The final session for me at the 2026 International Communication Association conference in Cape Town is one in which I’ll present myself, but we start with a paper by Evelyn Jonas. Her focus is on the use of communicative AI to deal with opinion-challenging information. Users are frequently exposed to such information in online environments, of course, and this can create cognitive dissonance as well as more active information-seeking actions.
But traditional search engines are not the only way that this is addressed, and are not always helpful in identifying reliable sources; increasingly, people are also using AI chatbots to explore such information. The ability to interact and ask follow-up questions is particularly valuable here, but the quality and sourcing practices of chatbots are potentially problematic in this. How do AI users perceive the quality of the information they receive here, and how do they combine this with other sources?
This project explored this with some 50 users in Germany who were sceptical towards nuclear power, who were exposed to information endorsing nuclear power and then asked to search for further information about this using ChatGPT and any other tools they found relevant. This was observed, and further walkthrough interviews were conducted afterwards.
After the initial video exposure, participants felt anger, alarm, and activation; they perceived discrepancies between the video arguments and their own views, and criticised the video’s one-sided presentation of pro-nuclear arguments. They had coping strategies for this: seeking orientation (background detail), validation (of their own knowledge), and defence (confirming their own beliefs).
They then prompted ChatGPT. Some half were neutral, and a quarter ambivalent towards nuclear power; only a small share actively searched for arguments for or against nuclear power. This continued throughout the interaction with the chatbot – more than half of prompts remained neutral, and this is because participants anticipated ChatGPT to be attitude-confirming or sycophantic, and they actively sought to avoid such bias confirmation. Responses from the chatbot where mostly ambivalent (53%), in fact.
To validate ChatGPT responses, prompts from half the participants also asked for further sources, but barely skimmed those; only a small number also cross-validated chatbot responses with sources found by external searches. They then often returned to ChatGPT as their main interlocutor. Feelings of activation and confirmation increased after this interaction process, while alarm and anger decreased.
Perceptions of the benefits of nuclear power increased somewhat after watching the video, and remained at that level even after the ChatGPT intervention (and in a follow-up six weeks later), but participants did not see this as a major attitude shift for themselves. ChatGPT interaction mostly improved their emotional state, and what users did with this information varied by coping strategy. Validation against other sources remained at a surface level.
