Deepfakes and the Liar’s Dividend: From Deception to Denial in the Erosion of Public Trust

Abstract

Generative artificial intelligence can now generate fake audio and video at a relatively low cost, and it is very difficult to tell whether it is real; thus, public trust in what they see and hear is starting to erode. Communication and political science research have shown that, in the context of public and political life, the most serious trust-related harm caused by deepfakes is not widespread deception of the audience but rather the "liar's dividend": as the public becomes more aware that any recording may be fake, political and other parties can dismiss genuine evidence as false. Exposure to deepfakes more frequently results in uncertainty rather than a fixed false belief, and this uncertainty erodes trust in the news and in the evidentiary value of recordings. An increasing number of studies in law, communication and political science support the above model, although their extent is still unknown. Three documented cases in the United States are relevant to this issue: an AI-cloned political robocall, the rejection of an authentic campaign photograph as "AI", and a "deepfake defence" presented in court. The liar's dividend is empirically real but limited; it has been more reliably employed against textual evidence than against video evidence, and the main damage has been to public trust in all evidence.