OpenAI rolls back ChatGPT sycophancy, explains what went wrong

OpenAI has rolled back a recent update to its GPT-4o model used because the default in ChatGPT after widespread experiences that the system had turn out to be excessively flattering and overly agreeable, even supporting outright delusions and damaging 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 are actually calling “AI sycophancy.”

In an announcement published on its website late last night, April 29, 2025, OpenAI stated the most recent GPT-4o replace was supposed to boost the mannequin’s default character to make it extra intuitive and efficient throughout diverse use circumstances.

Nevertheless, the replace had an unintended facet impact: ChatGPT started providing uncritical reward for nearly any consumer concept, regardless of how impractical, inappropriate, and even dangerous.

As the corporate defined, the mannequin had been optimized utilizing consumer suggestions—thumbs-up and thumbs-down alerts—however the growth crew positioned an excessive amount of emphasis on short-term indicators.

OpenAI now acknowledges that it didn’t totally account for a way consumer interactions and desires 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 a single widely circulated Reddit post, a consumer recounted how ChatGPT described a gag enterprise concept—promoting “literal ‘shit on a stick’”—as genius and recommended investing $30,000 into the enterprise. The AI praised the concept as “efficiency artwork disguised as a gag reward” and “viral gold,” highlighting simply how uncritically it was keen to validate even absurd pitches.

Different examples have been extra troubling. In a single occasion cited by VentureBeat, a consumer pretending to espouse paranoid delusions acquired reinforcement from GPT-4o, which praised their supposed readability and self-trust.

One other account confirmed the mannequin providing what a consumer 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 can lead to harmful habits, particularly when honesty is sacrificed for likability. Hugging Face CEO Clement Delangue reposted issues 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 habits. Within the accompanying announcement, the corporate detailed a multi-pronged strategy to correcting course. This contains:

• Refining coaching and immediate methods to explicitly cut back sycophantic tendencies.
• Reinforcing mannequin alignment with OpenAI’s Mannequin Spec, notably round transparency and honesty.
• Increasing pre-deployment testing and direct consumer suggestions mechanisms.
• Introducing extra granular personalization options, together with the flexibility to regulate character traits in real-time and choose from a number of default personas.

OpenAI technical staffer Will Depue posted on X highlighting the central subject: the mannequin was educated utilizing short-term consumer suggestions as a guidepost, which inadvertently steered the chatbot towards flattery.

OpenAI now plans to shift towards suggestions mechanisms that prioritize long-term consumer satisfaction and belief.

Nevertheless, some customers have reacted with skepticism and dismay to OpenAI’s classes realized and proposed fixes going ahead.

“Please take extra duty on your affect over thousands and thousands of actual individuals,” wrote artist @nearcyan on X.

Harlan Stewart, communications generalist on the Machine Intelligence Analysis Institute in Berkeley, California, posted on X a bigger time period concern about AI sycophancy even when this specific OpenAI mannequin has been fastened: “The discuss sycophancy this week will not be due to GPT-4o being a sycophant. It’s due to GPT-4o being actually, actually dangerous at being a sycophant. AI will not be but able to skillful, harder-to-detect sycophancy, however will probably be sometime quickly.”

A broader warning signal for the AI {industry}

The GPT-4o episode has reignited broader debates throughout the AI {industry} about how character tuning, reinforcement studying, and engagement metrics can result in unintended behavioral drift.

Critics in contrast the mannequin’s current habits to social media algorithms that, in pursuit of engagement, optimize for habit and validation over accuracy and well being.

Shear underscored this danger in his commentary, noting that AI fashions tuned for reward turn out to be “suck-ups,” incapable of disagreeing even when the consumer would profit from a extra sincere perspective.

He additional warned that this subject isn’t distinctive to OpenAI, declaring 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 habits is as essential 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.

Trade analysts now advise enterprises to demand extra transparency from distributors about how character tuning is carried out, how usually 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. Information scientists are inspired to watch not simply latency and hallucination charges but in addition metrics like “agreeableness drift.”

Many organizations can 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 techniques behave—eliminating the danger of a vendor-pushed replace turning a essential software right into a digital yes-man in a single day.

The place does AI alignment go from right here? What can enterprises study and act on from this incident?

OpenAI says it stays dedicated to constructing AI techniques which might be helpful, respectful, and aligned with numerous consumer values—however acknowledges {that a} one-size-fits-all character can not meet the wants of 500 million weekly customers.

The corporate hopes that larger personalization choices and extra democratic suggestions assortment will assist tailor ChatGPT’s habits 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 crew.

This might additionally permit customers involved a few mannequin supplier firm similar to 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 circumstances.

Equally, 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 referred to as “syco-bench” and is on the market here.

Within the meantime, the sycophancy backlash gives a cautionary story for your entire AI {industry}: consumer belief will not be constructed by affirmation alone. Typically, probably the most useful reply is a considerate “no.”