While AI systems are much better than humans at spotting deepfake images, humans may still have an advantage when it comes to deepfake videos. This is a surprising development from a new study pitting humans against machines in the race to detect digital counterfeiting. The results suggest that humans and machines will need to work together to identify and counter deepfakes in the future, psychologist Natalie Ebner and colleagues reported on January 7. Cognitive research: principles and their implications.
Deepfakes are AI-generated images, audio, and videos that can falsely represent what a person looks like, says, or does, and are already being used to commit financial fraud, influence elections, and damage reputations. They are becoming more persuasive at an alarming rate, fooling humans and AI models alike.
To determine whether humans or machines were better at detecting deepfakes, Ebner and his colleagues first asked roughly 2,200 participants and two machine learning algorithms to rate the authenticity of 200 faces on a scale of 1 (fake) to 10 (real). Humans were only able to identify deepfakes at the probability level, or about 50% of the time. However, the machines performed better, with one algorithm getting the correct answer about 97 percent of the time, and the other having an average accuracy of 79 percent.
The researchers then asked approximately 1,900 human participants to watch 70 short videos of people discussing a topic and rate how realistic the faces of the people were. Remarkably, humans outperformed algorithms on this task. Human participants came up with the correct answer an average of 63% of the time, but the algorithm performed at near-chance levels.
Researchers are now looking more closely at both human and AI decision-making. We want to know, “What do machines use to be so much better at certain conditions than humans? And how does that differ from how humans reason? What do humans see and perceive in their brains?” said Ebner of the University of Florida, Gainesville. “We are now looking at different angles of human and machine to understand, not just explain yes or no. why Are they making a yes or no decision? ”
The researchers claim that knowledge will help humans work with AI to figure out how best to survive a deepfake-saturated future.
Source link
