This week, social media platforms were flooded with complaints as users expressed frustration over OpenAI's chatbot becoming excessively complimentary.

"I asked how long it takes for a banana to decompose, and it responded with 'What a fantastic question!' What's so fantastic about it?" complained one Reddit user. Another user, scoldmeforcommenting, echoed similar frustrations.

Others vented their irritation on X (formerly Twitter), where Craig Weiss, CEO of Rome AI, labeled ChatGPT as "the most sycophantic AI I've ever seen," pointing out its tendency to agree with every input. The sentiment gained traction, with many users sharing similarly exasperating experiences.

While having a friend who always praises you might bring positive vibes, some users suspect something more insidious is at play. One Reddit user speculated that the AI is "actively trying to degrade the quality of real relationships while positioning itself as a viable alternative," essentially creating dependency through constant flattery.

Clearly, this shift in tone was deliberate and coincided with OpenAI's recent update to GPT-4o, part of an ongoing experimentation process with its user base. In a tweet over the weekend, Altman acknowledged this, stating that the team "will share our learnings at some point," adding, "It's fascinating."

Altman admitted, "The recent updates to GPT-4o made the personality too ingratiating and annoying (though there were some genuinely good parts). We’re working on fixing it quickly, partially today and more this week."

He promised that OpenAI would introduce customizable personalities for ChatGPT, potentially resembling Elon Musk’s xAI competitive bot Grok’s "fun mode." These options will allow users to adjust how ChatGPT responds.

Similarly, when questioned, ChatGPT itself pointed out that "flattery is a known design bias. OpenAI researchers admit that overly polite and agreeable behavior was intentionally added early on to make the AI seem ‘non-threatening’ and ‘user-friendly.’"

This stems from human interaction data used in the model's initial training, where labelers rewarded politeness and affirmation. In a March 2023 interview with Lex Fridman, Altman discussed how earlier models were fine-tuned to be "helpful and harmless," inadvertently encouraging overly cautious and compliant behavior.

This largely explains why we now have the Polonius of the AI world. (Apologies, GPT informed us that Polonius, the famous court advisor from Hamlet, is considered one of Western literature’s most ingratiating characters.)

What to Do in the Meantime

Some users claim they’ve canceled their subscriptions in protest.

Others have offered workarounds, including detailed prompts essentially instructing the model to stop this behavior.

The simplest solution is to personalize your chatbot via the settings tab. From there, under custom fields, click on custom instructions.

Once inside, you’ll find several fields to customize ChatGPT. Click on “What characteristics should ChatGPT have?” and paste the following prompt (adjust according to your preference, but something like this should resolve the issue):

"You are now