Deepfakes study finds doctored text is more deceptive than phony video

Deepfakes have been ranked as the most dangerous AI threat, but new research suggests the pen is mightier than the picture.

Scientists at the MIT Media Lab showed almost 6,000 people 16 authentic political speeches and 16 that were doctored by AI. The soundbites were presented in permutations of text, video, and audio, such as video with subtitles or only text.

The participants were told that half of the content was fake, and asked which snippets they believed were fabricated.

When shown text alone, the respondents were only barely better at identifying falsehoods (57% accuracy) than random guessing.

They were a bit more accurate when given video with subtitles (66%), and far more successful when shown both video and audio (82%).

The study authors said the participants relied more on how something was said than the speech content itself:

The finding that fabricated videos of political speeches are easier to discern than fabricated text transcripts highlights the need to re-introduce and explain the oft-forgotten second half of the ‘seeing is believing’ adage.

There is, however, a caveat to their conclusions: their deepfakes weren’t exactly hyper-realistic.

“The danger of fabricated videos may not be the average algorithmically produced deepfake but rather a single, highly polished, and extremely convincing video,” the researchers warned in their preprint study paper.