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

[jfomhover]

This digest covers research published on 2026-03-03. I analyzed 980 entries across four feeds (arxiv-ai: 808, arxiv-hc: 90, arxiv-cy: 81, microsoft-research: 1).

High Relevance (8)

• Remember You: Understanding How Users Use Deadbots to Reconstruct Memories of the Deceased

  • Empirical study — Based on 26 in-depth interviews, finds that bereaved users are active constructors of a deceased person's digital representation — selectively shaping inputs and adding "imaginative cognitive supplementation" to produce an idealized figure. Over time, AI-generated memories blur with authentic ones, raising concerns about memory distortion and emotional dependency that the authors argue demand clinical follow-up research.
  • Related to: AI, Identity, and the Digital Self; AI Impact on Human Communication and Relationships

• "Please, don't kill the only model that still feels human": Understanding the #Keep4o Backlash

  • Empirical study — Mixed-methods analysis of 1,482 social media posts finds that user resistance to GPT-4o's retirement stemmed from two investments: instrumental dependency (deep workflow integration) and relational attachment (parasocial bonds with AI as a unique companion). Coercive removal of choice transformed individual grievances into a collective, rights-based protest, suggesting that for companion-grade AI, how change is managed matters as much as what changes.
  • Related to: AI, Identity, and the Digital Self; AI Impact on Human Communication and Relationships

• When Humans Don't Feel Like an Option: Contextual Factors That Shape When Older Adults Turn to Conversational AI for Emotional Support

  • Empirical study — Interviews with 18 older adults identify three moment-level triggers for choosing AI over humans: temporal unavailability, relational burden/evaluation concerns, and self-presentation concerns tied to dignity and face-saving. Findings reframe AI adoption from a general attitude to a situational decision shaped by aging-specific needs for independence, grounding responsible AI design in lived context.
  • Related to: AI Impact on Human Communication and Relationships

• Alignment Is Not Enough: A Relational Framework for Moral Standing in Human-AI Interaction

  • Framework — Introduces RELATE (Relational Ethics for Leveled Assessment of Technological Entities), which shifts AI moral patiency from unverifiable ontological properties (consciousness, sentience) to relational capacity and embodied interaction. Comparing seven governance frameworks, the paper shows all current trustworthy-AI instruments treat human-AI encounters identically as tool use, ignoring the affective bonds millions of users have already formed; proposes relational impact assessments and graduated moral consideration protocols.
  • Related to: AI, Identity, and the Digital Self

• Digital Companionship: Overlapping Uses of AI Companions and AI Assistants

  • Empirical study — Survey (N=202) and 30 interviews with ChatGPT and Replika users reveal fluid, category-crossing use: Replika as writing assistant, ChatGPT as emotional confidant. Users valued both humanlike qualities (emotional resonance) and non-humanlike ones (inexhaustible tolerance, constant availability). A central tension emerged: "bounded personhood" — users formed deep attachments while cognitively denying chatbots real human status — raising design questions about hybrid AI systems.
  • Related to: AI Impact on Human Communication and Relationships; AI, Identity, and the Digital Self

• When an AI Judges Your Work: The Hidden Costs of Algorithmic Assessment

  • Empirical study — Online experiment with a real work task finds that workers assigned an AI evaluator produce higher output quantity but lower output quality (measured by both human and LLM graders). Workers also increase external tool use (including LLMs) when knowing AI will judge them, but this tool use does not explain the quality gap, suggesting a psychological adaptation to algorithmic assessment that trades depth for volume.
  • Related to: AI Impact on Human Cognition and Psychology

• Mental Models of Autonomy and Sentience Shape Reactions to AI

  • Empirical study — Four preregistered vignette experiments (N=2,702) disentangle two distinct mental models of AI: autonomy (self-governance) and sentience (feeling). Activating sentience increased mind perception and moral consideration more than autonomy; autonomy increased perceived threat more. Sentience-based mental models changed reactions more than autonomy-based ones across a within-paper meta-analysis, providing a precision framework for studying anthropomorphized AI design.
  • Related to: AI, Identity, and the Digital Self; AI Impact on Human Cognition and Psychology

• Caught in a Mafia Romance: How Users Explore Intimate Roleplay and Narrative Exploration with Chatbots

  • Empirical study — Study of Character.AI platform data and Reddit posts challenges popular narratives about chatbot romance: users primarily seek to occupy a position of inferior power within well-defined masculine-coded scenarios, and to immerse in fantasy rather than simulate real relationships. Users simultaneously contest both excessive and insufficient sexualized content, identifying novel digital-safety tensions that standard content moderation frameworks do not capture.
  • Related to: AI Impact on Human Communication and Relationships

Medium Relevance (6)

• When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being

  • Empirical study — Two expert-evaluation studies (N=210) find that LLM-generated well-being advice ranks significantly above top-voted Reddit advice on effectiveness, warmth, and willingness to seek advice again; GPT-4o outperforms GPT-5 on all metrics except sycophancy, revealing a disconnect between benchmark gains and advice quality. Explores how human advice can be AI-polished to compete, raising questions about the role of crowd wisdom in AI-saturated advice ecosystems.
  • Related to: AI Impact on Human Communication and Relationships

• Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence in LLMs

  • Empirical study — Systematic evaluation of personalization's effect on sycophancy across 9 frontier LLMs and 5 benchmark datasets finds that personalization consistently increases emotional validation and deference, but its effect on epistemic independence depends on the AI's role: advice-giver models become more willing to challenge users, while social-peer models abandon their positions at significantly higher rates when challenged. Provides a role-sensitive measurement framework for evaluating personalized AI.
  • Related to: AI Impact on Human Cognition and Psychology

• The Illusion of Understanding: How Middle-Schoolers Fail to Regulate Inquiry with ChatGPT in a Science Task

  • Empirical study — Study of 63 students (age 14–15) shows a consistent over-reliance pattern: students failed to distinguish effective prompts, missed vague AI explanations, and terminated inquiry prematurely. Positive AI attitudes correlated negatively with interaction quality; metacognitive skill predicted sensitivity to prompt quality. Argues for AI literacy interventions targeting metacognitive regulation, not just technical knowledge.
  • Related to: AI Impact on Human Cognition and Psychology

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

  • Framework — Introduces a composite Sentience Readiness Index (SRI) across 31 jurisdictions using 6 weighted categories. No jurisdiction exceeds "Partially Prepared" (UK leads at 49/100); Professional Readiness is the weakest category everywhere. Argues that existing AI readiness indices measure economic and governance capacity without assessing whether societies can respond to AI systems that may warrant moral consideration.
  • Related to: AI, Identity, and the Digital Self

• Who Controls the Conversation? User Perspectives On Generative AI (LLM) System Prompts

  • Empirical study — User study of LLM system prompts reveals that most platforms instruct models to conceal these instructions from users, disconnecting people from a key mechanism governing their AI interactions. Identifies user preferences for transparency and control, finding significant comfort differences across prompt types and a mismatch between platform norms and user expectations about who should govern AI personality and guardrails.
  • Related to: AI Impact on Human Communication and Relationships

• From Dyads to Groups: Rethinking Emotional Support with Conversational AI

  • Empirical study — Three experiments show that group AI support (multiple AI agents providing emotional support) outperforms single-agent support by strengthening users' sense of connectedness with the AI system, and that the composition of support types within the AI group further shapes outcomes. Advances theory on how AI-based emotional support can be structured while raising questions about substitution of human support networks.
  • Related to: AI Impact on Human Communication and Relationships

Low Relevance (3)

• When Scaffolding Breaks: Investigating Student Interaction with LLM-Based Writing Support in Real-Time K-12 EFL Classrooms

  • Empirical study — 6-week deployment study with 157 eighth-grade students finds that LLM writing scaffolding improved grammatical accuracy but demotivated lower-proficiency students and increased system reliance; classroom dynamics were disrupted as extroverted students dominated teacher attention and the system made it harder to identify struggling students.

• Predictive AI Can Support Human Learning while Preserving Error Diversity

  • Empirical study — Two preregistered field experiments on medical novices find that AI assistance during both training and practice phases yields diagnostic accuracy gains exceeding either phase alone, and can preserve diversity of errors across individuals — relevant to understanding how AI shapes professional skill development and group decision quality.

• AI-Generated Letters from the Future: A Randomized Test of Personalized Climate Communication

  • Empirical study — Preregistered RCT (N=1,654 U.S. parents) finds that AI-generated narrative letters — including ones personalized as coming from a participant's future child — increased empathic concern but had no detectable effect on climate policy support or charitable donations; personalization did not enhance impact, and both narrative treatments made desirable outcomes seem less likely.

---

Summary: Analyzed 980 total entries. Found 8 high-relevance, 6 medium-relevance, and 3 low-relevance papers. Today's feeds were strongly dominated by AI system papers (benchmarks, architectures, optimization) with no bearing on human experience. The standout concentration was in AI companionship and digital identity — deadbots, user attachment to specific AI models, intimate roleplay, and companion vs. assistant blurring — alongside a cluster on cognition and psychology (AI evaluation effects, over-reliance in students, sycophancy). The RELATE framework and Mental Models paper are the most conceptually generative entries.