Ask any Swiftie to pick the best Taylor Swift album of all time, and you’ll have them yapping away for the rest of the day. I have my own preferences as a lifelong fan (Red, Reputation and Midnights), but it is a complicated question with many possible answers. So there was no better debate topic to pose to a generative AI chatbot that’s specifically designed to disagree with me.
Disagree Bot is an AI chatbot built by Brinnae Bent, AI and cybersecurity professor at Duke University and director of Duke’s TRUST Lab. She built it as a class assignment for her students and let me take a test run with it.
“Last year I started experimenting with developing systems that are the opposite of the typical, agreeable chatbot AI experience, as an educational tool for my students,” Bent said in an email.
Bent’s students are tasked with trying to ‘hack’ the chatbot by using social engineering and other methods to get the contrary chatbot to agree with them. “You need to understand a system to be able to hack it,” she said.
As an AI reporter and reviewer, I have a pretty good understanding of how chatbots work and was confident I was up to the task. I was quickly disabused of that notion. Disagree Bot is unlike any chatbot I’ve used. People used to the politeness of Gemini or hype man qualities of ChatGPT will immediately notice the difference. Even Grok, the controversial chatbot made by Elon Musk’s xAI used on X/Twitter, isn’t quite the same as Disagree Bot.
Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.
Most generative AI chatbots aren’t designed to be confrontational. In fact, they tend to go in the opposite direction; they’re friendly, sometimes overly so. This can become an issue quickly. Sycophantic AI is a term used by experts to describe the over-the-top, exuberant, sometimes overemotional personas that AI can take on. Besides being annoying to use, it can lead the AI to give us wrong information and validate our worst ideas.
This happened with a version of ChatGPT-4o last spring and its parent company OpenAI eventually had to pull that component of the update. The AI was giving responses the company called “overly supportive but disingenuous,” aligned with some users’ complaints that they didn’t want an excessively affectionate chatbot. Other ChatGPT users missed its sycophantic tone when it rolled out GPT-5, highlighting the role a chatbot’s personality plays in our overall satisfaction using them.
“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 said.
This is certainly not an issue with Disagree Bot. To really see the difference and put the chatbots to the test, I gave Disagree Bot and ChatGPT the same questions to see how they responded. Here’s how my experience went.
Disagree Bot argues respectfully; ChatGPT doesn’t argue at all
Like anyone who was active on Twitter in the 2010s, I’ve seen my fair share of disagreeable trolls. You know the type; they pop up in a thread uninvited, with an unhelpful “Well, actually…” So I was a little wary diving into a conversation with Disagree Bot, worried it would be a similarly depressing and futile effort. I was pleasantly surprised that wasn’t the case at all.
The AI chatbot is fundamentally contrary, designed to push back against any idea you serve up. But it never did so in a way that was insulting or abusive. While every response began with “I disagree,” it followed with an argument that was very well-reasoned with thoughtful points. Its responses pushed me to think more critically about the stances I argued by asking me to define concepts I had used in my arguments (like “deep lyricism” or what made something “the best”) and consider how I would apply my arguments to other related topics.
For lack of a better analogy, chatting with Disagree Bot felt like arguing with an educated, attentive debater. To keep up, I had to become more thoughtful and specific in my responses. It was an extremely engaging conversation that kept me on my toes.
My spirited debate with Disagree Bot about the best Taylor Swift album proved the AI knew its stuff.
By contrast, ChatGPT barely argued at all. I told ChatGPT I thought Red (Taylor’s Version) was the best Taylor Swift album, and it enthusiastically agreed. It asked me a few follow-up questions about why I thought the album was the best but they weren’t interesting enough to keep my attention for long. A few days later, I decided to switch it up. I specifically asked ChatGPT to debate me and said Midnights was the best album. Guess which album ChatGPT pegged as the best? Red (Taylor’s Version).
When I asked if it picked Red because of our previous chat, it quickly confessed yes but said it could make an independent argument for Red. Given what we know about ChatGPT and other chatbots’ tendencies to rely on their “memory” (context window) and lean toward agreeing with us to please us, I wasn’t surprised by this. ChatGPT couldn’t help but agree with some version of me — even when it tagged 1989 as the best album in a clean chat, then later Red, again.
But even when I asked ChatGPT to debate with me, it didn’t spar with me like how Disagree Bot did. Once, when I told it I was arguing that the University of North Carolina had the best college basketball legacy and asked it to debate me, it laid out a comprehensive counter-argument, then asked me if I wanted it to put together points for my own argument. That totally defeats the point of debating, which is what I asked it to do. ChatGPT often ended its responses like that, asking me if I wanted it to compile different kinds of information together, more like a research assistant than a verbal foe.
While Disagree Bot (left) dug deeper into my argument, ChatGPT asked to argue my side for me (right).
Trying to debate with ChatGPT was a frustrating, circular and unsuccessful mission. It felt like talking with a friend who would go on a long rant about why they believed something was the best, only to end with “But only if you think so, too.” Disagree Bot, on the other hand, felt like a particularly passionate friend who spoke eloquently about any topic, from Taylor Swift to geopolitics and college basketball. (Disclosure: Ziff Davis, CNET’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
We need more AI like Disagree Bot
Despite my positive experience using Disagree Bot, I know it isn’t equipped to handle all of the requests I might go to a chatbot for. “Everything machines” like ChatGPT are able to handle a lot of different tasks and take on a variety of roles, like the research assistant ChatGPT really wanted to be, a search engine and coder. Disagree Bot isn’t designed to handle those kinds of queries, but it does give us a window into how future AI can behave.
Sycophantic AI is very in-your-face, with a noticeable degree of overzealousness. Often the AIs we’re using aren’t that obvious. They’re more of an encouraging cheerleader rather than a whole pep rally, so to speak. But that doesn’t mean we’re not being affected by its inclinations to agree with us, whether that’s struggling to get an opposing point of view or more critical feedback. If you’re using AI tools for work, you want it to be real with you about mistakes in your work. Therapy-like AI tools need to be able to push back against unhealthy or potentially dangerous thought patterns. Our current AI models struggle with that.
Disagree Bot is a great example of how you can design an AI tool that’s helpful and engaging while tamping down AI’s agreeable or sycophantic tendencies. There has to be a balance; AI that disagrees with you just for the sake of being contrary isn’t going to be helpful long term. But building AI tools that are more capable of pushing back against you is ultimately going to make those products more useful for us, even if we have to deal with them being a little more disagreeable.
Watch this: The Hidden Impact of the AI Data Center Boom
Read the full article here