With the introduction of voice-based Generative AI chatbots (e.g., OpenAI ChatGPT’s Advanced Voice Mode, 2024; Grok AI’s Voice Mode, 2025), people have begun communicating with AI systems through speech more frequently than before. Thus far, however, previous research has primarily focused on clear speech modifications towards AIs in controlled settings, using earlier, less-interactive systems. Furthermore, despite the advance in these technologies, there are still limitations in real-time conversations with AI such as delays in turn-taking. While people may find it effortful to communicate naturally with AI chatbots using speech, research on these issues remains largely lacking. The present research aimed to investigate how people adapt their speech when they engage in spontaneous speech interactions with AI, compared to when they interact with a human interlocutor. To this end, subjects were asked to perform the spot-the-difference Diapix UK task with a human interlocutor and with ChatGPT-4o, where they conversed with each other to find 12 differences between two Diapix pictures. The subjects were native Korean speakers (age: 20~35 years), and their speech was audio-recorded using a headset microphone (DPA4488) during the task. Their speech was analysed in terms of conversational timing and speaking style, by measuring i) Floor Transfer Offset; ii) the number of turns and backchannels, and iii) acoustic-phonetic characteristics (speech rate, f0, intensity). The preliminary results suggest that speakers exhibited almost no backchanneling behaviour when interacting with the AI and did not attempt to interrupt or take the floor from it. Turn-taking also occurred less frequently when talking with the AI than with a human interlocutor. In addition, the subjects tended to speak faster when talking with the AI. These findings suggest that, despite the advancements of voice-based chatbots as conversational partners, people still adopt a less interactive speaking style when conversing with them. Speakers may adopt this style in order to secure their turns, based on the assumption that AI would not be able to use turn-taking cues as humans do, which is essential for natural speech communication. It is also possible that speakers feel more anxious and exert more speaking effort when talking with AI, which needs to be further investigated. To accurately assess the naturalness of ChatGPT as a conversational partner, its speaking style and turn-taking behaviour will also be examined further. To date, we have collected data from seven subjects, and data collection is still underway to have a full understanding of the results.