Webnart AI Applied Lab: Research Library

This library brings together research, books, reports, and public projects that help explain how artificial intelligence affects people, institutions, work, culture, and decision-making.

Each entry names the work, gives a brief description, and links directly to the publication or project page. Author names are included for proper attribution.

AI safety, agents, and technical governance

DAIOS Research — compartmentalized harm in multi-agent systems

This preprint studies a safety problem in which a harmful objective is divided into ordinary-looking subtasks and distributed among several AI agents. It tests whether additional context, safety prompts, character training, and action restrictions help agents recognize the larger danger. The work shows why agent safety cannot depend only on each model judging its immediate prompt. The page identifies DAIOS Research as the publishing research group; Andrew Melnychuk shared it in the LinkedIn discussion.

Research page

Daniele Bailo — researcher-curated knowledge governance and RockGPT

This paper presents a model for trustworthy scientific AI in the geosciences. Researchers actively curate and govern the knowledge made available to the system, rather than treating retrieval as a purely technical pipeline. RockGPT is used as a self-hosted example connecting provenance, transparency, domain expertise, and institutional control.

Paper

Heidi Bennett — an agent adapting its model of AI rollout and policy

This project experiments with an AI agent that updates its world model as it reasons about AI deployment and policy. It is relevant to questions about changing assumptions, self-models, and governance-sensitive behaviour. The repository provides a way to inspect the experiment alongside its permanent research record.

Research record · Code and experiment

John R. Smith — Stewardship Intelligence

This work proposes “Stewardship Intelligence” as a way to frame AI safety around responsibility, balance, and the relationship between a system and its environment. It uses the Transactional Homeostasis Framework and is primarily conceptual and systems-oriented, making it a useful lead for readers looking beyond narrow model-performance definitions of safety.

Research record

Liberation Labs — AI welfare, interpretability, and agent ethics

Liberation Labs describes a research program connecting mechanistic interpretability with AI welfare and ethics. It asks what evidence might matter when discussing experience, self-modelling, or welfare in AI, while also exploring governance and licensing for autonomous systems. The project is exploratory, but it maps several emerging questions around consciousness indicators and moral consideration. Thomas Edrington appears in the project’s research and coalition material, but the linked pages represent a broader cooperative program rather than one individual paper.

Research constellation · AI welfare guide

4-Gov — applied governance

The 4-Gov project is an applied governance resource for AI and advanced technologies. It is relevant to how governance can be built into deployment when applied research does not pass through a journal or preprint server. Dennis Palatov shared the project; the linked site should be consulted for its formal authorship and methods.

4-Gov

Human rights, bias, and representational harm

Lindsey Andersen — human rights across the generative AI value chain

This BSR report maps human-rights risks across the generative AI value chain, from data suppliers and foundation-model developers to downstream developers and deployers. It is designed for responsible-AI practitioners and connects impact assessment to international human-rights frameworks. Its practical contribution is to show how harm can travel through a supply chain rather than sit only in the final interface.

BSR report

Australian Human Rights Commission, Actuaries Institute, and UTS Human Technology Institute — discrimination in insurance and AI risk in finance

The first resource examines AI-driven discrimination in insurance pricing and underwriting through a human-rights lens. The second, from the Actuaries Institute and UTS Human Technology Institute, provides an operational framework for AI risk management in financial services, covering accountability, classification, quantification, controls, and vendor due diligence. Together they connect ethical AI to credit, insurance, governance, and customer outcomes. Chris Dolman is acknowledged as a contributor to the financial-services framework; he is not presented here as the sole author.

Human-rights guidance · Financial-services framework

Siobhan Mackenzie Hall — cultural bias in image generation

The World Wide Recipe paper develops a community-centred method for collecting culturally specific data and evaluating generative image systems. Using food as a lens, it shows how text-to-image models can produce inaccurate, flattened, or insensitive representations, especially for regions poorly represented in web-scraped data. The accompanying dataset supports research on representational erasure rather than treating bias as one abstract score.

Paper

Tiago Torrent — multilingual stereotypes in language models

SHADES is a multilingual dataset for evaluating stereotypes in language models across 16 languages and 20 regions. It pairs culturally grounded stereotypes with minimally contrasting statements and tests base and instruction-tuned models. It addresses a major weakness in bias evaluation: the tendency to make English-language findings stand in for multilingual and culturally specific behaviour.

SHADES

Sidney Wong — the social benefit of hate-speech detection

This systematic review examines 48 hate-speech detection systems across 37 publications and asks what social benefit they actually produce. Wong argues that technical research has had limited uptake by policymakers and civil-society organizations partly because ethical and social-impact frameworks are missing. The paper therefore connects model development to real users, institutions, and public-interest outcomes.

Review

Democracy, law, and public accountability

Giulia Sandri and Claudio Novelli — measuring AI dangers to democracy

This preprint develops a framework for identifying and comparing AI risks to information ecosystems, elections, and public administration. It treats AI governance as a delegation problem: institutions and citizens may delegate important functions to systems or vendors they cannot properly monitor. The authors combine principal-agent theory with the NIST trustworthy-AI characteristics and propose institutional assessability as a condition for democratic control.

Preprint

Martin Schmalzried — AGI, embodiment, consciousness, and the metaverse

This philosophical paper explores what embodied artificial general intelligence might mean and how it could relate to human consciousness and perception. It combines embodied cognition, theories of the self, interface theories of perception, the metaverse, decentralized systems, and open-source AI. It is speculative rather than an empirical safety evaluation, but it broadens the discussion toward ontology and embodiment.

Paper

Gry Hasselbalch — data ethics, power, and human agency

Data Ethics of Power examines how data systems redistribute power through policy, business, and everyday life. Human Power asks what a human-centred politics should protect in an AI-shaped society, focusing on creativity, intuition, emotion, life, defiance, love, and wisdom. Read together, the books connect institutional power with the human capacities that technology policy should preserve.

Human Power · Data Ethics of Power · Earlier essay

Education, coaching, and human development

Devika Toprani — Somagraphic Learning and pre-AI sense-making

Somagraphic Learning is a human-first framework built around “Attempt, Map, Refine.” The learner develops and maps their own reasoning before AI enters at the refinement stage. The project addresses how education can use AI without allowing the tool to replace initial sense-making, bodily awareness, or responsibility for the work.

Preprint

Caitlin E. McDonald — coaching’s core competencies in the age of generative AI

This article examines where generative AI might support coaching and where it could undermine the profession’s human aims. It considers reflective feedback, goal tracking, accessibility, privacy, bias, automation bias, poor ethical judgment, and sycophancy. The International Coaching Federation’s competencies provide a practical framework for deciding when AI supports learning and when it displaces it.

Article

Dr Rosalie Clarke — AI in the PhD job search

This article considers AI in career development for PhD students and postgraduate professionals. It places strategic use alongside authenticity, ethical judgment, and recruitment realities, and introduces a planned guide for researchers, universities, recruiters, and employers. It is an accessible practitioner resource rather than a conventional research paper.

Article · Minerva Recruitment

Daniel Leto — neural networks and epidemic studies

Leto’s linked work concerns the use and consequences of neural networks in epidemic research. The topic sits at the intersection of machine learning, public-health modelling, and scientific reliability: models may identify patterns or forecast spread, but their assumptions and errors can affect decisions during health crises. The two links should be read together to confirm the exact study titles and results.

ResearchSquare · Springer chapter

Workplace, scientific, and cultural applications

Julian Friedland — AI tools for workplace sustainability

This Organizational Dynamics article asks how AI can help organizations move beyond short-lived behavioural nudges in sustainability programs. It proposes combining nudges with “boosts” that strengthen people’s own understanding and decision-making, then offers a framework for choosing between them in AI-enabled workplaces. Its ethical interest lies in the difference between building capacity and quietly engineering behaviour.

Article

David C. Flynn — Literary Narrative as Moral Probe

This paper by David C. Flynn uses literary narrative, including unresolved moral scenarios, to test whether AI systems demonstrate ethical reasoning or merely produce correct-sounding responses. It reports a cross-system study and identifies failure modes such as self-misattribution and reliance on surface cues. The work is a useful contribution for readers interested in narrative methods and evaluating moral reasoning in AI.

Work

Sune Selsbæk-Reitz — meaning and the fluency trap

Promptism: Fluent Machines, Forgotten Questions, and the Fight for Meaning in the Age of AI examines what happens when systems become extremely good at producing plausible language. Its central concern is that fluency can be mistaken for understanding, knowledge, or wisdom, encouraging people to outsource questions that should remain human and interpretive.

Book

Angelo Adrian — artificial stupidity and AI overconfidence

This essay uses “artificial stupidity” to draw attention to the gap between impressive outputs and genuine understanding or judgment. It questions the tendency to treat AI as intelligent simply because it is fluent, fast, or financially valuable. It is best read as a critical public essay rather than as a peer-reviewed paper.

Essay

Michael A. Covington — an introduction to computer ethics

Covington’s short book is designed for computer users and managers who need an accessible introduction to computer ethics. It provides a starting point for thinking about privacy, responsibility, professional conduct, and human consequences before moving into more technical or academic literature.

Book

Deliberate Ensemble — a public archive of AI research and systems work

Deliberate Ensemble’s archive collects research papers, architecture documents, agent designs, governance specifications, and related publications. Its themes include multi-agent coordination, persistent identity and memory, verification, constitutional constraints, and the translation of research into working systems. Because the page is an index rather than one paper, it is best used as a map of a larger body of work. Sean David Ramsingh shared the archive; individual documents should be credited to their listed authors.

Research archive

Tiia Sõmer — earlier research record

The shared link is a Google Scholar citation record for earlier work by Tiia Sõmer. The LinkedIn comment says that the research is several years old and may be taken forward, but the captured material does not preserve enough bibliographic information to summarize its subject responsibly. This entry should remain a verification task until the paper title and research area are confirmed.

Scholar record

Internet Rules Lab — student avoidance of generative AI

The Internet Rules Lab study investigates students who choose not to use generative AI in academic or personal life. It asks how avoidance works in practice, how widespread it is, and what motivates it. The study is recruiting CU Boulder students for interviews, making it a useful qualitative complement to research that focuses only on adoption, productivity, or model performance. The page identifies Mirakle Wright as the lead student researcher, with Casey Fiesler, Jed Brubaker, and Jonathan Zong as faculty; Urmila Venkatesh recommended the study.

Study page