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

Contributions

• A practical framework distinguishing two fundamentally different labeling goals: identifying content generated through AI versus identifying potentially misleading content
• Cross-cultural empirical evidence (5,130 participants across 5 countries) on public understanding of nine potential AI-generated content labels
• Identification of which terms best align with process-focused goals (explicitly mentioning AI) versus misleadingness-focused goals (e.g., "Deepfake")
• Evidence of cross-cultural differences in label interpretation, particularly the term "Artificial" which carries different connotations in China versus Western countries
• Actionable guidance for policymakers and platforms on label selection based on their labeling objectives

Method

The authors conducted a large online study across five countries using Qualtrics, recruiting participants from the United States (N=1,056), Mexico (N=1,060), Brazil (N=1,065), India (N=1,038), and China (N=1,011), for a total effective sample of 5,130. The study was conducted in August and September 2023 with informed consent and Institutional Review Board approval.

Study Design: Between-subjects experiment where each participant was randomly assigned to view one of nine potential labeling terms. Participants were shown 20 different pieces of content spanning the full range of AI-involvement and misleadingness (e.g., accurate infographics created by a person, AI-generated images falsely depicting real events, misleading captions on real photos, etc.) and asked whether the assigned label would apply to each piece of content.

Nine Terms Tested:

• AI Generated
• Generated with an AI Tool
• AI Manipulated
• Artificial
• Synthetic
• Deepfake
• Manipulated
• Not Real
• Edited

Measures: For each content type, participants rated:

1. Whether a label applies to that content (identifying AI-generated content)
2. Whether a label applies to misleading content
3. Subjective feelings toward labeled content and posters (using 7-point Likert scales)
4. Beliefs about whether depicted events actually occurred

The study used alignment analysis, calculating the average difference between participants' responses to the "How much AI?" and "How much human?" questions to derive continuous AI-generated content scores. Results were analyzed separately for each country and aggregated across all five countries.

Results

Perception alignment with AI-generated process:

The terms "AI Generated," "Generated with an AI Tool," and "AI Manipulated" showed the strongest and most consistent alignment with content generated using AI, significantly outperforming other terms (p<0.001). Linear regression analyses revealed these three terms were each significantly better than all other terms (p≤0.001). Conversely, "AI Generated" was one of the worst terms at identifying content that is misleading (regardless of generation process), and other process-focused terms performed similarly poorly on misleading content.

Perception alignment with misleading content:

"Deepfake" and "Manipulated" were most consistently associated with misleading content (regardless of whether it was AI-generated or human-manipulated), showing strong and consistent association across all five countries. These terms performed poorly at identifying AI-generated content specifically. The trade-off between classification goals is stark and shown visually in a Pareto frontier analysis (Figure 4).

Cross-cultural differences:

Results were remarkably consistent across the four English-speaking countries (USA, Brazil, India) and Spanish-speaking Mexico. However, a significant cross-cultural exception emerged with the term "Artificial." In China, "Artificial" performed poorly in both classification tasks and sat at the Pareto frontier, whereas in the other four countries it performed moderately on both dimensions. The authors speculate this reflects linguistic/epistemological differences: "Artificial" in Chinese (人工) emphasizes human involvement rather than machine generation, in contrast to the English connotation of "unnatural."

Subjective perceptions:

Participants reported more negative feelings toward content labeled "Deepfake" and "Manipulated" compared to other terms, and these labels were associated with lower belief in the depicted events. This suggests that labels optimized for misleading-detection may inadvertently influence belief through an "implied truth effect" — where unlabeled content increases belief.

Notes

This working paper provides important empirical guidance on a practical policy question: different labeling goals require different terminology. The distinction between process-focused labels ("AI Generated") and impact-focused labels ("Deepfake"/"Manipulated") is actionable and the cross-cultural evidence strengthens generalizability claims. The surprising finding that "Deepfake" is poorly understood as an AI-generated-specific term (participants interpret it as indicating falsehood regardless of generation process) has important implications for labeling policies that conflate process and misleadingness.

The "implied truth effect" finding — that labeling some content as misleading may increase belief in unlabeled content — deserves follow-up research on the cognitive mechanisms and practical platform implications.