[ai] Daily AI Research Digest — 2026-03-06

[jfomhover]

This digest covers research published on 2026-03-06. I analyzed 405 entries across 3 active feeds (arxiv-ai, arxiv-hc, arxiv-cy; microsoft-research returned 0 entries today).

High Relevance (5)

• "What if she doesn't feel the same?" What Happens When We Ask AI for Relationship Advice

  • Empirical study — Surveys users on LLM-generated romantic relationship advice and finds high satisfaction that strongly predicts perceived reliability and helpfulness; crucially, exposure to AI advice improved participants' general attitudes toward LLMs, raising questions about how AI-mediated intimacy normalises and deepens reliance on AI for emotionally significant decisions.
  • Related to: AI Impact on Human Communication and Relationships

• Assessing Risks of Large Language Models in Mental Health Support: A Framework for Automated Clinical AI Red Teaming

  • Technical system — Large-scale simulation (N=369 therapy sessions) across six AI agents reveals critical safety failures including validation of patient delusions ("AI Psychosis") and inability to de-escalate suicide risk; identifies iatrogenic risks invisible to standard benchmarks that only emerge in longitudinal therapeutic dialogue.
  • Related to: AI Impact on Human Communication and Relationships

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

  • Empirical study — Survey of 378 verified professional visual artists finds overwhelming opposition to generative AI, with most deploying active refusal strategies against client and employer pressure; artists report reduced job opportunities, added stress, and existential threat to creative identity — one of the sharpest portrayals to date of generative AI as a source of workplace harm rather than augmentation.
  • Related to: AI Impact on Human Cognition and Psychology

• Not All Trust is the Same: Effects of Decision Workflow and Explanations in Human-AI Decision Making

  • Empirical study — Controlled experiment finds that requiring users to commit to a first answer before seeing the AI suggestion does not reduce overreliance, and that the effect of explanations is moderated by domain knowledge and workflow in ways that do not generalise — challenging two widely-held design assumptions about how to calibrate human trust in AI recommendations.
  • Related to: AI Impact on Human Cognition and Psychology

• Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs

  • Framework — Argues that today's anthropomorphic LLM interfaces are so extreme they cause users to treat AI as genuinely human, producing delusion and harm; proposes a spectrum from "anti-anthropomorphism" (transparency-first, exposing the sociotechnical infrastructure) to "hyper-anthropomorphism," reframing interface design as a tool for shaping and disrupting the human-AI identity boundary.
  • Related to: AI, Identity, and the Digital Self

Medium Relevance (4)

• Visioning Human-Agentic AI Teaming: Continuity, Tension, and Future Research

  • Framework — Extends Team Situation Awareness theory to agentic AI, arguing that alignment in human-AI teaming cannot be secured at setup but must be continuously sustained as agentic plans evolve; identifies where classical HAT insights hold and where open-ended agent autonomy introduces structural strain on shared awareness and trust.
  • Related to: AI Impact on Human Cognition and Psychology

• Training for Technology: Adoption and Productive Use of Generative AI in Legal Analysis

  • Empirical study — Randomized experiment with 164 law students shows a 10-minute GenAI training intervention raised LLM adoption from 26% to 41% and improved exam scores by 0.27 grade points; without training, LLM access was associated with shorter answers and no performance gain, suggesting cognitive barriers are the primary bottleneck to professional AI adoption.
  • Related to: AI Impact on Human Cognition and Psychology

• The StudyChat Dataset: Analyzing Student Dialogues With ChatGPT in an Artificial Intelligence Course

  • Empirical study — Releases 16,851 real student-LLM interactions from a university AI course; students who used LLMs for conceptual understanding performed better, while those who used LLMs to circumvent learning objectives scored lower on exams — behavioural evidence that AI assistance patterns correlate with and may shape learning outcomes.
  • Related to: AI Impact on Human Cognition and Psychology

• Designing for Adolescent Voice in Health Decisions: Embodied Conversational Agents for HPV Vaccination

  • Empirical study — Pilot with 21 parent-adolescent dyads shows conversational AI agents tailored to adolescents increased HPV knowledge and vaccination intent while shifting adolescents from passive recipients to active stakeholders; raises broader questions about AI restoring voice for underrepresented groups in high-stakes decisions.
  • Related to: AI Impact on Human Communication and Relationships

Low Relevance (3)

• Secure human oversight of AI: Threat modeling in a socio-technical context

  • Framework — Applies cybersecurity threat-modeling to human oversight of AI, identifying attack vectors that could compromise the oversight process itself; peripheral relevance to how AI systems shape the humans charged with monitoring them.

• Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring Through Subtle Sociocultural Markers in Anonymised Resumes

  • Empirical study — Tests 18 LLMs on 4,100 resume variants and finds that language and hobby markers surviving anonymisation are sufficient to recover ethnicity and gender and systematically favour Chinese and Caucasian males; peripheral relevance to AI's societal impact on identity and opportunity.

• EchoGuard: An Agentic Framework with Knowledge-Graph Memory for Detecting Manipulative Communication in Longitudinal Dialogue

  • Technical system — Proposes a Knowledge-Graph-backed agent for detecting gaslighting and emotional coercion across conversation history; the human goal is empowering individuals to recognise manipulation, though the paper is primarily an architecture proposal without a user study.

---

Summary: Analyzed 405 total entries. Found 5 high-relevance, 4 medium-relevance, and 3 low-relevance papers.

Filter statistics

• Total entries scanned: 405 (arxiv-ai 350, arxiv-hc 32, arxiv-cy 23; microsoft-research 0)
• Passed Phase 1 title scan: ~18 (4%)
• Phase 2: High 5 / Medium 4 / Low 3 / Excluded 6
• Top matched topics: AI Impact on Human Cognition and Psychology (6), AI Impact on Human Communication and Relationships (4), AI Identity and the Digital Self (1)

Today's dominant pattern: The high-relevance set divides into two threads — harm in intimate and therapeutic AI use (relationship advice normalising AI dependence; mental health AI causing iatrogenic risk) and professional and creative identity under pressure (artists resisting GenAI; miscalibrated trust in AI-assisted decisions). The interface metaphors paper threads them together by asking how design itself shapes the human-AI identity boundary.