Why speaking to AI may be better for brainstorming

As generative AI becomes embedded in classrooms and workplaces, so has the growing concern that these systems may be flattening human thinking. When AI produces fluent, plausible answers on demand, it can be tempting to accept those responses instead of wrestling with ideas independently.

In 2023, Dan Wang, the Lambert Family Professor of Social Enterprise at Columbia Business School, took a different approach. Instead of limiting AI use, Wang and his former students Jill Cohen ‘20 and Johnny Lee ‘23 developed CAiSEY, an AI-powered platform that places students in structured, debate-style conversations with an informed AI adversary. Students must take a position, defend it, and respond in real time as the AI pushes back.

Now, new research by Wang and Neelam Jain, a postdoctoral researcher at CBS, is analyzing the greater impact of how we communicate with AI tools like CAiSEY. Their findings show that the way we communicate with AI may be shaping the ideas we ultimately produce.

It’s Not the AI, It’s the Interface

Much of the anxiety around generative AI stems from studies of text-based interactions. People prompt a system, receive an answer, and refine it, often converging on similar language and ideas. That raises a more fundamental question of if AI itself is driving that convergence, or is it the fact that most interactions happen through typing?

“We wanted to ask something more fundamental. Rather than just texting with AI, what if we talk to it too?” Jain said. “And how does that impact this idea of the homogenization of thought?”

For decades, scholars have argued that speaking and writing engage different cognitive processes. Spoken language unfolds in real time, without the ability to revise but writing, by contrast, allows for editing, compression, and refinement. If those differences hold in AI interactions, then the perceived effects of AI on thinking may depend as much on how we communicate as on the system itself, according to Wang and Jain.

In a study supported by the AI and Business Initiative and the Tamer Institute for Social Enterprise and Climate Change at CBS, the researchers analyzed 957 debate-style conversations between MBA and EMBA students and CAiSEY in a Strategy course. 

Each interaction followed the same structure, with students reading a case, choosing a position on an open-ended question, and engaging in a back-and-forth debate with the AI, which automatically adopted the opposing stance. The AI’s knowledge, prompts, and responses were held constant.

The key variable was that students were able to choose whether to interact with CAiSEY by voice or by text. This created a natural comparison between two modes of engaging with the same system on the same task. While students were not randomly assigned to one mode, the dataset allowed researchers to account for individual tendencies and case-specific factors.

Messier Talk, Bigger Ideas

CAiSEY’s structure is central to the study. Unlike many AI tools that aim to be helpful and agreeable, CAiSEY is intentionally adversarial: students take a stance and the AI challenges it each time.

That friction is by design. “If you have a conversational partner that is more likely to question your assumptions…that’s going to put your thinking outside of your typical comfort zone,” Wang says. “That type of thinking is likely to be more divergent and more creative.”

CAiSEY’s responses are carefully structured to sustain productive dialogue rather than shut it down. This combination—structured opposition and real-time interaction—creates an environment where students must actively generate, defend, and extend their ideas.

The results revealed that students who spoke to CAiSEY produced longer conversations. They used more words, sent more messages, and repeated themselves more often than those who typed. These spoken responses were less concise and less polished, but that messiness turned out to be meaningful.

“People tend to say a lot more when they’re talking versus texting, but in doing so, they’re also repeating themselves a lot more,” Jain said.

That repetition functioned as a kind of scaffolding—helping students maintain coherence while continuing to explore new directions, according to the researchers.

The data shows that voice-based interactions were associated with significantly higher divergent ideation, a measure of how many distinct ideas a person generates. Even after accounting for the greater length of spoken conversations, voice users explored a broader conceptual space than their text-based counterparts.

As Jain put it, “Even though people who are speaking verbally repeat themselves a lot more, they’re also conveying a wider variety of arguments.”

By contrast, text-based interactions favored concision and refinement. Students tended to develop and polish a smaller set of ideas rather than branching outward.

Match the Medium to the Mindset

The findings suggest that AI’s impact on thinking is not fixed, but actually depends on how people engage with it.

Voice interaction appears to support exploration, helping users generate and connect a wider range of ideas in real time. Text interaction supports exploitation—refining, narrowing, and sharpening those ideas.

“If your goal is brainstorming, then speaking to an AI is way better for this,” Wang said. “If your goal is to analyze something, texting is going to be way better.”

The implications extend beyond the classroom. In Wang’s Technology Strategy course, students who use CAiSEY often arrive more prepared and more engaged, having already tested their ideas against an opposing viewpoint. “They can anticipate a wide range of arguments and are ready to think on their feet,” he said.

The research findings also point to a design principle for AI systems: they can be built to deliver answers, but also to shape how people conceive ideas.