Are Humans or AI Better at Detecting Deepfakes Videos?

Source: Olya Kobruseva/ Pexels

The technology to create realistic fake videos using AI is becoming increasingly sophisticated, making it difficult, if not impossible, to determine whether audio, images, or videos are real. Can humans or machines tell if a video is authentic, AI-generated, or altered? Has technology gotten to the point where there is no foolproof way to identify AI-altered videos?

Manipulated videos are not a new issue; it is important to note that they can be created without AI. The advancement of AI, specifically deep neural networks and generative adversarial networks, has created sophisticated tools for realistic fake videos. One type of AI-altered video is deepfakes.

There are several types of deepfake videos, including “face swapping,” in which a face is video-grafted onto another person in order to make that person look as if they are saying or doing things that they did not actually say or do. “Lip-sync” videos are where mouths are made to move to match an audio recording. “Puppet-master” videos are where videos of a person are animated based on the movements and facial expressions of someone else sitting in front of a camera. (See this example.)

AI models need to be trained on a lot of image and video data, so targets of deepfakes are typically celebrities and politicians who have a lot of publicly available footage available.

Actor Bruce Willis recently made headlines after it was reported that he had sold the rights to his face to a Russian deepfake company Deepcake, though Willis’s agent has denied these reports. Willis’s face has reportedly been used in a Russian commercial created using deepfake “face-swapping” technology.

Val Kilmer worked with software company Sonantic to use AI to create an emotional and lifelike model of his speaking voice prior to throat cancer treatment.

Respeecher, a voice cloning AI startup, has created an algorithm to replicate the 1977 voice of Darth Vader.

As images and video data of individuals become more widely accessible online, the issue of deepfakes may become a more widespread problem for public figures and private individuals. Deepfake tools in the wrong hands can potentially violate privacy rights, spread disinformation, and cause serious financial instability and political unrest.

Researchers and technologists are working on algorithms that automate the detection of fake visual content. AI models have become capable of outcompeting human experts in a wide range of activities, from chess to medical diagnosing, so AI has the potential to help solve this problem.

The need for accurate and automated detection of deepfakes has been concerning enough that large companies, including Facebook, Microsoft, and Amazon, offered $1,000,000 in prize money to the most accurate deepfake detection models during the Deepfake Detection Challenge contest from 2019 to 2020. The top-performing model was 82.56 percent accurate against the dataset of videos that was available publicly to participants. However, in the “black box dataset” of 10,000 unforeseen videos not available to participants ahead of time, the leading model only performed at 65.18 percent accuracy.

A recent study found that ordinary human observers and leading computer vision deepfake detection AI models are similarly accurate but make different types of mistakes. People who had access to machine model predictions were more accurate, suggesting that AI-assisted collaborative decision-making could be useful but will be unlikely to be foolproof.

Researchers found that when AI makes wrong predictions and humans have access to those models' predictions, humans end up revising their answers incorrectly. This suggests that machine predictions can affect human decision-making–an important factor when designing systems of human-AI collaboration.

The problem of falsified media existed long before these AI tools. Like any technological advance, people find both positive and negative applications. AI has created exciting new possibilities with applications in creative and filmmaking industries and, at the same time, raises the need for reliable detection, protection of privacy rights, and risk management against harmful use cases.

Current research suggests that humans and machine models are imperfect at detecting AI-altered videos. One answer may be a collaborative approach between AI and human detection in order to address the shortcomings of each. Since it is unlikely for any detection model to be foolproof, education about deepfake technology can help us become more aware that seeing is not always believing—a reality that was true long before the arrival of deepfake AI tools.

Copyright © 2022 Marlynn Wei, MD, PLLC. All rights reserved.