Altered neural processing of evaluative feedback from presumed conscious AI.
Antje Peters, Merit Ohlsen, Jens Bölte, Su Arkun, Laura Brockhoff, Thomas Straube, Sebastian Schindler
iScience August 21, 2026 DOI: 10.1016/j.isci.2026.117175 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Preregistered experimental EEG study Peer reviewed |
|---|---|
| Sample size | 40 |
| Population | Adult participants (N=40) who received feedback from human or AI sources |
| Intervention | ChatGPT |
| Measures | early posterior negativity (EPN), late positive potential (LPP), conscious ratings |
| Keywords | EEG/Erp Social evaluative feedback Machine consciousness Valence |
| Key findings | Participants rated human feedback as most conscious, with AICA-2 rated significantly more conscious than ChatGPT. Neural responses (EPN and LPP) were larger for human than AI feedback, and exploratory analyses showed differences between AICA-2 and ChatGPT, with the EPN effect correlating with attributed consciousness. |
Abstract
Interacting with artificial intelligence (AI) is increasingly prominent in everyday life. Modern AI can produce intelligent, witty, and even sarcastic responses, leading people to assume AI is conscious. This raises the question whether AI is perceived as conscious, and how people respond to such machines behaviorally and neurally. This preregistered electroencephalogram (EEG) study investigated the processing of social evaluative feedback from a human and two putative AI algorithms differing in attributed consciousness. Participants (N = 40) received feedback, supposedly sent by either another human, a conventional AI based on a large language model (ChatGPT), or an innovative AI capable of conscious processing (AICA-2). Human received the highest conscious ratings, while AICA-2 was rated significantly higher than ChatGPT. Event-related potentials showed enlarged early posterior negativity (EPN) and late positive potential responses to human vs. AI feedback. Exploratory analyses revealed differences between AICA-2 and ChatGPT, whereas the EPN effect correlated with attributed consciousness.