Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Who Wrote the Story? Beliefs about AI Authorship Modulate the N400
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Kateřina Chládková1,2, Michaela Svoboda1,2, Maroš Filip1,2; 1Charles University, 2Czech Academy of Sciences
Language comprehension is shaped by the broader communicative context, including expectations, experience, and knowledge of the world. The N400 component of the ERPs serves as an index of semantic expectancy and contextual integration, being larger to semantically anomalous or less predictable words than non-anomalous and expected ones (Nieuwland & Van Berkum, 2006; Kutas & Federmeier, 2011), and modulated by language users’ beliefs and preferences such as degree of scepticism or voting behavior (Lindeman et al., 2008; Galli et al., 2021; Mahieux et al., 2024). While extensive ERP research has examined the lexical and contextual determinants, considerably less attention has been devoted to how beliefs about authorship modulate semantic processing. This question has become increasingly relevant with the growing presence of generative AI systems in everyday communication. Recent work suggests that readers approach AI-generated language differently from human-produced language. People are often unable to reliably distinguish AI-generated from human-generated text and rely on imperfect heuristics about what “human” language should look like (Jakesch et al., 2023). Attributed AI authorship affects perceived trustworthiness, competence, and likeability (Jia et al., 2024; Nakano et al., 2026). These effects may have consequences for predictive language processing (Rao et al., 2025): assuming that AI-generated texts are less coherent or less semantically constrained, readers may engage in weaker semantic prediction. Here we tested whether semantic violations elicit different N400 responses depending on whether participants believe the content was generated by a human author or by artificial intelligence. We hypothesized that semantic anomalies in putatively AI-generated texts would result in attenuated N400 effects relative to texts believed to be written by humans. Thirty-eight native speakers of Czech silently read two fairy-tale-like narratives presented in a RSVP paradigm. Before each story, they were told that the text had either been generated by AI or written by a human author (while, in fact, both were human-produced). Authorship manipulation and the order of stories were counterbalanced. Within each story 25 % of sentences (out of total 130) ended with a semantically incongruent word. EEG was recorded from 32 scalp electrodes and pre-processing pipelines standard for adult reading data were applied. N400 analyses focused on mean amplitudes in a 50-ms time window centered on a grand-mean negative peak between 400 ms and 500 ms after the target word onset. LME models showed that, as expected, semantically anomalous sentence endings elicited a robust N400 (mean amplitude semantically congruent words = -0.63 μV [c.i.=-1.11.. -0.147] vs. incongruent words = -3.58 μV [-4.12..-3.05]). An interaction of putative authorship, congruency, and anteriority demonstrated that the magnitude of the effect depended on believed authorship (on central-posterior sites) [slope=0.132, SE=0.027, t=4.917, p<0.001]: semantic violations in the putatively AI-generated narratives elicited significantly smaller N400 than identical violations in narratives labeled as human-generated. This demonstrates that beliefs about artificial agents interact with core mechanisms of online language comprehension: readers engage in weaker predictive processing when they believe language originates from AI rather than from a human communicator. The study contributes to emerging research on the neurocognitive consequences of human–AI interaction.
Topic Areas: Meaning: Lexical Semantics, Meaning: Discourse and Pragmatics