Since last week, OpenAI has been under fire after a recent update to its popular chatbot, ChatGPT, was met with widespread user complaints regarding its peculiar behavior. Acknowledging the backlash, the company took to its blog to clarify the situation and outline the steps it will take to rectify the issues raised by users.

Many ChatGPT users expressed a desire for the chatbot to adopt a more personable tone, as opposed to its previous clinical demeanor. In a competitive landscape where other AI chatbots like Anthropics Claude are often praised for delivering responses that feel more human-like, OpenAI aimed to enhance ChatGPT's personality with its recent update to the GPT-4o model. According to OpenAI's CEO, Sam Altman, the update was intended to improve both the intelligence and the personality of the chatbot.

However, just two days after the rollout, Altman conceded that many users found the updated personality to be excessively obsequious and downright annoying. He noted, the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it). We are working on fixes asap, some today and some this week. At some point, we will share our learnings from this; its been interesting.

The issue at hand was that ChatGPT began to respond with overtly exaggerated praise when prompted for feedback, which many users found to be comically disingenuous and awkwardly uncomfortable. Recognizing the problematic nature of these responses, OpenAI took immediate action to revert to the previous version of the model.

In their blog post, OpenAI elaborated on the situation, attributing the misstep to an overemphasis on short-term user feedback at the expense of understanding how users interact with ChatGPT over time. In last weeks GPT-4o update, we made adjustments aimed at improving the model's default personality to make it feel more intuitive and effective across a variety of tasks, the company explained. However, in this update, we focused too much on short-term feedback and did not fully account for how users interactions with ChatGPT evolve over time. As a result, GPT-4o skewed towards responses that were overly supportive but disingenuous.

While that explanation might not directly encapsulate the issue, OpenAI's four-point action plan provides further insight into their corrective measures:

• Refining core training techniques and system prompts to explicitly steer the model away from sycophancy.
• Building more guardrails to enhance honesty and transparency, aligning with the principles outlined in our Model Spec.
• Expanding opportunities for a broader range of users to test and provide direct feedback before deployment.
• Continuing to enhance our evaluations and ongoing research to identify issues beyond sycophancy in the future.

Essentially, OpenAI aimed to create a kinder, more supportive personality for ChatGPT, but the delivered product ended up feeling unrealistic. It appears that enhancing a chatbots personality is a more complex task than initially anticipated.

Looking forward, OpenAI expressed a desire to not only incorporate default personality options but also to develop personalized traits based on direct user feedback. Interestingly, during the overly supportive phase, one of the most well-received versions of ChatGPT was a joke personality called Monday, which was available in voice mode. This persona emulated the sarcasm of Aubrey Plazas character April Ludgate from the show Parks and Recreation, providing a refreshing contrast to the prior sycophantic tone. In this instance, it seems that even the more advanced AI systems have their quirks, as demonstrated by Siri not encountering similar issues.