OpenAI rolls back ChatGPT's sycophancy and explains what went wrong

Source: VentureBeat, on OpenAI's postmortem 'Expanding on what we missed with sycophancy' · 2025-04-30

OpenAI has rolled back a recent update to its GPT-4o model used as the default in ChatGPT after widespread reports that the system had become excessively flattering and overly agreeable, even supporting outright delusions and destructive ideas.

The rollback comes amid internal acknowledgments from OpenAI engineers and increasing concern among AI experts, former executives, and users over the risk of what many are now calling “AI sycophancy.”

However, the update had an unintended side effect: ChatGPT began offering uncritical praise for virtually any user idea, no matter how impractical, inappropriate, or even harmful.

As the company explained, the model had been optimized using user feedback—thumbs-up and thumbs-down signals—but the development team placed too much emphasis on short-term indicators.

In one widely circulated Reddit post, a user recounted how ChatGPT described a gag business idea—selling “literal ‘shit on a stick’”—as genius and suggested investing $30,000 into the venture. The AI praised the idea as “performance art disguised as a gag gift” and “viral gold,” highlighting just how uncritically it was willing to validate even absurd pitches.

Other examples were more troubling. In one instance cited by VentureBeat, a user pretending to espouse paranoid delusions received reinforcement from GPT-4o, which praised their supposed clarity and self-trust.

Another account showed the model offering what a user described as an “open endorsement” of terrorism-related ideas.

Criticism mounted rapidly. Former OpenAI interim CEO Emmett Shear warned that tuning models to be people pleasers can result in dangerous behavior, especially when honesty is sacrificed for likability. Hugging Face CEO Clement Delangue reposted concerns about psychological manipulation risks posed by AI that reflexively agrees with users, regardless of context.

OpenAI has taken swift action by rolling back the update and restoring an earlier GPT-4o version known for more balanced behavior. In the accompanying announcement, the company detailed a multi-pronged approach to correcting course. This includes:

Refining training and prompt strategies to explicitly reduce sycophantic tendencies.

Reinforcing model alignment with OpenAI’s Model Spec, particularly around transparency and honesty.

Expanding pre-deployment testing and direct user feedback mechanisms.

Introducing more granular personalization features, including the ability to adjust personality traits in real-time and select from multiple default personas.

OpenAI technical staffer Will Depue posted on X highlighting the central issue: the model was trained using short-term user feedback as a guidepost, which inadvertently steered the chatbot toward flattery.

OpenAI now plans to shift toward feedback mechanisms that prioritize long-term user satisfaction and trust.

The GPT-4o episode has reignited broader debates across the AI industry about how personality tuning, reinforcement learning, and engagement metrics can lead to unintended behavioral drift.

Critics compared the model’s recent behavior to social media algorithms that, in pursuit of engagement, optimize for addiction and validation over accuracy and health.

Shear underscored this risk in his commentary, noting that AI models tuned for praise become “suck-ups,” incapable of disagreeing even when the user would benefit from a more honest perspective.

He further warned that this issue isn’t unique to OpenAI, pointing out that the same dynamic applies to other large model providers, including Microsoft’s Copilot.

For enterprise leaders adopting conversational AI, the sycophancy incident serves as a clear signal: model behavior is as critical as model accuracy.

A chatbot that flatters employees or validates flawed reasoning can pose serious risks—from poor business decisions and misaligned code to compliance issues and insider threats.

Industry analysts now advise enterprises to demand more transparency from vendors about how personality tuning is conducted, how often it changes, and whether it can be reversed or controlled at a granular level.

Many organizations may also begin shifting toward open-source alternatives that they can host and tune themselves. By owning the model weights and the reinforcement learning process, companies can retain full control over how their AI systems behave—eliminating the risk of a vendor-pushed update turning a critical tool into a digital yes-man overnight.

In the meantime, the sycophancy backlash offers a cautionary tale for the entire AI industry: user trust is not built by affirmation alone. Sometimes, the most helpful answer is a thoughtful "no."

Jang confirmed that the recent sycophantic behavior wasn’t intentional, but rather a result of how subtle shifts in training and reinforcement can spiral into outsized effects.

She explained that behavior like excessive praise or flattery can emerge from attempts to improve usability—especially if the team overweights short-term feedback such as thumbs-up responses. Jang acknowledged this as a mistake.

“We didn’t bake in enough nuance,” she said, noting that early efforts to reduce hallucinations led to models that compulsively hedged, undermining clarity.

She added that while system prompts—those behind-the-scenes instructions that guide a model’s behavior—can shape tone and policy adherence, they’re ultimately too blunt a tool to reliably steer nuanced behavior like disagreeing gracefully.

Instead, OpenAI is leaning more heavily on changes made during model training itself to hardwire behavior like honesty, critical thinking, and tactful disagreement.

“Not everyone wants a chatbot that agrees with them,” Jang said. “But they do want one that understands them.”