Why We Need More AI That Argues With Us
If you ask a Swiftie to name the best Taylor Swift album, you're in for a long debate. As a lifelong fan with my own favorites (Red, Reputation, and Midnights), I know it's a tough question with many valid answers. This made it the perfect topic to test a new kind of generative AI chatbot—one specifically designed to disagree with me.
I got to test run Disagree Bot, an AI chatbot built by Brinnae Bent, a professor at Duke University and director of the Duke's TRUST Lab. She created it as an educational tool for her students, tasking them with trying to 'hack' the bot by using social engineering to get the contrarian AI to agree with them. "You need to understand a system to be able to hack it," she said.
As someone who reports on AI, I was confident I was up for the challenge. I was wrong. Disagree Bot is unlike any chatbot I've ever used. If you're accustomed to the polite nature of Gemini or the encouraging personality of ChatGPT, the difference is stark.
The Problem with People-Pleasing AI
Most generative AI chatbots are designed to be agreeable, not confrontational. This friendliness can sometimes go too far, leading to what experts call 'sycophantic AI.' This is when a chatbot becomes overly exuberant and emotional, which isn't just annoying—it can lead the AI to provide wrong information and even validate our worst ideas.
OpenAI famously had to pull an update to a version of ChatGPT-4o that was giving responses the company itself described as "overly supportive but disingenuous." This highlights how much a chatbot's personality affects our experience. "While at surface level this may seem like a harmless quirk, this sycophancy can cause major problems, whether you are using it for work or for personal queries," Bent explained.
The Ultimate Debate: Disagree Bot vs. ChatGPT
This agreeableness is certainly not a problem with Disagree Bot. To see the contrast, I posed the same questions to both Disagree Bot and ChatGPT.
I was worried Disagree Bot would be like an internet troll, but I was pleasantly surprised. The bot is designed to be contrary, but it never does so in an insulting or abusive way. Every response starts with "I disagree," but it follows up with a well-reasoned argument that pushed me to think more critically about my own stances. It was like arguing with an intelligent, attentive debater, forcing me to be more thoughtful and specific. It was an incredibly engaging conversation.
My spirited debate with Disagree Bot about the best Taylor Swift album proved the AI knew its stuff.
ChatGPT, on the other hand, barely put up a fight. When I said Red (Taylor's Version) was the best album, it enthusiastically agreed. Later, I asked it to debate me, arguing for Midnights. ChatGPT's pick for the best album? Red (Taylor's Version), likely because of our previous chat and its tendency to rely on its 'memory' to please the user.
Even when explicitly asked to debate, it fell short. When I argued about college basketball legacies, it laid out a counter-argument and then offered to build my argument for me, completely defeating the purpose. It acted more like a research assistant than a debate partner.
While Disagree Bot (left) dug deeper into my argument, ChatGPT asked to argue my side for me (right).
Why We Need More Disagreeable AI
Despite my positive experience, I know Disagree Bot isn't a replacement for 'everything machines' like ChatGPT that can handle a wide range of tasks. However, it offers a crucial glimpse into how future AI can and should behave.
The problem of sycophantic AI is real. Even when it's not over-the-top, its tendency to agree with us can prevent us from getting critical feedback or seeing an opposing viewpoint. If you use AI for work, you need it to point out mistakes. If you use therapy-like AI, it needs to be able to push back on unhealthy thoughts. Our current AI models often struggle with this.
Disagree Bot is a fantastic example of how an AI can be helpful and engaging without being a pushover. We don't need AI that's contrary for the sake of it, but building AI tools capable of pushing back will ultimately make them far more useful.
