Can dialogues with AI systems help humans better discern visual misinformation?

TL;DR

This study addresses how AI dialogue about AI-generated images can help humans discern visual misinformation. Using 80 participants and 1,310 dialogues, the authors find a strong short-term improvement in accuracy from ~60% to ~90% ($p<0.001$), but no durable transfer to unseen content without AI, which returns to ~60% ($p=0.88$). The work highlights the persuasive power of AI dialogue for immediate belief revision yet underscores a gap in lasting, generalizable learning, motivating interventions that promote durable understanding. It also evaluates prompting strategies, showing potential but also limitations in achieving transfer across novel examples, informing the design of AI-assisted media literacy tools.

Abstract

The widespread emergence of manipulated news media content poses significant challenges to online information integrity. This study investigates whether dialogues with AI about AI-generated images and associated news statements can increase human discernment abilities and foster short-term learning in detecting misinformation. We conducted a study with 80 participants who engaged in structured dialogues with an AI system about news headline-image pairs, generating 1,310 human-AI dialogue exchanges. Results show that AI interaction significantly boosts participants' accuracy in identifying real versus fake news content from approximately 60\% to 90\% (p$<$0.001). However, these improvements do not persist when participants are presented with new, unseen image-statement pairs without AI assistance, with accuracy returning to baseline levels (~60\%, p=0.88). These findings suggest that while AI systems can effectively change immediate beliefs about specific content through persuasive dialogue, they may not produce lasting improvements that transfer to novel examples, highlighting the need for developing more effective interventions that promote durable learning outcomes.

Figures (3)

• Figure 1: Examples of real and fabricated news stories alongside performance metrics.
• Figure 2: Prompting Strategies Description for News headline and image Credibility Assessment. The optimal strategy (Prompt 3) integrates approaches for artifact detection and persuasion to enhance human-AI dialogue to distinguish between real and fake news headline image pairs.
• Figure 3: Interaction based on the AI system trained on Persuasion. After participants report whether they have seen the news or not, they interact with the AI system. In step 1: participants rate their beliefs which is treated as initial accuracy. Step 2 shows one interaction and response from the system based on the participant's response. After three rounds of interaction, participants rate their belief which is treated as final accuracy.