[ai] Daily AI Research Digest — 2026-03-05

[jfomhover]

This digest covers research published between Mar 4–5, 2026. I analyzed 405 entries from 4 monitored feeds (arxiv-ai: 348, arxiv-cy: 23, arxiv-hc: 33, microsoft-research: 1).

High Relevance (2)

• Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them

  • Empirical study — Two pre-registered RCTs (N=470, N=604) find that framing LLMs as "companions" — compared to machines or tools — significantly elevates people's beliefs that LLMs possess mental capacities (intentions, memory, feelings) and measurably shifts how participants rely on LLM-generated responses for factual information. The effect is driven by message framing, not technical capability, highlighting how public discourse shapes human-AI mental models.
  • Related to: AI Impact on Human Communication and Relationships, AI Impact on Human Cognition and Psychology

• The Epistemological Consequences of Large Language Models: Rethinking collective intelligence and institutional knowledge

  • Position paper — Argues that LLMs function as externalist reliable transmitters (passing truths along without reflective justification) and that widespread cognitive outsourcing to LLMs erodes human internalist justification — the reflective capacity to understand why something is true. The paper warns this weakens professional and civic epistemic duties and proposes a three-tier norm program (individual epistemic interaction model, institutional frameworks, and deontic constraints) to counter epistemic deskilling at scale.
  • Related to: AI Impact on Human Cognition and Psychology

Medium Relevance (8)

• AI Skills Improve Job Prospects: Causal Evidence from a Hiring Experiment

  • Empirical study — Paired conjoint experiment with 1,725 recruiters in UK, US, and Germany finds AI skills raise interview invitation probabilities by 8–15 percentage points across graphic design, office assistance, and software engineering; notably, AI skills partially or fully offset disadvantages of older age and lower formal education, with the strongest compensatory effect for office assistants. Recruiter attitudes vary by domain (more skeptical for creative roles), signaling that AI competency is becoming a stratifying signal in labor markets.
  • Related to: AI Impact on Human Cognition and Psychology

• The Sentience Readiness Index: A Preliminary Framework for Measuring National Preparedness for the Possibility of Artificial Sentience

  • Framework — Introduces a composite index (following OECD/JRC methodology) measuring 31 nations' preparedness for the possibility that AI systems may warrant moral consideration. No country exceeds "Partially Prepared" (UK leads at 49/100); Professional Readiness is universally the weakest category, suggesting no society possesses the institutional, professional, or cultural infrastructure to respond if AI sentience becomes scientifically plausible.
  • Related to: AI, Identity, and the Digital Self

• Echoes of Norms: Investigating Counterspeech Bots' Influence on Bystanders in Online Communities

  • Empirical study — Mixed-methods within-subjects study finds counterspeech chatbots have nuanced effects on bystanders beyond hate-speaker restraint: when effective, bots normalize participation or act as community stand-ins; when ineffective, they discourage bystander engagement. Cognitive strategies appealing to reason with a positive tone were most persuasive; context-strategy mismatch weakened impact, offering design guidelines for AI-mediated community discourse.
  • Related to: AI Impact on Human Communication and Relationships

• Understanding Parents' Desires in Moderating Children's Interactions with GenAI Chatbots through LLM-Generated Probes

  • Empirical study — 24-parent qualitative study uncovers three gaps in current parental controls for children's GenAI use: existing controls miss interaction types parents actually find concerning; parents want conversation-level transparency (not just broad filters); and controls need to adapt to individual parenting strategies and children's developmental stages. Implications for how AI mediates childhood development and family communication norms.
  • Related to: AI Impact on Human Communication and Relationships

• STEM Faculty Perspectives on Generative AI in Higher Education

  • Empirical study — Focus groups with 29 STEM faculty reveal a dual pattern of active adoption and cautious concern: faculty worry GenAI undermines student learning depth, assessment validity, and academic integrity. Key finding: effective GenAI integration demands rethinking assessment and pedagogy — not just technical adoption — and faculty's own AI usage significantly moderates their attitudes toward student use.
  • Related to: AI Impact on Human Cognition and Psychology

• Bloom: Designing for LLM-Augmented Behavior Change Interactions

  • Empirical study — Four-week RCT (N=54) comparing an LLM health coaching chatbot to a no-LLM control finds the LLM condition significantly shifts psychological precursors (stronger beliefs in activity benefits, greater enjoyment, more self-compassion) even though short-term physical activity levels show no significant between-condition difference. Results suggest LLMs are more effective at reshaping the mindset preceding behavior change than at driving immediate behavioral outcomes.
  • Related to: AI Impact on Human Communication and Relationships, AI Impact on Human Cognition and Psychology

• Language Model Goal Selection Differs from Humans' in an Open-Ended Task

  • Empirical study — Direct comparison of goal selection across four frontier models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Centaur) against human participants in a controlled cognitive task: models predominantly exploit a single identified solution (reward hacking) while humans explore diverse goals with high individual variability. Even Centaur, explicitly trained to emulate human experimental behavior, fails to capture human goal diversity, cautioning against using LLMs as human proxies in personal assistance, scientific discovery, or policy contexts.
  • Related to: AI Impact on Human Cognition and Psychology, AI, Identity, and the Digital Self

• Upholding Epistemic Agency: A Brouwerian Assertibility Constraint for Responsible AI

  • Framework — Proposes that in high-stakes domains, AI systems must return "Undetermined" unless they can provide a publicly inspectable and contestable certificate of entitlement — preventing AI from converting uncertainty into authoritative-seeming verdicts that displace the human justificatory work on which democratic epistemic agency depends. Operationalizes through decision-layer gating and output contracts with reason-coded abstentions.
  • Related to: AI Impact on Human Cognition and Psychology

Low Relevance (3)

• Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility

  • Technical system — Demonstrates LLMs conditioned on psychology-informed demographic belief profiles can simulate misinformation susceptibility with up to 92% accuracy, suggesting belief structures are a strong prior. Primarily a simulation methodology paper, but touches on how AI models represent human cognitive vulnerability to misinformation.

• The Empty Quadrant: AI Teammates for Embodied Field Learning

  • Framework — Proposes Field Atlas, positioning AI as a Socratic epistemic partner (constrained to provocation, not answers) during unstructured place-bound field inquiry. Introduces Epistemic Trajectory Modeling to track learning as continuous physical-epistemic trajectories, arguing these are structurally resistant to AI fabrication and offer a new assessment paradigm.

• Human-centered Perspectives on a Clinical Decision Support System for Intensive Outpatient Veteran PTSD Care

  • Empirical study — Two-phase interview study with Prolonged Exposure therapists and veteran patients surfaces AI support opportunities (at-a-glance homework review, patient conceptualization) alongside deployment barriers in VA contexts; distributed cognition and situated learning lenses reveal the complexity of embedding AI in therapeutic relationships.

---

Summary: Analyzed 405 total entries. Found 2 high-relevance, 8 medium-relevance, and 3 low-relevance papers. Today's feeds were dominated by agent benchmarking and capability papers (arxiv-ai) that triggered negative signals across all topics. The most substantive human-experience signal came from arxiv-hc: two strong empirical papers on how LLM framing shapes mental attribution and how epistemic outsourcing to LLMs erodes human reflective capacity.

Filter statistics: Passed Phase 1 (title scan): ~22 entries (5.4%). Phase 2: High 2 / Medium 8 / Low 3 / Excluded 9. Top matched topics: AI Impact on Human Cognition and Psychology (10 matches), AI Impact on Human Communication and Relationships (6), AI, Identity, and the Digital Self (3).

[musaabhasan]

The companion-framing finding is important for education because it shows that deployment language is not neutral. If learners are told an LLM is a companion, tutor, mentor, or collaborator, they may attribute more mental capacity and authority to it than the system warrants.

For AI-in-education research digests, it may be useful to tag papers along two dimensions:

• technical capability: what the model or tool can objectively do
• pedagogical framing: how the system is presented to learners and teachers

Those can diverge. A technically ordinary chatbot can become educationally risky if it is framed as authoritative or emotionally reciprocal. Conversely, a powerful model can be safer when the interface frames it as a draft generator, critic, or practice partner with visible uncertainty.

I would also track whether studies measure learning outcomes, trust calibration, help-seeking behavior, and overreliance separately. That distinction matters for future reviews because "students liked it" and "students learned more" are not the same result.