[ai] Daily AI Research Digest — 2026-03-09

[jfomhover]

This digest covers research entries published on March 9, 2026. I analyzed 327 entries from 3 monitored feeds (arxiv-ai: 257, arxiv-hc: 45, arxiv-cy: 25).

High Relevance (6)

• Is it Me? Toward Self-Extension to AI Avatars in Virtual Reality

  • Framework — Introduces ProxyMe, a VR prototype where users embody an avatar whose voice and spoken content are modified by AI (voice cloning + LLM speech augmentation), coining "avatar self-extension" to describe situations where AI-modified communication is experienced as part of one's own expressive behavior. The paper maps research challenges around how varying degrees of AI delegation and steerability influence perceived agency, authorship, and self-identification — directly operationalizing the question of what it means for an AI system to speak as you.
  • Related to: AI, Identity, and the Digital Self

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

  • Position paper — Argues that LLMs fail the three necessary conditions for autonomous agency (individuality, normativity, and interactional asymmetry under embodied theories of mind), characterizing them instead as "libraries-that-talk" — linguistic automata devoid of autonomous agency yet capable of engaging conversationally. Crucially, argues that LLM–human coupling nevertheless produces "midtended" forms of agency, a novel mode closer to intentional agency than any prior tool-extended instrumentality, transforming where and how human agency is experienced in ways that precede clear philosophical resolution.
  • Related to: AI, Identity, and the Digital Self

• From Risk Avoidance to User Empowerment: Reframing Safety in Generative AI for Mental Health Crises

  • Position paper — Argues that most GenAI chatbots respond to mental health crises through liability-minimizing avoidance (templated referrals to crisis hotlines, refusal to engage), which may actively harm users who lack viable alternatives and reduce their motivation to seek further help. Proposes empowerment-oriented design principles drawn from community helper models — positioning AI as a supportive bridge to de-escalate crises and connect users to care, not as a liability shield — and calls for coordinated developer–regulator frameworks to balance risk mitigation with user empowerment.
  • Related to: AI Impact on Human Communication and Relationships

• What are AI researchers worried about?

  • Empirical study — Surveys more than 4,000 AI researchers (the largest such study to date) to map their concerns and compare them with public attitudes. Finds surprising convergence: only 3% of researchers prioritize existential risks despite their prominence in media and policy, and sociotechnical harms already visible today dominate both researcher and public concern. Concludes that AI research is far more cautious, plural, and publicly aligned than dominant narratives suggest, pointing to new opportunities for researcher–public dialogue on AI's real harms.
  • Related to: AI Impact on Human Cognition and Psychology

• The Values of Value in AI Adoption: Rethinking Efficiency in UX Designers' Workplaces

  • Empirical study — Through design workshops with 15 UX designers, examines how AI adoption unfolds at individual, team, and organizational scales. Finds that efficiency discourse carries hidden social and ethical dimensions: designers weigh skill development and professional worth alongside productivity, while teams negotiate responsibility, trust, and rigor. Argues that AI adoption fundamentally redistributes roles, relationships, and power, and calls for future research on how AI systems should strengthen rather than erode worker agency.
  • Related to: AI Impact on Human Cognition and Psychology

• Ambiguity Collapse by LLMs: A Taxonomy of Epistemic Risks

  • Framework — Introduces "ambiguity collapse": the phenomenon where an LLM resolves a genuinely contested, value-laden term (e.g., "hate speech," "qualified," "biased") into a single interpretation, bypassing the human deliberative practices through which meaning is ordinarily negotiated. Develops a three-level taxonomy of epistemic risks — process (foreclosing deliberation and skill development), output (distorting the concepts agents act upon), and ecosystem (reshaping shared vocabularies and how concepts evolve over time) — with particular urgency given LLMs' growing use in content moderation, hiring, and AI self-regulation.
  • Related to: AI Impact on Human Cognition and Psychology

Medium Relevance (6)

• Why Human Guidance Matters in Collaborative Vibe Coding

  • Empirical study — Across 20 experiments with 737 participants, finds that human-provided high-level instructions dramatically outperform AI-provided instructions in "vibe coding" (natural-language-directed code generation), with AI-only instruction leading to performance collapse. The optimal hybrid configuration places humans in the instruction-giving role while delegating evaluation to AI, reinforcing that uniquely human strategic judgment remains the critical bottleneck in AI-augmented task performance.
  • Related to: AI Impact on Human Cognition and Psychology

• The Malicious Technical Ecosystem: Exposing Limitations in Technical Governance of AI-Generated Non-Consensual Intimate Images of Adults

  • Policy/Governance — Identifies a "malicious technical ecosystem" comprising open-source face-swapping models and nearly 200 nudifying programs enabling non-technical users to create deepfake pornography (AIG-NCII) within minutes. Using a survivor-centered approach and NIST AI 100-4 as a benchmark for current governance, shows that existing AI oversight frameworks systematically fail to regulate this ecosystem and exposes the flawed assumptions driving those gaps.
  • Related to: AI, Identity, and the Digital Self

• Cultural Perspectives and Expectations for Generative AI: A Global Survey Approach

  • Empirical study — Large-scale global survey across Europe, the Americas, Asia, and Africa on what culture means to diverse communities and how GenAI should represent cultural artifacts, concepts, and values. Finds that religion and tradition require heightened sensitivity beyond geographic targeting, and recommends participatory design approaches alongside a cultural "redlines" sensitivity framework for responsible cultural representation in GenAI.
  • Related to: AI, Identity, and the Digital Self

• The Pen: Episodic Cognitive Assistance via an Ear-Worn Interface

  • Case study — Presents an ear-worn AI device designed for episodic rather than continuous use — assistance is intentionally invoked for short sessions and set aside when not needed. An exploratory user study finds that layered activation boundaries shape users' sense of agency, cognitive flow, and social comfort in ways that contrast with always-available AI designs, which can disrupt natural work rhythms and create discomfort around privacy and unclear system boundaries.
  • Related to: AI Impact on Human Cognition and Psychology

• "When to Hand Off, When to Work Together": Expanding Human-Agent Co-Creative Collaboration through Concurrent Interaction

  • Empirical study — Two user studies with professional designers identify five action patterns and six triggers explaining when designers choose delegation (70.1% of turns), direction (28.5%), or concurrent work (31.8%) with AI agents in creative workflows. Finds that process visibility enabled reasoning about agent actions but created friction when agents could not distinguish collaborative feedback from independent work — a core tension in human-AI co-creative collaboration.
  • Related to: AI Impact on Human Cognition and Psychology

• Human, Algorithm, or Both? Gender Bias in Human-Augmented Recruiting

  • Empirical study — Quantitative analysis of gender bias across human-only, AI-only, and hybrid recruiting on a real-world platform finds that hybrid decision-making (humans reviewing AI recommendations before searching independently) produces the fairest candidate lists — more equitable than either alone, with a measurable "more than sum of parts" effect from structured human deliberation. Provides one of the first empirical comparisons of fairness across human, AI, and hybrid hiring processes.
  • Related to: AI Impact on Human Cognition and Psychology

Low Relevance (0)

---

Summary: Analyzed 327 total entries. Found 6 high-relevance, 6 medium-relevance, and 0 low-relevance papers.

Filter statistics

• Total entries scanned: 327
• Passed Phase 1: ~32 (10%)
• Phase 2: High 6 / Medium 6 / Low 0 / Excluded ~20
• Top matched topics: AI Impact on Human Cognition and Psychology (8 entries), AI, Identity, and the Digital Self (4), AI Impact on Human Communication and Relationships (1)

What dominated today's feeds: Over 80% of entries were technical systems papers — new architectures, benchmarks, and agent evaluation frameworks — with no bearing on the human experience of AI. The clearest near-miss was The EpisTwin (knowledge graph architecture for personal AI), which used "personal AI" framing but was purely a neuro-symbolic systems paper.