“As an AI language model, I cannot”: Investigating LLM Denials of User Requests

Published: May 11, 2024

Users ask large language models (LLMs) to help with their homework, for lifestyle advice, or support in making challenging decisions. Yet LLMs are often unable fulfil these requests, either as a result of technical inabilities policies restricting responses. To investigate the effect denying user we evaluate participants' perceptions different denial styles. We compare specific styles (baseline, factual, diverting, and opinionated) across two studies, respectively focusing on LLM's limitations social policy restrictions. Our results indicate significant differences users' denials between The baseline denial, which provided participants brief without any motivation, was rated significantly higher frustration lower usefulness, appropriateness, relevance. In contrast, found that generally appreciated diverting style. provide design recommendations LLM better meet peoples' expectations.

Language: Английский

---

Shaping Human-AI Collaboration: Varied Scaffolding Levels in Co-writing with Language Models

Published: May 11, 2024

Advances in language modeling have paved the way for novel human-AI co-writing experiences. This paper explores how varying levels of scaffolding from large models (LLMs) shape process. Employing a within-subjects field experiment with Latin square design, we asked participants (N=131) to respond argumentative writing prompts under three randomly sequenced conditions: no AI assistance (control), next-sentence suggestions (low scaffolding), and next-paragraph (high scaffolding). Our findings reveal U-shaped impact on quality productivity (words/time). While low did not significantly improve or productivity, high led significant improvements, especially benefiting non-regular writers less tech-savvy users. No cognitive burden was observed while using scaffolded tools, but moderate decrease text ownership satisfaction noted. results broad implications design AI-powered including need personalized mechanisms.

Language: Английский

---

Public attitudes toward chatgpt on twitter: sentiments, topics, and occupations

Social Network Analysis and Mining, Journal Year: 2024, Volume and Issue: 14(1)

Published: May 20, 2024

Language: Английский

---

ChatGPT as a voting application in direct democracy

Swiss Political Science Review, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 18, 2025

Abstract The study examines the potential role of ChatGPT as a tool for popular voting. It assesses ChatGPT's positions on four voting objects (three initiatives and one referendum) by simulating various Swiss voter profiles (neutral, centrist, left, right, progressist, conservative) comparing these to its default stance. Additionally, responses arguments against each object were analyzed consistency with patterns. compared decisions real outcomes, fixing temperature collecting 100 answers per question ensure reliability. Results showed that varied model version (GPT‐3.5 or GPT‐4) language (French German), indicating cultural iteration influences. did not consistently align left‐progressist stance, highlighting impact language, versions, contextual factors political interpretations.

Language: Английский

---

The AI-mediated communication dilemma: epistemic trust, social media, and the challenge of generative artificial intelligence

Synthese, Journal Year: 2025, Volume and Issue: 205(3)

Published: March 12, 2025

Language: Английский

---