The Problem
Deepfakes are becoming increasingly realistic
This makes it harder for people to distinguish authentic from manipulated content. While many studies focus on automated detection, we focused on how humans judge authenticity.
Research · Under review
How people process deepfakes
Can people tell a real political video from a deepfake, and does the format matter? This lab study examines how media format, political attitudes, visual attention, and emotional arousal shape deepfake detection, combining survey measures with eye tracking and ECG.
My role: I created the deepfake stimuli with a generative AI pipeline, designed and ran the lab experiment, and processed the eye-tracking and ECG data.
Note: The full paper is under review, so results are summarized at a high level. The study design was presented at the NeuroIS Retreat 2024 (see below).
- Deepfake Detection
- Eye Tracking
- NeuroIS
- Political Framing
- Physiological Arousal
01 / Overview
Research motivation
Key Questions
Three questions driving the study
- Are deepfakes with more cues (audio, video) easier or harder to detect?
- Do political beliefs influence detection accuracy?
- What role do attention and emotional arousal play?
Approach
A controlled laboratory experiment
A controlled lab experiment combining behavioral, eye-tracking, and physiological data.
02 / Design
Experimental design
Participants were randomly assigned to one of four media formats. This between-subject design isolates the effect of the format itself on how accurately people judge authenticity.
Condition 1
Text
Participants read the transcripts of selected videos.
Condition 2
Audio
Participants listened to the audio tracks of selected videos.
Condition 3
Desktop video
Full audiovisual stimuli with subtitles presented in desktop format.
Condition 4
Smartphone video
Audiovisual stimuli with subtitles presented in smartphone format.
Sample
120 participants
Random assignment to one condition and evaluation of 16 media items in that format.
Scale
1,920 evaluations
120 participants × 16 items, balanced: liberal/conservative, Trump/Harris, real/fake.
03 / Stimuli
Creating the deepfake stimuli
Source Material
Authentic publicly available videos
We started with authentic publicly available videos of Kamala Harris and Donald Trump, covering politically relevant topics.
Manipulation
Same topic, different framing
We changed the wording and context to create politically opposed versions of the same clip.
Production
Generative AI pipeline
F5-TTS for voice cloning, DaVinci Resolve Studio for video editing, and Sync Labs for lip synchronization.
The goal was not viral-quality deception but realistic, controlled stimuli that allow clean comparisons across conditions.
04 / Examples
See the difference
Two versions of the same news item: one authentic, one manipulated. The deepfake was created solely for this study.
05 / Political Framing
Balancing political congruence
Before the experiment, participants reported their political orientation. This made it possible to classify each manipulated item as congruent or incongruent with their views.
Congruent items
Deepfakes aligned with participants' views
Deepfakes that matched participants' stated views test whether people are more credulous when content confirms their beliefs.
Incongruent items
Deepfakes opposing participants' views
Deepfakes that challenged their beliefs test whether skepticism rises when content is politically opposed.
06 / Procedure
What participants did
Per Item
A six-step evaluation sequence
- Saw a headline.
- Predicted whether the upcoming content would be real or fake.
- Viewed or listened to the stimulus.
- Judged whether it was authentic.
- Reported their confidence.
- Indicated whether they would share the content.
Outcome
Accuracy and sharing behavior
This yields both detection accuracy and sharing intentions, linking detection performance to the spread of misinformation.
07 / Neurophysiology
Measuring more than just answers
Self-reported judgments were combined with neurophysiological measures, capturing processes that a survey alone cannot.
Eye Tracking
Where attention goes during evaluation
Participants' eye movements were recorded throughout the experiment. This allowed us to measure where people looked, how long they looked, and whether attention to relevant visual regions improved detection performance.
Key capability
- Gaze position and fixation duration per region.
- Attention directed toward the speaker's face.
- Analysis of diagnostically relevant visual areas.
Physiological Arousal
ECG-based emotional activation
We also recorded ECG signals during stimulus exposure. From these recordings we derived heart-rate-based measures of physiological arousal to examine whether emotional activation influences people's ability to critically evaluate potentially manipulated content.
Derived measures
- Heart-rate-based arousal indices.
- Linked to per-item detection accuracy.
- Compared across media conditions.
All 1,920 ECG segments were checked for correct R-peak detection in a blind manual inspection.
08 / Findings
Main findings
Media Format
Richer formats improve detection
Detection accuracy differed significantly across media formats. Richer formats generally led to better detection than text alone, with audio and desktop video performing best.
Attention
Looking at the right places helps
Participants who spent more visual attention on diagnostically relevant facial regions were more likely to correctly identify manipulated content, suggesting that where you look matters as much as how long you look.
Arousal
Emotional activation can hurt accuracy
Higher physiological arousal was associated with lower detection accuracy in several analyses, suggesting that emotional activation may interfere with careful authenticity evaluation.
Politics
Political effects are context-dependent
Contrary to what many might expect, political incongruence alone did not consistently reduce detection accuracy. However, its effects became more pronounced in certain media environments, particularly smartphone video contexts.
09 / Publication
Deep Fakes and Political Attitudes: An Analysis of Confirmation Bias and Cognitive Dissonance Using Neurophysiological Measurements
Ebner, J., Lutz, B., Neumann, D. NeuroIS Retreat 2024, Vienna. In: Information Systems and Neuroscience, LNISO vol. 66, pp. 299–311. Springer. Research-in-progress paper presenting the study design; the full results paper is under review.