philip lelyveld The world of entertainment technology

What label should be applied to content produced by generative AI?

PhilNote: this Stanford / MIT academic study quantified the public perception of 9 ways of labeling AI content. Spoiler alert; "Across all five countries, we find that participants most consistently associated ``AI Generated,'' ``Generated with an AI tool'' and ``AI manipulated'' with content that is generated using AI, regardless of whether that content is misleading - and thus are not consistently associate with misleading content."

Abstract

With the rise of generative AI, there has been a recent push for disclosing if content is produced by AI. However, it is not clear what the right term(s) are to use for such labels. In this paper, we investigate how the public understands the mapping between nine potential labeling terms (selected through consultation with academics, technology companies, and policy organizations) and twenty different types of content that vary in the extent to which they are AI-generated, and the extent to which they are potentially misleading. To do so, we conducted a study with samples from the U.S. (N=1056), Mexico (N=1060), Brazil (N=1065), India, (N=1038), and China (N=1031). Participants were randomly assigned to one of the nine terms, and were then shown descriptions of each of the 20 content types and asked whether the term applied to that type of content. Across all five countries, we find that participants most consistently associated ``AI Generated,'' ``Generated with an AI tool'' and ``AI manipulated'' with content that is generated using AI, regardless of whether that content is misleading - and thus are not consistently associate with misleading content. Conversely, participants most consistently associated ``Deepfake'' and ``Manipulated'' with content that is misleading, regardless of whether the misleading content was generated using AI. These patterns are robust across demographic subsets based on age, gender, education, digital literacy, and familiarity with generated AI. Interestingly, ``Artificial'' does fairly well at both classification tasks in the US, Brazil, India, and Mexico, but does very poorly in China because the term has a different connotation when translated. Finally, we examined participants' subjective assessments of how using a given term as a label would affect their belief in the veracity of the labeled content, and their attitudes towards the content and the poster. We observe a negative correlation between these two outcomes, such that the terms that participants thought would most effectively reduce belief were also the terms that created the most negative feelings towards the content and poster. These results have important implications for how and where generative AI disclosure is implemented, and suggest that platforms and civil society must decide carefully what their objective is for such disclosures.

See the full paper here: https://psyarxiv.com/v4mfz