Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Does the brain similarly process information from an AI vs. a human source during sentence comprehension?
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Olga Filimonova1, Benjamin R. Cowan2, Diego Blanco-Ortiz1,3, Alice Foucart1; 1Nebrija University, 2University College Dublin, 3Complutense University of Madrid
AI-generated content is now common in everyday communication, yet it is still unclear whether the brain processes similarly information attributed to an artificial rather than a human source. Previous ERP research has shown that the speaker’s identity (e.g., age, gender, accent) is rapidly integrated during online sentence comprehension and that it modulates how information is interpreted, but evidence on AI source remains limited. This event-related brain potential (ERP) study examined whether the processing and evaluation of world knowledge and emotional information changes when the message is attributed to a human or an AI source. Twenty-nine participants (18–35 years; M = 23.4, SD = 3.8) read sentences in Spanish while their brain activity was recorded. At the beginning of each trial, an icon indicated whether the sentence was generated by a human or an AI source. In one block, participants assessed the veracity of 60 true (a), 60 unknow (b) and 60 false statements (c). a) ‘The colour of the tongue of a dog is pink’ b) ‘The colour of the tongue of a giraffe is black’). c) ‘The colour of the tongue of a dog black’). In another block, they indicated how much they agreed with 60 moral (d) and 60 immoral statements (e) d) ‘Cheating when playing is dishonest’ e) ‘Cheating when playing is honest’ ANOVAs were conducted with the factors Content and Speaker. Behaviorally, participants rated true information as truer than unknown and false information (ps < .001) and they agreed more with moral statements than immoral ones (p < .001), independently of the speaker. At critical-word onset, ERP data in the word knowledge block revealed that unknown and false statements elicited larger N400 amplitudes than true statements (p = .002), and false statements elicited larger P600 amplitudes than unknown statements (p = .009), indicating greater processing difficulty and reanalysis. This effect did not interact with the Speaker factor. In the morality block, moral statements elicited larger P200 amplitudes than immoral statements at midline (p = .003), frontal (p = .033), and centro-parietal sites (p = .018), possibly reflecting anticipatory emotional processing. At occipital sites, this effect was significant in the human condition (p = .012), but not in the AI one (p = .497). At sentence onset, a significant effect of Speaker was observed. P200 amplitudes were larger for human-attributed statements than for AI-attributed statements across regions (world knowledge block: p = .054; morality block: p = .037). In addition, in the world knowledge block, the AI source elicited larger N400 amplitudes than the human source across regions (p = .005). These preliminary findings suggest that an AI source, compared with a human source, may modulate early attention processes and increase semantic processing difficulty at initial stages (i.e. beginning of the sentence). Although the speaker identity did not seem to affect overall processing at later stages, it may reduce prediction mechanisms when processing emotional language. These interpretations should be confirmed by future research.
Topic Areas: Meaning: Discourse and Pragmatics,