Designing Effective Digital Literacy Interventions for Boosting Deepfake Discernment

Paper Title

Designing Effective Digital Literacy Interventions for Boosting Deepfake Discernment

Publication Info

• Topic area: Digital literacy interventions to improve deepfake image detection.
• Keywords: Deepfake images, digital literacy, misinformation, AI-generated content, discernment accuracy, behavioral interventions, visual misinformation, gamification, feedback, AI literacy.

Background and Problem

• Problem / challenge: People struggle to discern real images from AI-generated deepfake images, which can erode trust, spread disinformation, and influence public opinion. Existing digital literacy interventions are limited in addressing visual misinformation, particularly deepfake images.
• Significance: Improving the ability to detect deepfake images is critical to maintaining trust in authentic content, reducing the spread of misinformation, and safeguarding democratic processes.
• Motivation and related work: Previous research has shown that digital literacy interventions can improve textual misinformation detection, but their application to visual misinformation remains underexplored. Existing approaches, such as reverse image search or media literacy tips, are often cognitively demanding and impractical for everyday use. Additionally, interventions risk increasing skepticism toward authentic content.

Solution

• Proposed approach: A comparative evaluation of five lightweight digital literacy interventions designed to improve deepfake image discernment while maintaining trust in real images.
• Novelty:
  1. Systematic evaluation of five intervention formats: textual, visual, gamified, feedback-based, and knowledge-based.
  2. Assessment of both immediate and long-term effects on deepfake and real image discernment.
  3. Examination of whether interventions increase skepticism toward authentic content.
  4. Analysis of sharing intentions to understand behavioral implications.
• Procedure and key techniques:
  1. Conducted a large-scale experiment with N = 1,200 U.S. participants, randomly assigned to one of five intervention conditions or a control group.
  2. Participants completed a discernment task (classifying 15 images as real or fake) and indicated their willingness to share the images on social media.
  3. Follow-up conducted two weeks later to assess long-term effects.
  4. Interventions included:
     • Textual: Short descriptions of common deepfake errors.
     • Visual: Textual descriptions paired with example images.
     • Gamified: Interactive game with points and feedback.
     • Feedback: Repeated classification tasks with immediate feedback.
     • Knowledge: Explanation of AI image generation processes.

Results

• Concrete findings:
  • The Textual and Visual interventions improved deepfake detection accuracy by 7.5 and 13 percentage points, respectively (both p < 0.01).
  • No significant improvements were observed for Gamified, Feedback, or Knowledge interventions.
  • Intervention effects did not persist significantly at the two-week follow-up.
  • None of the interventions increased skepticism toward real images.
• Advantage over baselines:
  • Visual intervention outperformed the control group with the highest improvement in deepfake detection (+13 percentage points).
  • Textual intervention provided measurable benefits with minimal cognitive load.
• Experiments / evaluation:
  • Metrics: Accuracy in discerning real vs. fake images, sharing intention, and self-reported confidence.
  • Dataset: 15 images per participant, including real and deepfake images (both viral and non-viral).
  • Validation study with N = 600 participants confirmed findings.
• Limitations and future work:
  • Interventions did not produce lasting effects beyond the immediate session.
  • Study focused only on deepfake images; transferability to other modalities (e.g., video, audio) is untested.
  • Real-world applicability in fast-paced online environments remains to be evaluated.
  • Future research should explore repeated exposure, personalized interventions, and their impact on trust in institutions.

Summary

This study evaluates five digital literacy interventions for improving deepfake image detection. The Textual and Visual interventions significantly enhanced discernment accuracy immediately after training, with the Visual intervention showing the greatest improvement. Importantly, none of the interventions increased skepticism toward real images. However, effects did not persist at the two-week follow-up, highlighting the need for sustained or repeated interventions. These findings provide actionable insights for scalable, user-focused strategies to combat visual misinformation while maintaining trust in authentic content.