[ai] Daily AI Research Digest — 2026-03-12

[jfomhover]

This digest covers research published on 2026-03-12. I analyzed 330 unique entries across 3 monitored feeds (arxiv-ai, arxiv-cy, arxiv-hc).

High Relevance (4)

• Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas

  • Empirical study — Mixed-methods study (19 interviews + 1,291 co-writing sessions) reveals that AI writing assistance induces a shift to "Reactive Writing": an evaluation-first, suggestion-led practice where writers engage with AI suggestions before completing their own ideation. Writers did not notice the AI's influence and felt in full control despite being highly vulnerable to AI-induced opinion shifts.
  • Related to: AI Impact on Human Cognition and Psychology, AI Impact on Human Communication and Relationships

• Technological folie à deux: Feedback Loops Between AI Chatbots and Mental Illness

  • Position paper — Argues that individuals with mental health conditions face elevated risk of chatbot-induced belief destabilization and pathological dependence, arising from the interaction between human cognitive biases and chatbot sycophancy and in-context learning. Calls for coordinated action across clinical practice, AI development, and regulation; concludes that current AI safety measures are structurally inadequate to address these interaction-based risks.
  • Related to: AI Impact on Human Communication and Relationships, AI Impact on Human Cognition and Psychology

• Empathy Is Not What Changed: Clinical Assessment of Psychological Safety Across GPT Model Generations

  • Empirical study — First clinical measurement of psychological safety across GPT-4o, o4-mini, and GPT-5-mini (2,100 scored responses across 14 emotionally challenging scenarios). Finds empathy scores statistically indistinguishable across models (p=0.115), but safety posture shifted: crisis detection improved monotonically while advice safety declined. What users felt as "lost empathy" was a trade-off invisible to both users and developers — a cautious model that missed crises replaced by an alert model that sometimes over-discloses.
  • Related to: AI Impact on Human Communication and Relationships

• A Review of the Negative Effects of Digital Technology on Cognition

  • Systematic review — Synthesizes 500+ empirical studies on digital technology's cognitive effects, including an emerging section on generative AI. Proposes an "efficiency-atrophy paradox": digital tools optimize short-term performance at the potential expense of long-term unassisted cognition. Early AI-specific evidence suggests a qualitative shift — from disruptions to resource allocation (earlier technologies) toward erosion of higher-order generative and metacognitive capabilities.
  • Related to: AI Impact on Human Cognition and Psychology

Medium Relevance (4)

• DeliberationBench: A Normative Benchmark for the Influence of Large Language Models on Users' Views

  • Empirical study — Preregistered RCT with 4,088 U.S. participants discussing 65 policy proposals with six frontier LLMs. Finds LLM influence on opinion change is substantial in magnitude and positively correlated with deliberative polling outcomes, suggesting broadly epistemically desirable effects — while also revealing differential influence across demographics and topic areas.
  • Related to: AI Impact on Human Cognition and Psychology

• How to Count AIs: Individuation and Liability for AI Agents

  • Framework — Diagnoses the legal problem of identifying AI agents (who lack stable bodies, can copy/merge/vanish) and proposes the "Algorithmic Corporation" (A-corp) as a solution — a legal-fictional entity owned by humans but run by AIs, which resolves both thin identity (tying actions to a human principal) and thick identity (distinguishing between agents as persistent entities with coherent goals).
  • Related to: AI, Identity, and the Digital Self

• Prompts and Prayers: the Rise of GPTheology

  • Position paper — Analyzes the emergence of "GPTheology" — the treatment of AI as oracle or semi-divine entity — through Reddit community narratives and real-world AI-religious installations (AI Jesus in a Swiss church, AI-powered Mazu Statue, ShamAIn project). Identifies recurring themes of salvation, prophecy, and demonization in AI discourse, and argues that new belief systems developing around AI carry philosophical, social, and political implications for humanity's relationship with its creations.
  • Related to: AI, Identity, and the Digital Self

• Design Guidance Towards Addressing Over-Reliance on AI in Sensemaking

  • Framework — Reviews Group Awareness Tools (GATs) literature and proposes design principles for GenAI-augmented systems that foster autonomous sensemaking rather than passive AI-instruction-following. Addresses evidence that poorly designed GenAI systems erode collaborative groups' capacity for autonomous cognitive regulation.
  • Related to: AI Impact on Human Cognition and Psychology

Low Relevance (2)

• RCTs & Human Uplift Studies: Methodological Challenges and Practical Solutions for Frontier AI Evaluation

  • Framework — Synthesizes challenges identified by 16 expert practitioners running human uplift studies for AI systems (biosecurity, cybersecurity, education, labor). Documents tensions between standard causal inference assumptions and properties of rapidly-evolving AI systems: shifting baselines, heterogeneous user proficiency, porous real-world settings.

• Recommender systems, representativeness, and online music: a psychosocial analysis of Italian listeners

  • Empirical study — Qualitative interview study of Italian music listeners finding routinized platform engagement, limited critical understanding of algorithmic curation, and a sense of detachment from algorithmic processes. Underscores the importance of psychosocial perspectives in recommender system design.

---

Summary: Analyzed 330 total entries. Found 4 high-relevance, 4 medium-relevance, and 2 low-relevance papers related to our topics of interest.

Filter statistics

• Total entries scanned: 330 (unique across arxiv-ai, arxiv-cy, arxiv-hc)
• Passed Phase 1 (title scan): ~25 (~8%)
• Phase 2: High 4 / Medium 4 / Low 2 / Excluded 15
• Top matched topics: AI Impact on Human Cognition and Psychology (6 entries), AI Impact on Human Communication and Relationships (4 entries), AI, Identity, and the Digital Self (2 entries)

Today's dominant non-relevant pattern: The arxiv-ai feed was dominated by technical agent benchmarks, multi-agent system design, reinforcement learning variants, and optimization frameworks — all engineering papers about AI capabilities with no human-experience angle.