A Representative Study on Human Detection of Artificially Generated Media Across Countries

View on arXiv ← Back to list

Authors: Joel Frank, Franziska Herbert, Jonas Ricker, Lea Schönherr, Thorsten Eisenhofer, Asja Fischer, Markus Dürmuth, Thorsten Holz

Published: 2023-12-10 19:34:52+00:00

Comment: Security and Privacy 2024 (S&P 24)

AI Summary

This paper presents the first comprehensive cross-country (USA, Germany, China) and cross-media (audio, image, text) survey to evaluate human ability to detect AI-generated media. The study involved 3,002 participants and found that state-of-the-art forgeries are nearly indistinguishable from real media, with most participants guessing randomly. Key personal factors like generalized trust, cognitive reflection, and self-reported familiarity with deepfakes significantly influence detection decisions across all media categories.

Abstract

AI-generated media has become a threat to our digital society as we know it. These forgeries can be created automatically and on a large scale based on publicly available technology. Recognizing this challenge, academics and practitioners have proposed a multitude of automatic detection strategies to detect such artificial media. However, in contrast to these technical advances, the human perception of generated media has not been thoroughly studied yet. In this paper, we aim at closing this research gap. We perform the first comprehensive survey into people's ability to detect generated media, spanning three countries (USA, Germany, and China) with 3,002 participants across audio, image, and text media. Our results indicate that state-of-the-art forgeries are almost indistinguishable from real media, with the majority of participants simply guessing when asked to rate them as human- or machine-generated. In addition, AI-generated media receive is voted more human like across all media types and all countries. To further understand which factors influence people's ability to detect generated media, we include personal variables, chosen based on a literature review in the domains of deepfake and fake news research. In a regression analysis, we found that generalized trust, cognitive reflection, and self-reported familiarity with deepfakes significantly influence participant's decision across all media categories.

Key findings

State-of-the-art AI-generated media across audio, image, and text are almost indistinguishable from real media, with participants often resorting to random guessing and frequently rating fakes as human-generated. Human detection accuracy for images was often worse than random, while German participants showed slightly better performance for audio, potentially due to lower generative quality of German audio. Generalized trust, cognitive reflection, and self-reported familiarity with deepfakes consistently influenced detection outcomes across all media types.

Approach

The researchers conducted a large-scale online survey with 3,002 participants across three countries (USA, Germany, China) to assess their ability to discern human- versus machine-generated audio, image, and text media. They then performed exploratory and Bayesian multilevel-binomial linear mixed effect model regression analyses, including personal and cognitive variables, to understand influencing factors on detection accuracy.

Datasets

Audio: LJSpeech dataset, CSMSC dataset, HUI dataset. Image: Subset of Nightingale and Farid [26]'s dataset (generated with StyleGAN2), StyleGAN2 training dataset. Text: National Public Radio (NPR), Tagesschau, China Central Television (CCTV).

Model(s)

UNKNOWN

Author countries

Germany

← Previous