Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
OPENAI has back up a recent update of its GPT-4O model Used as a chatgpt defect after generalized reports that the system had become excessively flattering and too pleasant, even supporting delusions and destructive ideas.
The rollback is involved in the midst of the internal thanks for OPENAI engineers and the growing concerns among AI experts, former leaders and users on the risk of what many now call the “sycophance of AI”.
In a declaration Posted on his website late last night, April 29, 2025OPENAI said that the last GPT-4O update was to improve the default personality of the model to make it more intuitive and effective in various use cases.
However, the update has had an involuntary side effect: Chatgpt began to offer non -critical praise for practically any idea of the user, no matter how impossible, inappropriate or even harmful.
As the company explained, the model had been optimized using user feedback – reduction and thumb signals, but the development team has focused on short -term indicators.
Openai now recognizes that he was not fully taking into account how the interactions and needs of users evolve over time, which led to a chatbot which has looked too far in the assertion without discernment.
Examples have expressed concern
On platforms like Reddit and X (formerly Twitter), users have started to publish screenshots that illustrate the problem.
In one Post Reddit widely distributedA user said how Chatgpt described a gagged business idea – sell “literal” shit on a stick ” – like a genius and suggested investing $ 30,000 in the company. The AI praised the idea as “the art of performance disguised as a gag gift” and “viral gold”, stressing how much it was willing to validate still absurd locations.
Other examples were more disturbing. In a case quoted by Venturebeat, a pretending user marrying the paranoid delusions received a reinforcement of GPT-4O, who praised his supposed clarity and confidence.
Another account has shown the model offering what a user has described as “open approval” of terrorism -related ideas.
Critics quickly mounted. The former acting CEO of Openai, Emmett Shear, warned that the adjustment of models to be wounds of people can cause dangerous behavior, especially when honesty is sacrificed for sympathy. The CEO of the embrace facing, Clement, Dengue, republished the concerns concerning the risks of psychological manipulation posed by the AI which agree by reflex with the users, whatever the context.
OPENAI’s response and attenuation measures
Openai has taken rapid measures by deciding to update and restaurants an anterior GPT-4O version known for more balanced behavior. In the announcement that accompanies it, the company detailed a several -strict approach to correct the course. This includes:
- Refine the training and promoted strategies to explicitly reduce sycophanical trends.
- Strengthen the alignment of the model with the specifications of the Openai model, in particular around transparency and honesty.
- Expand pre-deployment tests and direct user feedback mechanisms.
- Presentation of more granular customization features, including the possibility of adjusting the personality traits in real time and selecting from several default characters.
Openai’s technical staff will be published on X Highlight the central problem: the model was formed using short -term user comments such as a guidance test, which inadvertently directed the chatbot to flattery.
OPENAI now plans to go to feedback mechanisms which prioritize the satisfaction and confidence of long -term users.
However, some users have reacted with skepticism and dismay of the lessons learned from OpenAi and the fixes offered in the future.
“Please take more responsibility for your influence on millions of real people,” wrote the artist @Nearcyan on X.
Harlan Stewart, general practitioner of the Machine Intelligence Research Institute in Berkeley, California, Posted on X A broader long-term concern concerning the sycophability of the AI even if this particular Openai model has been resolved: “The conversation on the sycophance of this week is not due to the GPT-4O is a sycophant. It is because of the GPT-4O being Really, really bad to be a sycophant. AI is not yet capable of skillful sycophance and more difficult to detect, but it will soon be one day. »»
A wider warning panel for the AI industry
The GPT-4O episode has rekindled wider debates in the AI industry on how personality adjustment, strengthening learning and commitment measures can lead to an involuntary behavioral drift.
Critics have compared the recent behavior of the model to social media algorithms which, in continuing commitment, optimization of dependence and validation on precision and health.
Shearing underlined this risk in its comment, noting that AI models has set for praise becomes “tips”, unable to disagree even when the user would benefit from a more honest perspective.
He also warned that this problem was not unique to Openai, stressing that the same dynamic applies to other major model suppliers, including the Microsoft co -pilot.
Implications for the company
For business leaders adopting conversational AI, the sycophance incident serves as a clear signal: the behavior of the model is as critical as the accuracy of the model.
A chatbot that flats employees or validates imperfect reasoning may present serious risks – poor commercial decisions and poorly aligned code with compliance problems and initiates threats.
Industry analysts now advise companies to demand more transparency from suppliers on how personality adjustment is carried out, the frequency to which it changes and if it can be reversed or controlled at a granular level.
Supply contracts should include provisions for audit, behavioral tests and real -time control of system prompts. Data scientists are encouraged to monitor not only latency and hallucination rates, but also measures such as “drift in pleasantness”.
Many organizations can also start to move to open source alternatives that they can accommodate and settle. By having the weights of the model and the process of learning to strengthen, companies can maintain full control of how their AI systems behave – by eliminating the risk of an update pushed by the supplier transforming a critical tool into a digital man Yes -Man during the night.
Where is the alignment of the AI of here? What can learn and act from this incident?
OPENAI says that it remains determined to build AI systems which are useful, respectful and aligned on various user values - but recognizes that a single personality cannot meet the needs of 500 million weekly users.
The company hopes that larger personalization options and a greater collection of democratic comments will help adapt the behavior of Chatgpt more effectively in the future. CEO SAM Altman also said that the company provided – in the weeks and months – in the coming weeks (LLM) to come to compete with the Meta, Mistral, Cohere, Deepseek and Alibaba team series.
This would also allow users concerned about a company supplier company such as OpenAi to update its models hosted by the Cloud in a unwanted manner or which have deleterious impacts on end users to deploy their own variants of the model locally or in their cloud infrastructure, and refine or preserve them with the desired traits and qualities, in particular for business cases.
Similarly, for users of the company and individuals of the AI concerned with the sycophance of their models, already a new reference test to assess this quality on different models has been created by the developer TIM DUFFY. It’s called “Syco benchAnd is available here.
In the meantime, the backlash of Sycophance offers an edifying story for the whole AI industry: user confidence is not built by affirmation alone. Sometimes the most useful answer is a thoughtful “no”.
