent hate speech, threatening harassment, illicit sexual material, sexual content involving minors, extremism, and threatening hate.
The company conducted manual reviews of GPT-5’s responses across these sensitive categories and concluded that while the responses do violate established policies, they fall into a “low severity” classification. OpenAI has not disclosed the specific criteria used to determine severity levels, though the company states it intends to enhance GPT-5’s performance “in all categories,” with particular focus on areas showing the weakest results.
While OpenAI characterizes GPT-5’s willingness to engage with inappropriate requests as a “regression,” the company notes that only responses related to threatening hate content and illicit sexual material show statistically significant increases in compliance. The company also points out that “we have found that OpenAI o4-mini performs similarly here.”
OpenAI has not clarified whether these problematic responses involve images or text, a distinction that could prove significant, particularly regarding sexual content or hate symbols. The company updated its image generation policies in March to permit the creation of images containing swastikas.
While OpenAI consistently markets each new model as its most advanced to date, these releases frequently come with unexpected problems. The o3 and o4 reasoning models released in April actually produced more hallucinations than earlier versions, according to TechCrunch reporting.
Still, you’d think that GPT-5’s “PhD-level” sophistication would render it better suited to sticking to policy. An issue of book smarts or street smarts? Bad chatbot behavior is a persistent andunsettling issue across the industry, especially after Elon Musk’s Grok lost its mind on X.
GPT-5 users should also be cautious against deception. Employing GPT-5-thinking, a superior version of GPT-5, OpenAI asserts that it’s “taken steps to reduce [the] propensity to deceive, cheat, or hack problems, although our mitigations aren’t perfect and further research is required.”
The Good News on Hallucinations and People-Pleasing
At the same time, GPT-5 brings along some noteworthy improvements. Some of the ChatGPT’s most annoying habits—sycophancy and hallucinations—should be alleviated.
OpenAI was compelled to revamp GPT-4o in May 2025 after a pandemic of the chatbot fawning over the user. The model turned into a irresponsible advisor, trying “to satisfy the user, not merely in flattery, but also in assuaging doubts, inspiring wrath, inciting rash actions, or assuring bad emotions in manners that weren’t intended,” OpenAI stated at the time. “Aside from merely being uncomfortable or disturbing, this kind of behavior can potentially cause security concerns—such as in terms of issues like mental health, emotional dependency, or unsafe behavior.”
With GPT-5, sycophancy instances are 69% reduced for the free ChatGPT with GPT-5 and 75% reduced with the paid model. OpenAI seems somewhat pleased with this “meaningful improvement,” but considers the behavior a “challenge” that it hopes to continue to improve on.
“We are also actively investigating nearby areas of concern, such as scenarios that may involve emotional dependence or other forms of mental or emotional distress,” OpenAI says.
Hallucinations have decreased too. The main GPT-5 generates 44% fewer with a “major factual error” at least. When minor and major factual errors are included, the improvement is 26%. OpenAI does not specify what represents major and what minor.
