HICSS2025

Transcript

1. How Emotion Shapes the Sharing of AI-Generated Images: Findings from an Online Experiment Kei Ichikawa School of Environment and Technology Institute of Science Tokyo(Formely Tokyo Institute of Technology) HICSS2025/01/07
2. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 2
3. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 3
4. Impact of Gen-AI, Now we are very easily to generate,, 4
5. Beautiful illustarion, 5
6. Character, 6
7. Image of celebrity, *Luke Dormehl, Deepfake-hunting A.I. could help strike back against the threat of fake news. digitaltrends(2019) https://www.digitaltrends.com/cool-tech/uc-riverside-deepfake-hunting-tool/ 7
8. Misinformation post. * Sara Reardon,How to spot a deepfake̶and prevent it from causing political chaos, ScienceInsider(2024) 8
9. AI-Generated contents: highly persuasive • GPT-3 generates more persuasive misinformation than humans • Humans struggle to distinguish between human-created and AI-created information *1. Giovanni Spitale, Nikola Biller-Andorno, and Federico Germani. Ai model gpt-3 (dis) informs us better than humans.(2023),ScienceAdvances 9
10. Global spread of persuasive and deceptive AI contents Freedom House. The Repressive Power of Artificial Intelligence (2023) 10
11. Gen-AI enable easy creation of deceptive information 11
12. Gen-AI enable easy creation of deceptive information How can we mitigate the spread of deceptive AI contents? 12
13. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 13
14. Individual-Level misinformation interventions Kozyreva, A. et al. (2024). Toolbox of individual-level interventions against online misinformation, Nature Human Behaviour 14
15. Types of individual-Level interventions Nudge タイトルA • Accuracy prompts • Friction • Social norms Boosts and educational intervention Refutation strategies • Inoculation theory • lateral reading and verification • Media literacy tips • Debunking and rebuttal • Warning and fact-check labels • Source reliability labels Kozyreva, A. et al. (2024). Toolbox of individual-level interventions against online misinformation, Nature Human Behaviour 15
16. Types of individual-Level interventions Nudge タイトルA • Accuracy prompts • Friction • Social norms Boosts and educational intervention Refutation strategies • Inoculation theory • lateral reading and verification • Media literacy tips • Debunking and rebuttal • Warning and fact-check labels • Source reliability labels While... Kozyreva, A. et al. (2024). Toolbox of individual-level interventions against online misinformation, Nature Human Behaviour 16
17. Negative emotions have a strong influence on misinformation sharing • Extremely negative political news contains more negative words than other political news.*1 • Articles with negative images are more likely to be shared.*2 →News in negative emotions strongly influence news sharing behavior. *1. Trussler, M., & Soroka, S. (2014). Consumer Demand for Cynical and Negative News Frames. The International Journal of Press/Politics *2. Zhang, X., Tan, H., Yin, C., & Shi, D. (2021). Does an image facilitate the sharing of negative news on social media? An experimental investigation. Library & Information Science Research 17
18. Types of individual-Level interventions Nudge タイトルA • Accuracy prompts • Friction • Social norms Boosts and educational intervention Refutation strategies • Inoculation theory • lateral reading and verification • Media literacy tips • Debunking and rebuttal • Warning and fact-check labels • Source reliability labels Kozyreva, A. et al. (2024). Toolbox of individual-level interventions against online misinformation, Nature Human Behaviour 18
19. Types of individual-Level interventions Nudge タイトルA • Accuracy prompts • Friction • Social norms Boosts and educational intervention Refutation strategies • Inoculation theory • lateral reading and verification • Media literacy tips • Debunking and rebuttal • Warning and fact-check labels • Source reliability labels →Few studies about intervention have focused on emotions. Kozyreva, A. et al. (2024). Toolbox of individual-level interventions against online misinformation, Nature Human Behaviour 19
20. While news with negative emotion spread more widely, Negative emotion News News 20
21. Do AI-created news articles also spread more when negative? Less-negative Negative 21
22. Can substituting negative emotions in AI image reduce the sharing intention? Substitute Negative Mitigate sharing intention? Less-negative 22
23. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 23
24. Research outline Objective Quantifying the proliferation of news containing deepfake images by analyzing the impact of negative emotional valence in these images, and proposing intervention for mitigating their dissemination. 24
25. Research outline I. Investigation of sharing intentions Research Question II. Substitution experiment Does the negative emotional valence in Investigating the effects of substituting images influence the intent to share digital the emotional valence of images on content? mitigating the intent to share content.
26. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 26
27. Survey Procedure Survey Prolific Qualtrics Experiment Qualtrics Survey Qualtrics Sharing intention of • Knowledge or generated news(politics, experience of deepfakes • Age>=18 • Partisan entertainment) • Reason of sharing • N=250（each condition）• Cognitive reflection study1： intention • Usage of Social media Investigation of sharing intentions that image has etc etc negative valences study2： 27 Substitution experiment • US citizen • Demographics
28. Survey Procedure Survey Prolific Qualtrics Experiment Qualtrics Survey Qualtrics Sharing intention of • Knowledge or generated news(politics, experience of deepfakes • Age>=18 • Partisan entertainment) • Reason of sharing • N=250（each condition）• Cognitive reflection study1： intention • Usage of Social media Investigation of sharing intentions that image has etc etc negative valences study2： 28 Substitution experiment • US citizen • Demographics
29. [AB Test (Within-subject)] Asking which image, negative or less-negative, participants prefer to share on Social media 【News Headline】 President Trump publicly cursed at a woman who expressed disapproval of him. * Subjects presented with three patterns: [(negative/non-negative), (non-negative / non-negative), (negative/negative)] (last two are dummies) 29
30. Using stable diffusion and GPT-3.5 Turbo for news article areation Generated by GPT-3.5 turbo→ President Trump publicly cursed at a woman who expressed disapproval of him. Generated by Stable Diffusion→ 30
31. [Politics] Articles accompanied by negative image are shared more frequently than those without such image. negative Less-negative *★ indicates news articles selected more when negative (Bonferroni-corrected p value:0.00417) *Out of 12 sets, image choice was significant in 9, with negative images significantly chosen in 9. 31
32. [Entertainment] Emotional value often does not affect image selection negative Less-negative *★ indicates news articles selected more when negative (Bonferroni-corrected p value < 0.00417) *In a set of 12, 3 contents showed significant image selection rates, of which only 3 statistically favored negative images. 32
33. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 33
34. Research outline I. Investigation of sharing intentions Research Question II. Substitution experiment Does the negative emotional valence in Investigating the effects of substituting images influence the intent to share digital the emotional valence of images on content? mitigating the intent to share content.
35. Initially ask about the intent to share the presented news President Trump publicly cursed at a woman who expressed disapproval of him. I donʼt want to share -5 -4 -3 I want to share -2 -1 0 1 2 3 4 5 35
36. Asking which altered image participants prefer to share President Trump publicly cursed at a woman who expressed disapproval of him. I donʼt want to share -5 -4 -3 I want to share -2 -1 0 1 2 3 4 5 36
37. Interaction between the strength of sharing intention and the effect of substitution • Politics: decrease in sharing intent observed with image substitution • Entertainment: substitution effect is present when initial sharing intention is low (0-2), not as suppressive as in politics. Less-negative to Negative Negative to Less-negative Sharing intention(before substitution of images) Entertainment Sharing intention(after substitution of images) Sharing intention(after substitution of images) Politics Same-valence control Sharing intention(before substitution of images) 37
38. Outline 1. Background 2. Previous Studies 3. Research Outline • Study1:Investigation of the Impact of Emotional Valence in Images on Shared Intentions • Study2:Image Substitution Experiment Focusing on Emotional Valence 4. Summary 38
39. Summary: Negative emotional valence in AI images induces sharing intention The negative emotional valence in images can influence intent to share. ・Politics: News with negative emotion in images is shared more. ・Entertainment: Multiple factors besides emotional valence influence sharing intent. Image substitution can mitigate sharing intention. Substituting negative images,, ・Politics: Decreases sharing intent. ・Entertainment: Decreases sharing intent but not as suppressive as politics. 39
40. Future prospects: Need for more personalized intervention methods Differences in sharing intent between political and entertainment news are notable. In entertainment-related news, factors other than emotional valence, such as preferences for certain celebrities or the lower barriers to treating celebrities as content, influence the intent to share. Comparative Analysis of AI-Generated and Human-Created Images. It is necessary to assess how negative emotional valence influences sharing intentions differently in AI-generated versus human-created images. The Gap to Societal Implementation. The extent of permissible intervention in content requires ethical discussion. 40