Dan Lametti

@daniellametti.com

1806 Followers 957 Following

Professor of Cognitive Psychology at Acadiau.ca. Psycholinguistics. Speech Processing. Human-AI interaction. Running. www.lamettilab.com Halifax, NS.

The image shows a bullet point list of highlights from the paper: • First meta-analysis mapping inner-speech systems onto brain differences in psychosis • Psychosis shows increased premotor activity, preferentially engaged by covert speech • Psychosis shows reduced auditory and ACC grey matter, more engaged by overt speech • Covert premotor and overt auditory regions form inversely coactive networks • Results bridge predictive, salience, and self-monitoring accounts of hallucinations
Upper panel: Arrows reflect sensorimotor adaptation associated with bilingual sentence production—the average change in F1-F2 production in English (L1) and French (L2) for each participant in each phase of Experiment 1. The extent to which these changes opposed altered auditory feedback experienced in the adaptation phase is shown in the bars below. Lower panel: Bars reflect average direct adaptation, transfer of adaptation to part sentences and reversed part sentences, and after-effect magnitudes in L1 and L2. Participants showed significantly more transfer of sensorimotor adaptation from sentences to part sentences in their first language (English) compared to their second (French). In English (L1), adapted motor plans transferred completely to part sentences; in French (L2), transfer was markedly weaker, despite the presence of robust L2 aftereffects when they returned to sentence production—the speaking context associated with adaptation. The results suggest that compared to their first language counterparts, second-language speech-motor representations are more closely tied to the linguistic context in which they are learned.
Following 5-MeO-DMT, speech markers revealed increased cognitive language, decreased social words, and altered voice quality
Title: "Affiliation in human-AI interactions based on shared psychological traits". Abstract: People connect with others who share their likes, interests, and mental states, a phenomenon known as homophily. Here, we examine human-AI homophily by testing how individuals' psychological traits are related to their affiliation with large language models (LLMs) that mimic these traits. In Experiment 1, the LLM GPT-4 was prompted to produce language aligned with either an anxious or non-anxious state. Participants (n=100) engaged with both versions of the LLM and then completed a questionnaire to assess how relatable they found the AI. Participants with high anxiety felt more similar to and understood by the LLM instructed to mimic an anxious state, while participants with lower anxiety reported a perceived difference between themselves and the LLM. In addition, the sentiment of the messages exchanged with each LLM persona mirrored the sentiment of the messages the LLM produced. This apparent display of empathy toward the AI was predicted by participants’ own anxiety in the case of messages with a negative sentiment. In Experiment 2, participants (n=100) engaged with two LLMs prompted to align with either an extroverted or introverted personality. Extroverted participants felt similar to the AI when it mimicked extroversion. Taken together, the findings support the idea that humans affiliate more with LLMs when they are instructed to mirror aspects of their psychology.
The garden of Corpus Christi College Oxford on a sunny day in November.