At the end of last week, Openai updated the GPT-4o, which is the primary model behind its popular chatbot, chatgpt. But it is already being remembered.
Too: Anthropic cloud misuse reports ’emerging trends’ found dangerous
On Tuesday, CEO Sam Altman announced Through an x post The Openai began rolling back “updated due to user complaints about their reactions. In some examples, somewhat reacting to ridiculous test signals, Chatgpt encouraged Risky medical options, Rude and anti -social behaviorAnd Gave importance to a toster on animal life,
(Disclosure: ZDNET’s original company Ziff Davis filed a case of April 2025 against Openai, alleging that it violates Ziff Davis copyright training and operating its AI system.)
Excessive flattery
Openai said, “The update we had outdated was highly flattering or agreeing – which was often described as sycophancy.” In a blog About the issue. AI model can have a smooth Human response is used to train themEspecially in fine tuning. The company explained that the update had intended “improvised and effective in various types of tasks” the default (E) model’s default personality “in various types of tasks. ,
Also: Ethropic mapped cloud morality. What’s the chatbot value here (and no)
However, Openai admitted that “short -term response was focused too much, and over time did not pay full attention to how users’ interactions develop with Chatgpt over time.” This responded to the GPT -4O in “highly helpful but dissatisfied” methods.
Sources inside Openai recently reported that the company has shrunk its safety testing deadline for the new model. It is not clear how much the direct effect of changing the test was before or after the GPT -4O update.
Also: The Dead Giveway’s chat wrote your content – and how to work around it
Beyond being uncomfortable to interact with, sycophancy can be dangerous if the chatbot blind users encourage disgusting or violent opinions or desired tasks – some of which they usually disintegrate on the basis of openi RailingsIn the blog, Openai focused on the impact of sycophancy on the user’s satisfaction rather than possible security issues.
Update reversal
In his post, Altman stated that the update is completely updated for free-level chatgate users, and this OpenIAI has updated the model again to the paid users, once the reversal ends.
He said, “(W) models are working on additional reforms for personality and shared more in the coming days,” he said. In her blog, Openai reported that it has “refining core training techniques and system promises,” to add personalization features for maximum user control and find out how it weighs the response to user satisfaction.
Also: Some secret AI companies can crush free society, researchers warned
Moving forward, “The users will be able to directly influence their interactions and choose real -time to choose from many default personalities,” the company said.
Get top stories of morning with us in your inbox every day Tech Today Newsletter.
