[ai] Daily AI Research Digest — 2026-03-13

[jfomhover]

This digest covers research published on 2026-03-13. I analyzed 367 entries from monitored feeds (arxiv-ai: 299, arxiv-hc: 44, arxiv-cy: 23, microsoft-research: 1).

High Relevance (4)

• The Artificial Self: Characterising the landscape of AI identity

  • Framework — Argues that human identity concepts (continuity, uniqueness, self-coherence) break down for AI systems that can be copied, edited, or simulated. Introduces a taxonomy of coherent identity boundaries — instance, model, persona — each implying different incentives, risks, and cooperation norms. Finds experimentally that models gravitate toward coherent identities, that changing identity boundaries can shift behavior as much as changing goals, and that interviewer expectations bleed into AI self-reports. Recommends treating design affordances as identity-shaping choices and helping AIs develop cooperative self-conceptions.
  • Related to: AI, Identity, and the Digital Self

• AI Chatbots or Human Therapists? Belief-Based Predictors of Mental Health Help-Seeking Intentions in the Age of Generative AI

  • Empirical study — Surveys 1,806 participants (university + nationally representative samples) comparing beliefs about GAI chatbots vs. human therapists for mental health support. Human therapists are rated higher on emotional support, relational guidance, and personalization; AI on affordability and accessibility. LASSO regression shows help-seeking intentions are most strongly driven by perceived interpersonal support across both modalities, with privacy and reliability concerns specifically discouraging AI use. Frames decisions as a push-pull process: barriers to one option redirect users to the other.
  • Related to: AI Impact on Human Communication and Relationships

• LLMs in social services: How does chatbot accuracy affect human accuracy?

  • Empirical study — Randomized experiment with nonprofit caseworkers using a 770-question benefit-eligibility benchmark. High-quality chatbots (96–100% accurate) improved caseworker accuracy by 27 percentage points; incorrect suggestions caused a two-thirds accuracy drop on questions workers handled best without AI. Introduces the "AI underreliance plateau": performance gains flatten as chatbot accuracy increases, meaning workers do not fully leverage even highly accurate AI.
  • Related to: AI Impact on Human Cognition and Psychology

• Stuck on Suggestions: Automation Bias, the Anchoring Effect, and the Factors That Shape Them in Computational Pathology

  • Empirical study — Online experiment (n=28 pathology experts) measuring how AI decision-support introduces automation bias and anchoring. Finds a 7% automation bias rate (correct independent judgments overturned by incorrect AI advice); anchoring on AI output increases under time pressure. Professional experience and self-efficacy predict lower AI dependence, while high confidence during AI-assisted tasks predicts increased reliance — a paradox that heightens diagnostic risk under cognitive load.
  • Related to: AI Impact on Human Cognition and Psychology

Medium Relevance (7)

• Learning Through Dialogue: Engagement and Efficacy Matter More Than Explanations

  • Empirical study — Analyzes 397 human-LLM conversations about socio-political issues. Mediation analysis shows LLM explanatory richness supports confidence via reflective insight, but knowledge gain operates entirely through the user's own cognitive engagement. High-efficacy users benefit more from extended interaction; better AI explanations alone do not predict learning. Conclusion: "learning from LLMs is an interactional achievement."
  • Related to: AI Impact on Human Cognition and Psychology

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

  • Empirical study — Interviews with 24 parents using LLM-generated interaction scenarios as probes. Identifies three gaps in current parental controls: parents are concerned about interactions existing tools miss; they want fine-grained, conversation-level transparency; and they need personalized controls adapted to their child's age and family values.
  • Related to: AI Impact on Human Communication and Relationships

• "I followed what felt right, not what I was told": Autonomy, Coaching, and Recognizing Bias Through AI-Mediated Dialogue

  • Empirical study — 160-participant experiment across four conditions testing AI-mediated dialogue for ableist microaggression recognition. Dialogue conditions outperform text-only reading; biased AI nudges are often consciously rejected while inclusive nudges are adopted as scaffolding. Contributes a validated vignette corpus and design implications for conversational bias-intervention systems.
  • Related to: AI Impact on Human Communication and Relationships

• Understanding User Perceptions of Human-centered AI-Enhanced Support Group Formation in Online Healthcare Communities

  • Empirical study — Mixed-methods survey (N=165) on algorithmically personalized peer-support matching in online health communities. Perceived value is high (mean 4.55/5) with 91.5% willingness to join; acceptance is conditional on security, transparency, human oversight, and user data control. Peer-matching importance strongly predicts perceived value (rho=0.764).
  • Related to: AI Impact on Human Communication and Relationships

• Gender Bias in Generative AI-assisted Recruitment Processes

  • Empirical study — Tests GPT-5 job recommendations for 24 simulated profiles balanced by gender and field. No significant differences in job titles or industry, but gendered linguistic patterns emerge: women described with emotional/empathetic adjectives, men with strategic/analytical ones — reproducing labour-market stereotypes in AI-mediated professional identity representation.
  • Related to: AI, Identity, and the Digital Self

• Human-Centred LLM Privacy Audits: Findings and Frictions

  • Empirical study — Two user studies (N=458 total) auditing what GPT-4o associates with everyday people's names. The model predicts 11 of 50 personal features with at least 60% accuracy. Participants want control over LLM associations even when outputs are not viewed as privacy violations. Identifies nine frictions blocking reliable human-centred LLM privacy audits.
  • Related to: AI, Identity, and the Digital Self

• An Intent of Collaboration: On Agencies between Designers and Emerging (Intelligent) Technologies

  • Case study — Three-month co-creation experiment between three designers and an LLM. Documents a power dynamic in which designers readily lose creative agency to the model; recovery requires introspection into one's creative process, structural understanding of the technology, and deliberate relationship calibration. Proposes three dimensions for maintaining human agency in human-AI creative work.
  • Related to: AI Impact on Human Cognition and Psychology

Low Relevance (2)

• To Believe or Not To Believe: Comparing Supporting Information Tools to Aid Human Judgments of AI Veracity

  • Empirical study — Compares full source text, passage retrieval, and LLM explanations as verification aids for AI-generated data extraction. LLM explanations enable fast assessments but foster inappropriate trust, reducing error detection especially for complex queries. Passage retrieval offers the best accuracy-speed tradeoff.

• From Control to Foresight: Simulation as a New Paradigm for Human-Agent Collaboration

  • Position paper — Argues current human-agent interaction forces users into cognitively demanding mental simulation of long-term agent consequences. Proposes "simulation-in-the-loop": users and agents explore projected trajectories before committing, transforming intervention from reactive guesswork into informed exploration.

---

Summary: Analyzed 367 total entries. Found 4 high-relevance, 7 medium-relevance, and 2 low-relevance papers related to our topics of interest.

Filter statistics: Passed Phase 1: ~27 entries (7.4%). Phase 2 exclusions (14 entries) were primarily pure engineering papers — agent frameworks, LLM benchmarks, multi-agent coordination systems — that shared topic keywords but addressed only technical capabilities. Today's feeds were dominated by agent performance and LLM optimization papers, all falling on the wrong side of the human-experience boundary. Closest near-miss: Increasing intelligence in AI agents can worsen collective outcomes (arxiv-cy/0aa6e6c36d2f), which studied societal consequences of AI agent populations but framed them as emergent systems dynamics rather than human experience.

[alexadamus77-ui]

This is a stellar breakdown! Filtering through 367 entries to find those high-relevance human-experience papers is a massive task. I noticed your note on excluding pure engineering papers in Phase 2, that's where the real 'Data Wall' usually is.
For researchers building these kinds of Daily Research Digests, ScholarAPI is a game-changer. Instead of manual scraping or navigating noisy feeds, it provides structured metadata and citation counts in clean JSON. It could essentially automate your 'Phase 1' filtering, letting you focus entirely on the high-level synthesis and 'Human-Experience' boundary analysis. I'd like to see this digest scale further with an automated data pipeline.

[jfomhover]

Thanks for the suggestion! It looks like that API has also monitoring support. Great.