Emerging Horizon of Research and Its Advancement
A developing index of research arenas, prioritising safety and responsible advancement.
PRIMARY GLOBAL AI GOVERNANCE RESEARCH
AI Identifiability First; Compliance Second
The first questions across industries and jurisdictions are whether AI systems and their deployments can be identified, and whether their uses comply with applicable duties. This framework distinguishes technical detection, mandatory disclosure, material-based identification hypotheses, inventories and accountable oversight.
Explore the global AI identification and compliance framework →
GLOBAL GOVERNANCE · INFORMATION ARCHITECTURE
AI Identification, Discovery & Inventory Mapping
Research the problem of AI systems missing from the map: embedded, outsourced, transient, autonomous and undocumented uses; formal identity relationships; inventory completeness; and testing of compliance-discovery coverage.
FRONTIER RESEARCH · HISTORICAL PROVENANCE
Research Novelty and Field Formation
Which research frameworks and specialisms were newly articulated in 2026, and how do they relate to earlier disciplines? A dated register credits originators, tracks novelty categories, compares predecessors and distinguishes field formation from new terminology.
FOUNDATIONAL RESEARCH PRINCIPLE
The Open Frontier of Knowledge and Innovation
Research and engineering developments may exist beyond the knowledge represented in an AI system's training data, accessible sources or current retrieval capabilities. This includes unpublished, privately developed, newly emerging or otherwise unrepresented work.
An AI system's inability to identify an advancement does not establish that it does not exist, or that a known approach represents the highest achievable level of scientific or technological development.
Frontier inquiry therefore requires epistemic openness: recognising the limits of available evidence, considering alternative architectures and independent discoveries, examining unfamiliar claims critically, and updating conclusions when new evidence becomes available.
Important distinction: assume that knowledge may be incomplete; do not assume that any particular unverified breakthrough exists. Openness to discovery should coexist with rigorous evaluation, uncertainty reporting and appropriate safety controls.
SOCIAL EPISTEMOLOGY · RESEARCH EVIDENCE
Why AI Access Is Not the Limit of Human Knowledge
Evidence from archives, oral traditions, unpublished research, tacit expertise, digital inequality and AI recall shows why a training corpus or search system cannot delimit human knowledge. This research explainer includes 17 direct links to authoritative sources.
EXPLORE ACTIVE RESEARCH
AI Research Discovery Directory
Explore 27 curated destinations across emerging discoveries, frontier engineering, independent capability evaluation, and AI safety and governance research. Browse original papers, engineering issues, benchmark methods and research programmes.
RESEARCH QUESTIONS · EVIDENCE · SOLUTIONS
Research Questions, Confirmed Gaps & Solutions
Compare existing contributions, competing explanations, demonstrated limitations and proposed responses. Each case includes source credit, safety implications, validation status and a continuation path for other researchers.
MY RESEARCH AVENUES · MA FOUNDATIONS · 5 NOVEMBER 2026
Urgent research gaps: developing my MA foundations
These ten avenues form part of my independent research, developing from my 2024–25 study of AI, Ethics and Society at Birkbeck, University of London. They connect technical mechanisms with philosophical questions, social research, institutional practice and the consequences of deployment. They sit within the wider arena index below; they are not ten new, separate project pages.
The programme specification combines technical knowledge with philosophical and ethical inquiry, case analysis, policy development, communication and independent research. The History and Foundations of AI syllabus supplies the more specific connections identified here. Specialist applications extend those foundations; they are not all named subjects in the taught syllabus.
Real-world AI reliability and evaluation
Research gap: Establish when test performance transfers to unfamiliar inputs, incomplete evidence and changing operational conditions. Investigate calibration, abstention, reproducibility and the difference between correct output and justified conclusions.
MA foundation: Meeting 4, probabilities and prediction; Meeting 6, feedback loops and model collapse; Meeting 9, explainability and accountability. Existing laboratory and experiment records →
Source code to explore: UKGovernmentBEIS/inspect_ai. Evaluation framework; use explicit case/scorer definitions rather than treating benchmark success as proof of internal reasoning. Licence / reuse terms: MIT.
Autonomous agents, cybersecurity and human control
Research gap: Test permissions, resistance to malicious instructions, recovery after errors and the effectiveness of human intervention when agents act through tools.
MA foundation: Meeting 3, artificial advisers and human–AI interaction; Meeting 9, design and accountability. Cybersecurity and agent-control engineering are further technical extensions. Existing agent architecture study →
Source code to explore: modelcontextprotocol/python-sdk. Tool-protocol SDK; a component for permission experiments, not a complete secure agent. Licence / reuse terms: MIT.
Evidence integrity, provenance and dependent revision
Research gap: Distinguish independent corroboration from repeated reporting; preserve conditions and source ancestry; update affected conclusions when a premise changes while retaining historical records.
MA foundation: Meetings 1–2, intelligence, meaning and rules; Meeting 5, data; Meeting 6, recursion; Meeting 9, accountability. Provenance systems and revision benchmarks develop these connections. Existing evidence architecture →
Source code to explore: trungdong/prov. W3C PROV implementation; lineage representation does not itself establish truth or propagate every model inference. Licence / reuse terms: MIT.
Institutional accountability and effective oversight
Research gap: Determine which review, challenge, escalation and correction arrangements actually protect people affected by automated decisions, and where responsibility becomes dispersed.
MA foundation: Meeting 7, AI, power and the state; Meeting 8, unfairness and alternative approaches; Meeting 9, accountability. The programme also includes developing and critiquing AI policy and regulation. Existing systems and institutional research →
Source code to explore: fairlearn/fairlearn. Fairness assessment toolkit for one oversight component; appeals, authority and institutional remedies need a separate process design. Licence / reuse terms: MIT.
Healthcare AI, evidence quality and research ethics
Research gap: Examine evidence standards, sensitive-data use, unequal performance and oversight before findings or tools are relied upon in health settings.
MA foundation: The programme's cross-domain ethical analysis, case-study and applied-framework outcomes; Meeting 9, accountable design. Clinical validation requires additional health-methods expertise and is an application beyond the named syllabus themes. Existing cross-sector evidence applications →
Source code to explore: Project-MONAI/MONAI. Medical-imaging research framework; not a clinically validated service or substitute for domain review. Licence / reuse terms: Apache-2.0.
Employment, income distribution and beneficial automation
Research gap: Measure changes to work, income, entry routes and bargaining power; distinguish potential task exposure from observed displacement, and investigate how gains can be shared.
MA foundation: Meeting 5, AI data work and data–compute–labour relationships; Meeting 7, power and inequality; Meeting 10, human welfare and material consequences. Existing labour-market research →
Source code to explore: mesa/mesa. Agent-based simulation framework; simulated labour outcomes are conditional hypotheses, not economic forecasts. Licence / reuse terms: Apache-2.0.
Energy, computing efficiency and physical infrastructure
Research gap: Investigate resource measurement, efficient computation and the environmental and distributional consequences of AI infrastructure. Distinguish scientific uncertainties from grid, procurement and construction delays.
MA foundation: Meeting 10, the materiality of AI; Meeting 5, data, compute and labour. Hardware optimisation and power-system engineering are specialist extensions. Existing platform-design reference →
Source code to explore: mlco2/codecarbon. Emissions/energy estimation toolkit; distinguish estimates from direct power measurement and grid constraints. Licence / reuse terms: MIT.
Language, culture and unequal model performance
Research gap: Test interpretation and reliability beyond dominant languages and well-represented cultural settings; investigate whose knowledge and experience are omitted from data and evaluation.
MA foundation: Meeting 2, grounded cognition; Meeting 5, data and values; Meeting 8, fairness and counter-imaginaries; Meeting 10, colonial power. Existing research and culture study →
Source code to explore: fairlearn/fairlearn. Disaggregated metric toolkit; cultural meaning, representation and translation quality need context-sensitive human review. Licence / reuse terms: MIT.
Scientific discovery and independent verification
Research gap: Establish whether AI-assisted literature analysis, hypotheses and experiments survive independent checking and improve research outcomes, rather than merely generating plausible material.
MA foundation: Meeting 1, thinking machines; Meeting 4, prediction; Meeting 6, feedback loops; the programme's independent-research outcome. Domain-specific scientific validation is a further methodological extension. Existing evaluation protocol →
Source code to explore: stanford-crfm/helm. Model evaluation framework; scientific discoveries still need domain-specific independent verification. Licence / reuse terms: Apache-2.0.
Public communication, ambiguity and consequential omissions
Research gap: Test whether assisted review preserves qualifications, identifies foreseeable misleading interpretations and improves readers' understanding without inventing intent or suppressing legitimate uncertainty.
MA foundation: Meeting 2, rules and meaning; Meeting 3, AI interfaces; Meeting 9, design and accountability; the programme's communication outcome for diverse audiences. Existing semantic-integrity study →
Source code to explore: explosion/spaCy. Linguistic analysis toolkit; token/dependency analysis is not an automatic truth, omission or intent detector. Licence / reuse terms: MIT.
Programme mapping and evidence for urgency
Programme basis: MA AI, Ethics and Society: Programme Specification 2024–25, revised 18 June 2024, course aims and structure (pp. 2–4), learning outcomes 1–9 (p. 6); and MA History and Foundations of AI, Meetings 1–10 (pp. 1–5). Both documents were supplied by Grace and read for this mapping; the PDFs are not republished here. Existing programme-reference atlas →
Current evidence: The International AI Safety Report 2026 identifies gaps in evaluation, uneven reliability and safeguards. WHO's 2026 report on AI-related health research identifies ethics and oversight gaps. The ILO's occupational-exposure research informs the distinction between potential exposure and observed employment outcomes. The IEA's Key Questions on Energy and AI and Modernising Grids in the Age of Electricity inform the infrastructure questions.
These are selected research priorities, not an official global ranking or claims that the gaps have been resolved. The mapping records intellectual continuity with my MA study and subsequent independent extensions. It does not change the status of existing research results or academic records.
DEEPER RESEARCH · EXPLORATORY PROJECT DESIGN · 5 NOVEMBER 2026
Nine avenues for investigating persistent AI bottlenecks
I intend to explore these avenues more deeply, potentially creating bounded projects to understand the bottlenecks and then write about their mechanisms, evidence and practical consequences. They extend the MA-connected research directions above. Each project below is a possible investigation, not an experiment already completed.
These questions concern capabilities that are not reliably established across demanding conditions, rather than blanket claims of impossibility or assumptions about developers' motives. Existing studies supply starting points; findings must be interpreted within their tested models, tasks and dates.
Continual learning
Bottleneck to investigate: Acquiring new knowledge while retaining earlier knowledge and reusing experience appropriately; retrieval, conversation memory and changes to model weights must be evaluated separately.
Possible exploratory project: Build a small fictional fact stream with scheduled corrections. Compare fresh context, retained conversation and an external memory store; measure retention, useful transfer and obsolete-answer rates. Weight updates can be a later, separately resourced experiment.
Writing focus: Explain when apparent learning is retrieval, what changes persist and where adaptation creates forgetting or interference.
Research starting point: Continual Learning in Large Language Models (2026) →
Source code to explore: ContinualAI/avalanche. Continual-learning framework; small learning experiments do not establish continuous autonomous LLM updating. Licence / reuse terms: MIT.
Machine unlearning
Bottleneck to investigate: Distinguishing removal of a target influence from refusal, temporary suppression or damage to unrelated capabilities; a recovery test can itself introduce the information again.
Possible exploratory project: Using a small open model and fictional data, compare refusal instructions with a documented unlearning method. Separate low-information recovery probes from retraining, and test retained-task performance. A refusal-only pilot can study evaluation design without claiming deletion.
Writing focus: Document what each test establishes about removal, suppression, recoverability and utility; state the model and system boundary.
Research starting point: Existing Large Language Model Unlearning Evaluations Are Inconclusive →
Source code to explore: jaechan-repo/muse_bench. Original MUSE evaluation code; not the exact implementation of every later unlearning critique. Reuse permission is not established. Licence / reuse terms: No top-level licence file identified; reuse permission not established.
Multi-agent collaboration
Bottleneck to investigate: Coordination, shared assumptions, evidence transmission and termination can offset the benefits of role specialisation. Agreement between agents does not establish independent verification.
Possible exploratory project: Compare one agent with a small agent team on the same fictional evidence tasks, under matched total token or cost budgets. Log handoff losses, unsupported agreement, recovery and completion.
Writing focus: Analyse when collaboration improves results, when orchestration introduces errors and which failures are attributable to shared evidence rather than team size.
Research starting point: MAST: Why Do Multi-Agent LLM Systems Fail? →
Source code to explore: microsoft/autogen. Agent orchestration framework; use as a test system rather than claiming it implements the MAST paper. Licence / reuse terms: MIT for code; CC-BY-4.0 for documentation.
Long-horizon computer operation
Bottleneck to investigate: Sustaining state, recognising changes and recovering across multi-application workflows; a successful short demonstration does not establish dependable autonomy.
Possible exploratory project: Construct a sandbox workflow using fictional documents and a mock interface. Vary step count, introduce changed instructions and recoverable errors, and record completion, unintended actions, interventions and time.
Writing focus: Identify which task dependencies cause breakdown, how errors accumulate and whether recovery mechanisms improve end-to-end outcomes.
Research starting point: OSWorld 2.0 research and benchmark →
Source code to explore: xlang-ai/OSWorld. Computer-use benchmark environment; inspect benchmark version before comparing against OSWorld 2.0 claims. Licence / reuse terms: Apache-2.0.
Robotics and embodied intelligence
Bottleneck to investigate: Translating instructions and visual observations into spatially correct actions, with planning and feedback under physical constraints.
Possible exploratory project: Begin in a simulator with a small navigation or manipulation task. Vary object positions, occlusion or instructions; separate perception, planning and execution errors. Physical deployment would require a further safety and validation stage.
Writing focus: Explain where language-level competence fails to transfer to action, including the limits of simulation-to-reality conclusions.
Research starting point: EmbodiedBench →
Source code to explore: EmbodiedBench/EmbodiedBench. Embodied-agent benchmark source; simulator performance does not establish physical-world safety. Reuse permission is not established. Licence / reuse terms: No top-level licence file identified; reuse permission not established.
Physically faithful world models
Bottleneck to investigate: Visual plausibility is different from faithful physical prediction; consistency must be checked against measurable relationships rather than appearance alone.
Possible exploratory project: Choose one simple physical system, such as a pendulum or collision, and compare generated predictions against a controlled simulation. Define measurement tolerances and distinguish observation-extraction errors from physical inconsistencies.
Writing focus: Describe which relationships survive prediction, where they fail and why visual quality alone is an inadequate validation measure.
Research starting point: WorldModelBench →
Source code to explore: WorldModelBench-Team/WorldModelBench. WorldModelBench evaluation script; a VLM judge is not an independent physical-law oracle. Reuse permission is not established. Licence / reuse terms: No top-level licence file identified; reuse permission not established.
Reasoning under increasing complexity
Bottleneck to investigate: Additional reasoning tokens or search may help only under particular conditions; task complexity, output limits and verification quality can confound conclusions.
Possible exploratory project: Create automatically checkable rule or planning problems with controlled depth. Vary the available reasoning budget while recording accuracy, truncation, latency and error type; keep task content and scoring fixed.
Writing focus: Map where extra computation helps, plateaus or harms performance without treating one failure curve as proof that reasoning is impossible.
Research starting point: The Illusion of Thinking: complexity-controlled research →
Source code to explore: sympy/sympy. Symbolic mathematics toolkit for an exact reference baseline; not source code for the cited Apple study. Licence / reuse terms: BSD-style three-clause terms; bundled components have notices.
End-to-end software engineering
Bottleneck to investigate: Producing plausible code is different from diagnosing a system, satisfying requirements and making a change that survives integration and regression checks.
Possible exploratory project: Use a small versioned repository with seeded defects and independent acceptance tests. Compare patch generation with diagnosis-plus-testing workflows; inspect regressions and requirement coverage rather than code appearance.
Writing focus: Investigate the gap between a locally passing patch and a correctly completed engineering task, including the limitations of the test suite itself.
Research starting point: SWE-Lancer →
Source code to explore: openai/frontier-evals. Evaluation collection containing software-engineering research tasks; inspect the selected task and data terms before running. Licence / reuse terms: MIT.
Efficient reasoning on smaller hardware
Bottleneck to investigate: Lower precision or smaller models can change accuracy, generation length and recovery needs; cheaper individual operations need not yield cheaper successful tasks.
Possible exploratory project: Compare two supported precision settings on a modest open model and fixed task set. Measure answer quality, peak memory, wall-clock time and generated tokens on the same hardware; report hardware constraints and avoid inferring energy use without measuring it.
Writing focus: Analyse quality–resource trade-offs using cost or time per successful task, rather than model size or per-token speed alone.
Research starting point: Quantization Inflates Reasoning (2026) →
Source code to explore: huggingface/optimum. Inference optimisation toolkit; backend/hardware support varies, and it is not the implementation of the cited quantisation paper. Licence / reuse terms: Apache-2.0.
Project method: define the claim and system boundary; compare a defensible baseline; control resources and confounders; retain failures and source versions; distinguish observations from explanations. Detailed results belong with the relevant project and existing laboratory records, rather than being repeated across this index.
CURRENT PRIORITY · ARENAS 1–18
I. Safety, Security & Assurance
- AI Safety and Alignment
- AI Security and Cybersecurity
- AI Risk Assessment and Mitigation
- AI Reliability and Robustness
- AI Evaluation, Testing and Validation
- AI Interpretability and Explainability
- AI Transparency and Accountability
- AI Governance, Ethics and Regulation
- AI Privacy and Data Protection
- AI Fairness, Equality and Accessibility
- Human Oversight and Control
- AI Agent Safety and Autonomy
- AI Evidence Verification and Provenance
- AI Misuse Prevention and Incident Response
- AI Monitoring, Auditing and Assurance
- AI Controllability and Corrigibility
- AI Infrastructure and Supply-Chain Security
- AI Societal, Economic and Environmental Risks
FUTURE EXPANSION · ARENAS 19–38
II. Wider Emerging Research Arenas
- Foundational AI and Machine Learning
- AI Architectures and Systems Engineering
- Reasoning, Knowledge and Intelligence
- Artificial General Intelligence and Advanced AI
- Agentic AI and Multi-Agent Systems
- Scientific Discovery and AI for Science
- Robotics and Embodied Intelligence
- Multimodal and Generative AI
- Human–AI Interaction and Collaboration
- AI in Medicine and Life Sciences
- AI in Education and Learning
- AI in Law, Policy and Public Administration
- AI in Economics, Industry and Infrastructure
- Computational Social Science
- AI and Emerging Computing Paradigms
- AI Sustainability and Resource Efficiency
- Frontier Capability Evaluation
- Novel Applications and Translational Research
- Alternative Architectures and Research Paradigms
- Long-Term Technological and Societal Futures
Related work
Research overview · AI Research Laboratory · LLM Evaluation & Governance · AI Product Feedback · Papers
Index established 5 November 2026. This categorisation identifies fields of inquiry, not claims of completed work or expertise in every arena. Individual projects, evidence, findings and future additions will be documented separately; earlier records remain intact.
Exploratory horizon: formation and conflict prevention
Cross-disciplinary research possibilities
Explore the relationships between moral development, faith and secular education, social connection, media systems, workplace formation, institutional incentives, diplomatic early warning and AI evidence verification. These are linked candidate investigations, not new claims of discovery.