Research · Fall 2025
AI (Mis)attribution
AI (Mis)attribution: Perceiver Expectations About Agent Origin as a Driver of Social Evaluations in Text-Based Interactions
Research question
Does the label change the judgment?
An online experiment asking whether people judge the same interaction differently when they believe artificial intelligence is involved.
Short answer
What did the study find?
Participants rated the conversation as more natural when they believed both speakers were human, even though the transcript was unchanged.
This online course study was last reviewed on . The interpretation and limitations below apply to this study's specific text-based setting.
Research question
Do people judge the same interaction differently when they believe artificial intelligence is involved?
Theoretical background
People interpret a message alongside beliefs about its source. Research on agency, social evaluation, and human–AI interaction suggests that labels can shape expectations before a conversation is evaluated.
Hypotheses
Study One predicted that origin labels would influence naturalness, humanness, and thoughtfulness ratings. The warm-versus-cold extension tested whether tone would alter evaluations differently for human-labeled and AI-labeled responses.
Study One design
The study was conducted online. One hundred twenty-three participants were assigned equally to Human–Human, Human–AI, or AI–AI conditions. Each person evaluated the same approximately 446-word transcript. Only the stated origin of the speakers changed.
Questionnaire and participant materials were created in Qualtrics. Data were analyzed in IBM SPSS.
Main finding
Naturalness differed across the three label conditions, F(2, 120) = 8.05, p < .001. The Human–Human condition received higher ratings than Human–AI and AI–AI. The two AI-involved labels did not differ significantly from each other.
This supports a bounded conclusion: in this study, the label surrounding the exchange affected how the unchanged conversation was evaluated.
Proposed and completed extension
The second study crossed responder identity with warm or cold tone. The strongest pattern was that warmth improved evaluations of a human-labeled response, while the same tone contrast did not change evaluations of an AI-labeled response in the same way.
Limitations
The research used convenience university samples and one text-based academic context. Those limits matter. The results should not be generalized to all people, all AI systems, or all forms of conversation.
Reflection
The study made context impossible to ignore. People were not only responding to words; they were responding to what they believed those words represented.
Method at a glance
The control was the conversation itself.
Held constant
- The same approximately 446-word transcript
- The online evaluation flow
- The response scales
Changed
- The stated origin of the two speakers
- Human–Human, Human–AI, or AI–AI label
Available evidence
What can—and cannot—be claimed.
Available outcomes
- Study One found higher naturalness ratings in the Human–Human condition than in either condition labeled as involving AI
Limitations
- Convenience university sample
- Single text-based academic interaction
- Results should not be generalized to every AI interaction
Responsible interpretation
A difference in this setting, not a universal rule.
The findings show that origin beliefs affected social evaluation in the study's text-based context. They do not establish that all people reject AI, that labels always dominate content, or that any system can control judgment.