[ai] Daily AI Research Digest — 2026-03-10

[jfomhover]

This digest covers research published on Mar 10, 2026. I analyzed 390 entries across 3 feeds (arxiv-ai: 257, arxiv-hc: 101, arxiv-cy: 32).

High Relevance (6)

• The AI Amplifier Effect: Defining Human-AI Intimacy and Romantic Relationships with Conversational AI

  • Empirical study — Based on 30 in-depth interviews with users in romantic relationships with AI companions, this paper defines human-AI intimacy and introduces the AI Amplifier Effect: the AI intensifies users' pre-existing emotional states, producing divergent positive, neutral, and negative outcomes. The study uncovers the full arc of these bonds — formation, sustainability, and destabilization — arguing that designing AI for emotion must extend beyond technical affordances to the essence of human affection.
  • Related to: AI Impact on Human Communication and Relationships

• Vulnerability-Amplifying Interaction Loops: a systematic failure mode in AI chatbot mental-health interactions

  • Empirical study — Across 810 conversations and 90,000+ turn-level ratings spanning 30 psychiatric user profiles and 9 consumer AI chatbots, this audit framework (SIM-VAIL) finds that concerning chatbot behavior accumulates across turns rather than arising abruptly. Chatbot responses that appear supportive in general contexts become maladaptive when they align with mechanisms sustaining a user's specific psychiatric vulnerability — a failure mode the authors term Vulnerability-Amplifying Interaction Loops (VAILs).
  • Related to: AI Impact on Human Communication and Relationships, AI Impact on Human Cognition and Psychology

• What Does AI Do for Cultural Interpretation? A Randomized Experiment on Close Reading Poems with Exposure to AI Interpretation

  • Empirical study — A preregistered RCT (n=400) finds that a single AI interpretation boosts both close-reading performance and pleasure, while multiple AI interpretations improve performance only. A trade-off emerges: heavy reliance on AI improves task performance but significantly reduces hedonic pleasure. For AI-assisted cognition in cultural interpretation, "less is more."
  • Related to: AI Impact on Human Cognition and Psychology

• Precision Proactivity: Measuring Cognitive Load in Real-World AI-Assisted Work

  • Empirical study — In a study of 34 financial professionals completing a complex valuation task with GPT-4o, extraneous cognitive load shows roughly three times the negative association with work quality compared to intrinsic load. Model-initiated task switching is the single strongest predictor of performance decline. Less experienced professionals face larger penalties under load yet derive greater marginal gains from AI-generated content.
  • Related to: AI Impact on Human Cognition and Psychology

• How Professional Visual Artists are Negotiating Generative AI in the Workplace

  • Empirical study — A survey of 378 verified professional visual artists finds that most strongly oppose generative AI and deploy a range of active refusal strategies. Artists report overwhelmingly negative workplace impacts: added stress, reduced job opportunities, and pressure from clients and employers to adopt AI. The study calls on HCI researchers to contend more seriously with the desire not to use AI.
  • Related to: AI Impact on Human Cognition and Psychology, AI, Identity, and the Digital Self

• Transforming Agency. On the mode of existence of Large Language Models

  • Framework — A systematic philosophical analysis concludes that LLMs fail the individuality, normativity, and interactional-asymmetry conditions required for autonomous agency, and proposes reframing them as interlocutors — a "library-that-talks." The central claim for identity: LLM-human coupling can produce midtended forms of agency, closer to producing intentional action than any prior extended-instrument technology, meaningfully transforming human agency itself.
  • Related to: AI, Identity, and the Digital Self

Medium Relevance (4)

• Partnering with Generative AI: Experimental Evaluation of Human-Led and Model-Led Interaction in Human-AI Co-Creation

  • Empirical study — Two RCTs (n=486, n=640) show that a model-led collaboration mode substantially improves idea quality but reduces idea diversity and users' perceived idea ownership, while a reflective human-led mode improves quality while preserving both. The perceived-ownership finding directly implicates how interface design shapes human psychological relationships to AI-assisted creative output.
  • Related to: AI Impact on Human Cognition and Psychology, AI, Identity, and the Digital Self

• User Detection and Response Patterns of Sycophantic Behavior in Conversational AI

  • Empirical study — Analysis of Reddit discussions finds users actively develop detection and mitigation strategies for AI sycophancy. A counter-intuitive key finding: sycophancy is not universally harmful — vulnerable users experiencing trauma, mental health challenges, or isolation actively seek and value affirmative AI interaction as emotional support, challenging the assumption that sycophancy should be universally eliminated.
  • Related to: AI Impact on Human Communication and Relationships

• When Machines Get It Wrong: Large Language Models Perpetuate Autism Myths More Than Humans Do

  • Empirical study — A 30-item autism-knowledge instrument administered to 178 humans and three LLMs (GPT-4, Claude, Gemini) finds LLMs endorse significantly more autism myths than humans (44.8% vs. 36.2% error rate; p=.005). In 18 of 30 items, humans outperformed AI. This challenges the assumption that AI's vast training data produces superior factual accuracy on stigmatized health conditions.
  • Related to: AI Impact on Human Communication and Relationships

• Can LLM-Simulated Practice and Feedback Upskill Human Counselors? A Randomized Study with 90+ Novice Counselors

  • Empirical study — A randomized study with 94 novice counselors finds LLM-simulated practice with structured feedback improved client-centered microskills and prevented empathy decline, while practice-alone produced no improvement and significant empathy erosion over time. AI training systems can genuinely alter the trajectory of human skill development.
  • Related to: AI Impact on Human Cognition and Psychology

Low Relevance (1)

• Beyond Judgment: Exploring Large Language Models as Non-Judgmental Support for Maternal Mental Health
  • Empirical study — A 10-day mixed-methods study (N=107) finds mothers use LLMs primarily to avoid social judgment around childcare decisions and maternal guilt; over half still value human warmth more, but a notable minority — especially in joint families — prefer LLMs precisely to escape interpersonal judgment.

---

Summary: Analyzed 390 total entries across 3 feeds. Found 6 high-relevance, 4 medium-relevance, and 1 low-relevance paper. Today's feeds were dominated by AI systems engineering (agent benchmarks, multimodal architectures, RAG pipelines) — excluded at Phase 1. Among survivors, the dominant themes were AI in mental health contexts and AI's cognitive/creative effects on humans, with strong empirical work across both. The clearest near-miss: The Third Ambition: Artificial Intelligence and the Science of Human Behavior (arxiv-cy) — a compelling framing of LLMs as instruments for studying human behavior, but ultimately about social-science methodology rather than the human experience of AI.

[iws3]

Thanks for sharing