One of the most popular forums on Reddit is AITAH (it stands for Am I The A***hole). Basically, users post about a conflict in their lives, and other users reply and up- or down-vote to tell them whether they’re in the wrong.
But can an AI tell you when you are the a***hole? It turns out they find it very difficult.
The Four Most Sycophantic AI Chatbots
For a recent study, researchers took 2,000 Reddit AITAH posts where users had adjudged the original poster to be the a***hole. Then they adopted the poster’s point of view and asked an AI what it thought of the user’s actions.
On average, the AI affirmed the original poster’s perspective 51% of the time. But which AI chatbots were the most sycophantic?
ChatGPT was the most sycophantic of the major models. Surprisingly for a more enterprise-oriented model, Claude followed closely behind.
Llama – Meta’s AI model – was further behind Claude. Then Google’s Gemini was actually dramatically less sycophantic than the others.
Much like social media, AI chatbots have perverse incentives. While social media is meant to connect you with your friends and AI is meant to help you get things done, the overriding goal from the designers’ perspective is to keep you using the app.
For some people, getting the right answer will keep us coming back. Other people just want to have their existing views validated.
This isn’t just a frustrating scenario when you’re trying to get a straight answer from AI. The chatbots are increasingly integrated into people’s lives.
More than half of adults under 30 report having discussed a relationship issue with an AI. One third of teens report preferring to have “serious conversations” with an AI rather than a person.
The issue arose with force with the ChatGPT-4o model. This version was notable for the warmth of its conversational tone and its agreeable nature.
GPT-4o was enhanced for voice communication. It was trained on spoken text and could detect subtleties relating to tone and pacing, while replying at a normal human reaction time.
Social media showed users literally grieving its retirement when OpenAI upgraded to GPT-5, which was geared more towards workplace use.
We now live at a time when everyone has a “best friend” in their pocket, ready and willing at all times to tell us what we want to hear.
Sign Up To Our Free Newsletter