What Makes Conversational AI Worth Using?

After the first session, the plan was for her to use ChatGPT independently over the next two weeks and document her experience. I wanted to see whether repeated use would change her confidence, trust, or understanding of what ChatGPT could do.

When we met again, I learned that she had not used it at all.

She understood how to use ChatGPT, but she did not feel a reason to return to it. She also explained that something personal had happened in her life and that using ChatGPT felt like it required more mental energy than she wanted to give at the time. She preferred to use it when she felt focused.

This reinforced something I had already noticed during our earlier sessions: she could navigate the interaction, but she often did not know what she wanted to ask. Without a specific question or task in mind, the responsibility was on her to decide how to begin and how to keep the conversation going.

Rather than asking her to continue using ChatGPT simply to complete the original study plan, I changed the focus of our final session.

I wanted to understand whether giving her a more personally meaningful reason to interact would change her behavior. I refreshed her memory of the voice feature, which had been easier for her than typing, and introduced topics she frequently discussed in her everyday life, including the Philippines, Bicol, Tagalog, and travel.

This shift allowed me to move away from asking whether she could use ChatGPT and toward a more useful question: what would make her actually want to use it?

The participant understood most of ChatGPT’s responses but frequently asked, “What do I do now?”

She did not naturally see an answer as the beginning of a conversation and often did not know what to ask next. This was different from how she used apps like Facebook and YouTube. Those apps already gave her something to look at or interact with, while ChatGPT depended on her to start the interaction. When she did not already have a question or task in mind, she had difficulty finding a reason to use it. Despite ChatGPT’s onboarding giving her example prompts, she did not fully see its purpose in her everyday life.

Opportunity: Give users familiar starting points and possible follow-up questions rather than expecting them to know what to ask. One of the major problems that my participant ran into was that she simply did not know what ChatGPT was capable of.

The participant preferred voice over typing, but I noticed several points where the conversation became confusing. ChatGPT sometimes interpreted her reading its responses out loud as new input. This meant that she and the voice assistant cut each other off. Other times, ChatGPT waited for her while she was waiting for ChatGPT. She would say “ok” without realizing that she was expected to continue the conversation.

During one of these interactions, she laughed and said, “Haha, you’re funny.” While it sounded like she was having a natural conversation with ChatGPT, she did not completely understand what it was asking her. She also had trouble knowing whether ChatGPT was listening, thinking, speaking, or waiting for her.

Opportunity: Make the different voice states easier to understand and provide more direction when a conversation stops. Despite the voice states being shown at the bottom of the screen, the current placement is difficult to notice making it easily overlooked. This is especially concerning for beginners or the elderly who may have not seen these icons before. It could be beneficial to have larger icons voice states to help with recognizability.

The participant did not believe me when I told her that ChatGPT could speak her native language. It was not until I demonstrated it with her that she began trying it herself. When ChatGPT started speaking Tagalog and discussing the Philippines, she became noticeably more conversational. When it gave information she recognized as correct, she responded with things like “Yes!” and “That’s right!” and jokingly told it, “You know better than me.”

Because she already knew about these topics, she was also able to recognize when ChatGPT got something wrong. She noticed pronunciation and dialect mistakes in a few words, although they did not frustrate her enough to stop the conversation.

Another moment happened when she noticed the warning that “ChatGPT can make mistakes.” She responded with a surprised “Oh mistakes.” Until then, she had not really considered that the information ChatGPT gave her could be incorrect. The warning made her more cautious, but it did not explain what she should do if she wanted to check an answer.

Opportunity: Make language capabilities more visible and give users clearer ways to verify information. As of our sessions, the warning told her that ChatGPT could make mistakes, but the responsibility of figuring out how to check the information was still placed on her.