Human detection of political speech deepfakes across transcripts, audio, and video

Matthew Groh, Aruna Sankaranarayanan, Nikhil Singh, Dong Young Kim, Andrew Lippman, Rosalind PicardNature Communications 2024 · OpenReview

Abstract

Recent advances in technology for hyper-realistic visual and audio effects provoke the concern that deepfake videos of political speeches will soon be indistinguishable from authentic video. We conduct 5 pre-registered randomized experiments with N = 2215 participants to evaluate how accurately humans distinguish real political speeches from fabrications across base rates of misinformation, audio sources, question framings with and without priming, and media modalities. We do not find base rates of misinformation have statistically significant effects on discernment. We find deepfakes with audio produced by the state-of-the-art text-to-speech algorithms are harder to discern than the same deepfakes with voice actor audio. Moreover across all experiments and question framings, we find audio and visual information enables more accurate discernment than text alone: human discernment relies more on how something is said, the audio-visual cues, than what is said, the speech content. With advances in generative AI, political speech deepfakes are becoming more realistic. Here, the authors show that people’s ability to distinguish between real and fake speeches relies on audio and visual information more than the speech content.

Introduction

As deepfake technology moves toward hyper-realism, there is a growing fear that synthetic media will become indistinguishable from reality, potentially dismantling shared standards of evidence in politics. This study investigates the "realism heuristic"—the idea that people might be more likely to trust audiovisual content because it resembles the real world more closely than text. The researchers specifically ask: Does the richness of video and audio make us more susceptible to lies, or does it provide the necessary perceptual cues to spot a fraud? Furthermore, they explore how factors like the source of the audio (AI-generated vs. voice actor) and the frequency of misinformation (base rates) affect our ability to tell truth from fiction.

Methodology

The authors conducted five pre-registered, randomized experiments involving 2,215 recruited participants, supplemented by an organic sample of over 41,000 individuals for one experiment. The experimental design systematically varied media modality (transcripts, silent video, audio-only, or full video), audio source (professional voice actors vs. state-of-the-art text-to-speech clones), and base rates of misinformation (20% vs. 80%).

The study utilized and expanded upon the Presidential Deepfake Dataset (PDD), containing 32 videos of Donald Trump and Joe Biden. The "Enhanced PDD" used 2023 state-of-the-art tools, including the wav2lip algorithm for lip-syncing, DeepFaceLab for visual touch-ups, and ElevenLabs for high-fidelity TTS voice cloning. Ordinary Least Squares (OLS) regressions were used to analyze how these variables impacted "discernment"—the ability to correctly identify both real and fabricated content.

Results

The Modality Advantage

Contrary to fears that video is a "perfect lie," the research found that audiovisual richness actually aids detection. Accuracy jumped from approximately 57% for text-only transcripts to 86% for full video in the first experiment. This suggests that humans rely heavily on how something is said (prosody, facial micro-expressions) rather than just what is said (speech content).

Figure 1: Accuracy across different media modalities. Human discernment improves significantly as more audiovisual information is provided, moving from transcripts to full video with audio.

The Text-to-Speech (TTS) Threat

While humans are generally good at spotting fakes, AI-generated voices are becoming more deceptive. Participants were significantly less accurate at spotting deepfakes using TTS audio (72% accuracy) compared to those using voice actors (83% accuracy). This indicates that state-of-the-art voice cloning is narrowing the gap between real and synthetic media.

Figure 2: Comparison of detection accuracy across PDD versions and audio sources. Note the significant drop in accuracy for deepfakes utilizing text-to-speech (TTS) in the Enhanced PDD group.

Additional Findings

  • Base Rate Insensitivity: Knowing that misinformation is common did not significantly improve detection; discernment remained stable regardless of whether participants were told fakes were rare or frequent.
  • Confidence and Accuracy: A strong positive correlation (z = 7.5) was found between a participant's confidence and their actual accuracy.
  • Passive Suspicion: Even when not explicitly prompted to check for authenticity, participants naturally expressed suspicion in free-text responses, suggesting a baseline level of "generative AI literacy."

Discussion and Limitations

The findings challenge the "seeing is believing" narrative by showing that extra sensory information provides critical "tells" that help people discern fabrication. However, several constraints remain. The stimuli focused on "talking head" political speeches; deepfakes involving complex movement or side profiles may yield different results. Furthermore, as algorithms like wav2lip evolve, the perceptual distortions humans currently use as cues may eventually disappear. Future research should explore the gap between the ability to detect a manipulation and the psychological influence of its underlying message.

Summary updated

Abstract

Recent advances in technology for hyper-realistic visual and audio effects provoke the concern that deepfake videos of political speeches will soon be indistinguishable from authentic video. We conduct 5 pre-registered randomized experiments with N = 2215 participants to evaluate how accurately humans distinguish real political speeches from fabrications across base rates of misinformation, audio sources, question framings with and without priming, and media modalities. We do not find base rates of misinformation have statistically significant effects on discernment. We find deepfakes with audio produced by the state-of-the-art text-to-speech algorithms are harder to discern than the same deepfakes with voice actor audio. Moreover across all experiments and question framings, we find audio and visual information enables more accurate discernment than text alone: human discernment relies more on how something is said, the audio-visual cues, than what is said, the speech content. With advances in generative AI, political speech deepfakes are becoming more realistic. Here, the authors show that people’s ability to distinguish between real and fake speeches relies on audio and visual information more than the speech content.
BibTeX
@inproceedings{groh2024human,
  author = {Matthew Groh and Aruna Sankaranarayanan and Nikhil Singh and Dong Young Kim and Andrew Lippman and Rosalind Picard},
  title = {{Human detection of political speech deepfakes across transcripts, audio, and video}},
  booktitle = {Nature Communications},
  year = {2024},
  url = {https://openreview.net/forum?id=9mIanUaVCL},
}
Figure