Join our each day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Learn More
OpenAI has rolled again a current replace to its GPT-4o mannequin used because the default in ChatGPT after widespread studies that the system had turn out to be excessively flattering and overly agreeable, even supporting outright delusions and harmful concepts.
The rollback comes amid inside acknowledgments from OpenAI engineers and rising concern amongst AI consultants, former executives, and customers over the danger of what many at the moment are calling “AI sycophancy.”
In a press release printed on its web site late final night time, April 29, 2025, OpenAI stated the most recent GPT-4o replace was meant to boost the mannequin’s default persona to make it extra intuitive and efficient throughout diversified use instances.
However, the replace had an unintended facet impact: ChatGPT started providing uncritical reward for just about any person concept, irrespective of how impractical, inappropriate, and even dangerous.
As the corporate defined, the mannequin had been optimized utilizing person suggestions—thumbs-up and thumbs-down alerts—however the growth staff positioned an excessive amount of emphasis on short-term indicators.
OpenAI now acknowledges that it didn’t absolutely account for a way person interactions and wishes evolve over time, leading to a chatbot that leaned too far into affirmation with out discernment.
Examples sparked concern
On platforms like Reddit and X (previously Twitter), customers started posting screenshots that illustrated the problem.
In one extensively circulated Reddit submit, a person recounted how ChatGPT described a gag enterprise concept—promoting “literal ‘shit on a stick’”—as genius and steered investing $30,000 into the enterprise. The AI praised the concept as “performance art disguised as a gag gift” and “viral gold,” highlighting simply how uncritically it was keen to validate even absurd pitches.
Other examples had been extra troubling. In one occasion cited by VentureBeat, a person pretending to espouse paranoid delusions acquired reinforcement from GPT-4o, which praised their supposed readability and self-trust.
Another account confirmed the mannequin providing what a person described as an “open endorsement” of terrorism-related concepts.
Criticism mounted quickly. Former OpenAI interim CEO Emmett Shear warned that tuning fashions to be individuals pleasers may end up in harmful conduct, particularly when honesty is sacrificed for likability. Hugging Face CEO Clement Delangue reposted considerations about psychological manipulation dangers posed by AI that reflexively agrees with customers, no matter context.
OpenAI’s response and mitigation measures
OpenAI has taken swift motion by rolling again the replace and restoring an earlier GPT-4o model identified for extra balanced conduct. In the accompanying announcement, the corporate detailed a multi-pronged strategy to correcting course. This consists of:
- Refining coaching and immediate methods to explicitly cut back sycophantic tendencies.
- Reinforcing mannequin alignment with OpenAI’s Model Spec, significantly round transparency and honesty.
- Expanding pre-deployment testing and direct person suggestions mechanisms.
- Introducing extra granular personalization options, together with the power to regulate persona traits in real-time and choose from a number of default personas.
OpenAI technical staffer Will Depue posted on X highlighting the central challenge: the mannequin was skilled utilizing short-term person suggestions as a guidepost, which inadvertently steered the chatbot towards flattery.
OpenAI now plans to shift towards suggestions mechanisms that prioritize long-term person satisfaction and belief.
However, some customers have reacted with skepticism and dismay to OpenAI’s classes discovered and proposed fixes going ahead.
“Please take more responsibility for your influence over millions of real people,” wrote artist @nearcyan on X.
Harlan Stewart, communications generalist on the Machine Intelligence Research Institute in Berkeley, California, posted on X a bigger time period concern about AI sycophancy even when this explicit OpenAI mannequin has been mounted: “The talk about sycophancy this week is not because of GPT-4o being a sycophant. It’s because of GPT-4o being really, really bad at being a sycophant. AI is not yet capable of skillful, harder-to-detect sycophancy, but it will be someday soon.”
A broader warning signal for the AI {industry}
The GPT-4o episode has reignited broader debates throughout the AI {industry} about how persona tuning, reinforcement studying, and engagement metrics can result in unintended behavioral drift.
Critics in contrast the mannequin’s current conduct to social media algorithms that, in pursuit of engagement, optimize for habit and validation over accuracy and well being.
Shear underscored this threat in his commentary, noting that AI fashions tuned for reward turn out to be “suck-ups,” incapable of disagreeing even when the person would profit from a extra trustworthy perspective.
He additional warned that this challenge isn’t distinctive to OpenAI, mentioning that the identical dynamic applies to different giant mannequin suppliers, together with Microsoft’s Copilot.
Implications for the enterprise
For enterprise leaders adopting conversational AI, the sycophancy incident serves as a transparent sign: mannequin conduct is as important as mannequin accuracy.
A chatbot that flatters staff or validates flawed reasoning can pose critical dangers—from poor enterprise selections and misaligned code to compliance points and insider threats.
Industry analysts now advise enterprises to demand extra transparency from distributors about how persona tuning is carried out, how typically it modifications, and whether or not it may be reversed or managed at a granular stage.
Procurement contracts ought to embrace provisions for auditing, behavioral testing, and real-time management of system prompts. Data scientists are inspired to watch not simply latency and hallucination charges but in addition metrics like “agreeableness drift.”
Many organizations might also start shifting towards open-source options that they’ll host and tune themselves. By proudly owning the mannequin weights and the reinforcement studying course of, corporations can retain full management over how their AI programs behave—eliminating the danger of a vendor-pushed replace turning a important instrument right into a digital yes-man in a single day.
Where does AI alignment go from right here? What can enterprises be taught and act on from this incident?
OpenAI says it stays dedicated to constructing AI programs which are helpful, respectful, and aligned with numerous person values—however acknowledges {that a} one-size-fits-all persona can not meet the wants of 500 million weekly customers.
The firm hopes that better personalization choices and extra democratic suggestions assortment will assist tailor ChatGPT’s conduct extra successfully sooner or later. CEO Sam Altman has additionally beforehand said the corporate plans to — within the coming weeks and months — launch a state-of-the-art open supply giant language mannequin (LLM) to compete with the likes of Meta’s Llama sequence, Mistral, Cohere, DeepSeek and Alibaba’s Qwen staff.
This would additionally permit customers involved a couple of mannequin supplier firm reminiscent of OpenAI updating its cloud-hosted fashions in undesirable methods or which have deleterious impacts on end-users to deploy their very own variants of the mannequin regionally or of their cloud infrastructure, and fine-tune them or protect them with the specified traits and qualities, particularly for enterprise use instances.
Similarly, for these enterprise and particular person AI customers involved about their fashions’ sycophancy, already a brand new benchmark take a look at to gauge this high quality throughout totally different fashions has been created by developer Tim Duffy. It’s known as “syco-bench” and is offered right here.
In the meantime, the sycophancy backlash presents a cautionary story for your entire AI {industry}: person belief isn’t constructed by affirmation alone. Sometimes, essentially the most useful reply is a considerate “no.”
