Research

Concepts I use to think about how AI reshapes institutions, judgment, and accountability.

“Governance that doesn’t ship isn’t governance.”

My research sits at the intersection of AI governance, institutional theory, and production systems. I study what happens when organisations actually deploy AI — not in theory, but in practice. The work combines philosophical frameworks (McMurtry’s Life-Value Onto-Axiology), empirical data (company-task mappings, worker preferences and institutional adoption patterns), and what actually happens when these systems run (BORG, Arctic Tracker, Gjöll).

The methodology — Diagnostic Sociology — applies emergency medicine pattern recognition to complex social systems. Sixteen years as a paramedic taught me to read vital signs, identify cascading failures, and intervene before systems collapse. I apply the same diagnostic discipline to AI’s societal effects.

The concepts I work with

Eight terms I use and argue for. The entries distinguish my proposals and applications from the earlier work they draw on. I did not coin Cognitive Debt — it comes out of software engineering and technology writing, and what I add is the extension to AI governance. I put all eight forward to be argued with, not as settled vocabulary.

8 conceptsMA thesis + use in production2015–2026

Diagnostic Concepts

Responsibility Fog

How accountability dissolves when decisions are distributed across algorithms, committees, and platforms.

When a hiring algorithm rejects a candidate, who is responsible? The developer? The deploying organisation? The vendor? The regulator who approved the framework? Responsibility Fog describes the systematic diffusion of accountability in AI-mediated decision-making — not as a bug, but as an emergent property of how institutions adopt algorithmic systems. I read my 2015 BA thesis on telecommunications data retention retrospectively as an earlier encounter with unclear responsibility in surveillance systems. My contribution is this framing of institutional AI adoption, situated alongside Dennis Thompson’s problem of many hands (1980), Mark Bovens’s work on accountability, and Andreas Matthias’s responsibility gap (2004).

In this work: Beyond Fragmentation (MA thesis, 2026)

Under contract: The Irreducible Human (under contract with De Gruyter Brill, with Dr. Kristian Guttesen)

Cognitive Debt

How judgment atrophies when we outsource thinking to machines.

The term Cognitive Debt originates in software engineering and broader technology discourse. In Beyond Fragmentation, the concept is expanded and formalised to describe the gradual erosion of human evaluative capacity through sustained algorithmic dependence. It captures something specific: not that AI makes us lazy, but that repeated delegation of judgment creates a compounding deficit — one that becomes visible only when the system fails and no one remembers how to think without it.

In this work: Term predates this work; extended to AI governance in Beyond Fragmentation (MA thesis, 2026)

Applied: Teaching curriculum design at UNAK; GRÓ Fisheries Training Programme; forthcoming nursing course on AI and women’s health

The Benevolent Cage

When protective governance becomes the constraint it was designed to prevent.

My use of the Benevolent Cage names a governance paradox: AI governance frameworks designed to protect human autonomy can themselves become mechanisms of control — setting boundaries so rigid that they prevent the adaptive capacity they claim to preserve. The concept interrogates the paradox of institutional AI policy: at what point does protection become containment?

In this work: Beyond Fragmentation (MA thesis, 2026)

Cognitive Complacency

What happens when comfort meets capability and produces nothing.

I use Cognitive Complacency as a diagnostic label. Where Cognitive Debt describes judgment erosion through AI dependence, Cognitive Complacency describes cases where well-resourced institutions reduce AI to convenience layers. In Rwanda, PATH launched a trial without clinical intervention in July 2025 to evaluate generative AI support for community health workers. Roasting marshmallows is my metaphor for settling for convenience.

In this work: Portfolio notes (February 2026)

The Augmentation Gap

When capability compounds, inequality stops being about income and becomes about agency.

If a small, augmented cohort can produce disproportionate output, inequality is no longer only a matter of wealth. It becomes unequal access to the ability to make things happen. I use the augmentation gap to name that divide. A few hyper-productive actors steer the economy while the majority matter less and less to what McMurtry calls the Money Sequence. The gap describes the danger better than the general fear of AI taking jobs, because it is about a concentration of power, not a loss of work.

In this work: In the Shadow of the Machine: The Order of Magnitude Year (Atti Convegno 2025, Quaderni di Immaginazione Attiva no. 1, Palermo: Carlo Saladino Editore, 2026, pp. 181–205)

Cited in: Arif Jameel, “We the Living Generation”: A Human Constitution for Artificial Intelligence — The Ceteris Paribus Law (2026), which says it gives the Law’s article on concentration of AI power “a concrete, named mechanism”

Frameworks & Methods

The Investment-Sentiment Gap

Shao et al. (2025): 41% of Y Combinator company-task mappings fall in zones of low worker desire for automation.

The Investment-Sentiment Gap is my interpretation of findings from the Stanford SALT Lab WORKBank study (Shao et al., 2025, arXiv:2506.06576). Their 41% concerns Y Combinator company-task mappings in the Low Priority and Automation Red Light zones, where worker desire for automation is low. Their separate 1.26% measures the share of Claude.ai usage associated with the ten occupations with the highest desire for automation. These are different denominators and do not establish a dollar investment ratio.

In this work: Beyond Fragmentation (MA thesis, 2026)

Methodology: Secondary analysis of Shao et al.’s WORKBank data, Y Combinator company-task mappings and Claude.ai usage

The VALOR Framework

A governance model grounded in life-value rather than market logic.

VALOR is my application of McMurtry’s framework to AI governance. It provides a structured alternative to existing AI governance models (EU AI Act risk tiers, NIST RMF) by grounding assessment in McMurtry’s Life-Value Onto-Axiology — asking not “what is the risk level?” but “does this deployment enable or disable the means of life?” The framework has been applied to three production systems: two at the University of Akureyri, and Gjöll, which is my own.

In this work: Beyond Fragmentation (MA thesis, 2026)

Used in: BORG (AI governance platform), Arctic Tracker (conservation analytics), Gjöll (fire safety database)

Diagnostic Sociology

Emergency medicine pattern recognition applied to AI’s societal effects.

A proposed methodology, not an established field. It formalises the transfer of diagnostic thinking from clinical practice to social systems analysis. Just as a paramedic reads vital signs, identifies cascading organ failure, and triages intervention — Diagnostic Sociology reads institutional adoption patterns, identifies systemic governance failures, and proposes targeted interventions before collapse becomes irreversible. I put it forward as a working method and expect it to be tested rather than taken on trust.

In this work: Applied throughout Beyond Fragmentation, institutional advisory work, teaching methodology

Thesis

Beyond Fragmentation: A Life-Value Alternative for AI Governance

MA in Social Sciences · University of Akureyri · 2026

90 ECTS

Supervisor: Professor Giorgio Baruchello

An empirically grounded study analyzing systemic failures in AI governance. Introduces Responsibility Fog, the Benevolent Cage, and the VALOR Framework, and extends Cognitive Debt — a term from software-engineering and technology discourse — to AI governance. Develops the Investment-Sentiment Gap interpretation through secondary analysis of Shao et al.’s WORKBank study. Worked through three production systems and a Neo4j knowledge graph mapping the relationships between governance actors, frameworks, and failure modes.

QuartoNeo4jPythonXeLaTeXCypherGitHub Actions

Explore the thesis as an interactive concept graph

Book under Contract

The Irreducible Human: Life, Value, and Meaning in the Age of AI

De Gruyter Brill (under contract) · Co-authored with Dr. Kristian Guttesen

A philosophical framework for AI governance grounded in what makes life worth living. Builds on the theoretical foundations of Beyond Fragmentation to propose a comprehensive alternative to market-driven AI governance — one rooted in the conditions that enable human flourishing rather than the conditions that enable capital accumulation.

The Systems Behind the Concepts

This research is not purely theoretical. Each framework has been tested against production systems serving real users.

BORG

University of Akureyri’s hybrid AI governance platform. Multi-provider (Claude, GPT, Gemini), institutional knowledge retention, policy enforcement. The system that operationalises the governance theory.

Arctic Tracker

Conservation analytics platform integrating 492,095 (thearctic.is, 4 November 2026) CITES trade records, IUCN assessments, and illegal seizure data for 43 Arctic species. The preprint “Trading for Conservation” was posted on Research Square on 17 December 2025, by Tom Barry and Magnus Smári Smárasson (names as listed). The Tracker withholds individual dataset authorship until publication.

Gjöll

Open-access fire safety database documenting identified fatal fire incidents in Iceland in 1968–2025 (113 incidents, 145 lives). Deposited at GAGNÍS with DOI 10.34881/I5WGJU. Key finding: the fire death rate per person-year is several-fold lower in buildings from 1999 onward. This describes construction-period cohorts, not a proven regulatory effect; the decline began before 1999.

Sumarhús

My own workshop: the machines, standing agents and tooling I built to do my work on infrastructure I control. Not a product. It is where the empirical side of the governance work actually happens.

Speaking & Presentations

How Icelanders are thinking about AI (Anthropic, 2026)

On camera in Anthropic’s film about Iceland’s national AI education pilot, credited on screen as “Magnús, University of Akureyri”. Filmed in Akureyri in May 2026.

The Oxford Artificial Intelligence Summit 2025: Autonomous AI Agents

Course Tutor, University of Oxford, 27-28 June 2025. The workshop materials I wrote for it are public.

Atlas of the Future Conference

Palazzo Adriano, Sicily (August 2025) — with Prof. Giorgio Baruchello and Prof. Rachael Lorna Johnstone

Research Arc

The intellectual thread is continuous.

2012–2015

BA thesis examines the legality of mandatory telecommunications data retention → a precursor to Responsibility Fog in my retrospective reading

2007–2022

Sixteen years of emergency services develops pattern recognition under pressure → foundation for Diagnostic Sociology

2022–present

Building and operating AI systems in production → empirical grounding for governance theory

2024–2026

MA thesis formalises Responsibility Fog, Cognitive Debt, Investment-Sentiment Gap, VALOR Framework → carried into production systems

2026

The Irreducible Human (under contract with De Gruyter Brill) extends the theoretical framework to book-length treatment

Interested in this work?

I’m a researcher and builder focused on how AI reshapes institutions, judgment, and accountability. If your work intersects with AI governance, societal impacts, or responsible deployment — I’d be glad to hear from you.

magnus@smarason.is LinkedIn