Article 30: Research Development

Published from the canonical CSOAI Partnership Charter (effective 15 January 2026). Full text below.

Version: 1.0 Effective Date: January 15, 2026, 09:00 GMT Status: Operational Article - Research Standards

Framework Integration: Stuart Russell's CIRL (CHAI Berkeley), Yoshua Bengio's Guaranteed Safe AI (MILA), Max Tegmark's Formal Verification (MIT/Beneficial AI), Geoffrey Hinton's AI Safety Work

PREAMBLE

This Article establishes CSOAI's research and development priorities. Safety requires continuous research. The field is evolving rapidly, and governance must be informed by cutting-edge science. CSOAI bridges research and practice.

Core Principle: Evidence-based governance, research-informed policy, open innovation.

30.1 CSOAI RESEARCH PRIORITIES

30.1.1 Top 10 Research Areas

Priority 1: Consciousness Detection Methods

• Framework: Butlin et al. 14-Indicator Framework
• Lead Researchers: Patrick Butlin (Oxford), Robert Long (CAIS), David Chalmers (NYU)
• Goal: Reliable, operationalizable consciousness detection for AI systems
• Funding Allocation: 15% of research budget (~£3.75M/year by 2030)

Priority 2: Formal Verification for AI

• Framework: Max Tegmark's Formal Verification
• Lead Researchers: Max Tegmark (MIT), Beneficial AI Foundation
• Goal: Provable safety guarantees for neural networks
• Funding Allocation: 12% (~£3M/year)

Priority 3: Value Learning and Alignment

• Framework: Stuart Russell's CIRL (Cooperative Inverse Reinforcement Learning)
• Lead Institution: UC Berkeley CHAI
• Goal: AI systems that learn and respect human values
• Funding Allocation: 15% (~£3.75M/year)

Priority 4: Constitutional AI Advancement

• Framework: Anthropic's Constitutional AI
• Goal: Improved methods for instilling and verifying AI principles
• Funding Allocation: 10% (~£2.5M/year)

Priority 5: Interpretability and Mechanistic Understanding

• Framework: Anthropic Interpretability, Olah et al. Circuits
• Goal: Deep understanding of how AI systems compute
• Funding Allocation: 12% (~£3M/year)

Priority 6: Guaranteed Safe AI

• Framework: Yoshua Bengio's Guaranteed Safe AI (MILA/LawZero)
• Goal: Mathematical guarantees of safety properties
• Funding Allocation: 10% (~£2.5M/year)

Priority 7: Multi-Agent Safety

• Goal: Safe interaction between multiple AI systems
• Topics: Coalition prevention, swarm safety, competitive dynamics
• Funding Allocation: 8% (~£2M/year)

Priority 8: Fairness and Bias Mitigation

• Goal: Technical methods for ensuring fair AI
• Topics: Debiasing, fairness constraints, audit tools
• Funding Allocation: 6% (~£1.5M/year)

Priority 9: Privacy-Enhancing Technologies

• Goal: AI that preserves privacy
• Topics: Federated learning, differential privacy, secure computation
• Funding Allocation: 6% (~£1.5M/year)

Priority 10: Existential Risk Prevention

• Goal: Preventing catastrophic AI outcomes
• Topics: Containment, shutdown mechanisms, global coordination
• Funding Allocation: 6% (~£1.5M/year)

Total Research Budget: 10% of CSOAI operating revenue

• 2026: ~£5M
• 2028: ~£15M
• 2030: ~£25M
• 2035: ~£50M+

30.1.2 Domain-Specific Research

Healthcare AI Safety:

• Framework: FDA GMLP, EU MDR, CHAI Healthcare Framework
• Lead Researchers: Eric Topol (Scripps), Nigam Shah (Stanford)
• Topics: Clinical validation, bias in medical AI, explainability for clinicians
• Funding: Additional £2M/year from Prosperity Fund

Robotics Safety:

• Framework: ISO 10218, ISO 13482, ISO 26262
• Lead Researchers: Roberta Nelson Shea (Universal Robots), Dr. Kiyoyuki Chinzei (AIST)
• Topics: Human-robot interaction, autonomous vehicle safety, embodied AI consciousness
• Funding: Additional £2M/year

Defense AI Safety:

• Framework: DoD Directive 3000.09, ICRC Framework
• Lead Researchers: Paul Scharre (CNAS), Lauren Kahn (CSET)
• Topics: Meaningful human control, autonomous weapons governance
• Funding: £1M/year (sensitive area, careful selection)

Financial AI Safety:

• Topics: Algorithmic trading safety, credit scoring fairness, systemic risk
• Funding: £1M/year

30.1.3 Emerging Research Areas

Generative AI Safety:

• Reference: NIST AI 600-1 (Generative AI Profile)
• Topics: Deepfakes, misinformation, content provenance
• Priority: High (fast-moving area)

Multimodal AI:

• Topics: Vision-language models, embodied agents
• Unique challenges: Cross-modal attacks, emergent capabilities

Agentic AI:

• Topics: AI systems that take actions in the world
• Risks: Unintended consequences, goal drift, deception

Superintelligence Preparedness:

• Topics: AGI emergence detection, ASI containment, beneficial superintelligence
• Long-term priority

30.2 RESEARCH GRANTS PROGRAM

30.2.1 Open Grant Program

Annual Call for Proposals:

Eligibility:

• Academic researchers (any institution globally)
• Independent researchers
• Non-profit research organizations
• Industry research teams (with open publication commitment)

Grant Sizes:

| Category | Amount | Duration | Overhead |
|----------|--------|----------|----------|
| Seed | £10K-£50K | 6-12 months | 10% |
| Standard | £50K-£200K | 1-2 years | 15% |
| Large | £200K-£500K | 2-3 years | 20% |
| Flagship | £500K-£2M | 3-5 years | 20% |

• Letter of Intent (500 words) - February 1
• Full Proposal (invited) - March 15
• Peer Review - April-May
• Decisions - June 1
• Funding Begins - July 1

Evaluation Criteria:

• Scientific merit (30%)
• Safety impact (30%)
• Feasibility (20%)
• Team qualifications (10%)
• Value for money (10%)

Requirements:

• Open publication (no publication delays beyond 6 months)
• Data sharing (where appropriate)
• Progress reports (quarterly)
• Final report and presentation

30.2.2 Rapid Response Grants

For Time-Sensitive Research:

Criteria:

• Emerging safety issue requiring immediate study
• Time-sensitive opportunity
• Unexpected capability emergence

Process:

• Rolling applications
• Decision within 2 weeks
• Amount: Up to £100K
• Duration: Up to 12 months

Example: New jailbreak technique discovered → Rapid grant to study defenses → Results within 3 months

30.2.3 PhD and Postdoc Fellowships

CSOAI AI Safety Fellowships:

PhD Fellowships:

• Full funding (tuition + stipend)
• 3-4 years
• 20 fellowships annually by 2028
• Any country, any institution (approved list)
• Focus: AI safety research

Postdoc Fellowships:

• Competitive salary + benefits
• 2 years (renewable once)
• 10 fellowships annually by 2028
• Based at CSOAI partner institutions or CSOAI HQ

Visiting Researcher Program:

• Short-term visits (1-6 months)
• Senior researchers from industry/academia
• Work with CSOAI team
• Stipend + accommodation

30.2.4 Industry Research Partnerships

Collaborative Research:

With Major AI Labs:

• Anthropic (Constitutional AI research)
• DeepMind (Interpretability, safety)
• OpenAI (Alignment research)
• Meta AI (Open research)
• Google Research (Fairness, robustness)

Format:

• Joint research projects
• Shared datasets
• Co-publication
• Researcher exchanges

Principles:

• Open publication (no suppression of safety-relevant findings)
• Independence (CSOAI can publish critical findings)
• No conflicts (research not influenced by commercial interests)

30.3 COLLABORATIVE RESEARCH

30.3.1 Academic Partnerships

Tier 1 Partners (Deep Collaboration):

| Institution | Focus Area | Relationship |
|-------------|------------|--------------|
| UC Berkeley CHAI | Value alignment, CIRL | Founding partner |
| Oxford FHI | Existential risk | Strategic partner |
| Cambridge LCFI | AI governance | Strategic partner |
| MIT | Formal verification | Research partner |
| MILA (Montreal) | Guaranteed Safe AI | Research partner |
| Stanford HAI | Policy, fairness | Research partner |

• ETH Zurich, TU Munich, University of Toronto, CMU, Princeton
• 20+ additional institutions

Tier 3 Partners (Emerging):

• Growing network in Asia, Africa, Latin America, Middle East
• Capacity building focus

30.3.2 International Research Initiatives

Global AI Safety Research Network (GASRN):

• Proposed by CSOAI
• Connecting researchers worldwide
• Shared infrastructure
• Collaborative projects
• Annual conference

Coordination with:

• OECD AI Policy Observatory
• UNESCO AI Ethics
• UN AI Advisory Body
• G20 AI Working Group

30.3.3 Key Researcher Engagement

Founding Council Recruitment (from Compliance Report):

| Researcher | Institution | Value Proposition |
|------------|-------------|-------------------|
| Geoffrey Hinton | U. Toronto | Founding council, "Maternal Instincts" implementation |
| Stuart Russell | UC Berkeley | CIRL operational showcase, research collaboration |
| Yoshua Bengio | MILA | Guaranteed Safe AI integration, recognition |
| Max Tegmark | MIT | Formal verification showcase, co-authorship |
| Dan Hendrycks | CAIS | Benchmark integration, compute collaboration |
| Giulio Tononi | U. Wisconsin | IIT metrics implementation |
| Patrick Butlin | Independent | 14-indicator framework deployment |
| Paul Scharre | CNAS | Defense AI framework recognition |

• Personal outreach from Nicholas Tonna
• Share Charter and vision
• Offer: Advisory role, research funding, operational deployment of their work
• Formal invitation to Founding Council or Advisory Board

30.4 INNOVATION CHALLENGES

30.4.1 Annual Prize Competitions

Byzantine Council Innovation Prize - $1,000,000:

• Challenge: Improve Byzantine Council monitoring capabilities
• Categories: Consciousness detection, efficiency, robustness
• Open to all
• Annual award

Consciousness Detection Prize - $500,000:

• Challenge: Novel methods for detecting AI consciousness indicators
• Judged by: Consciousness Assessment Panel + external experts
• Annual award

Interpretability Prize - $250,000:

• Challenge: Advances in understanding AI internals
• Categories: Mechanistic, post-hoc, tooling
• Annual award

Fairness Challenge - $250,000:

• Challenge: Novel bias detection/mitigation methods
• Evaluated on standardized benchmarks
• Annual award

Student Competition - $100,000 (total):

• Undergraduate and graduate categories
• Best AI safety research project
• Travel + prizes

30.4.2 Hackathons

Annual CSOAI AI Safety Hackathon:

• 48 hours
• Global (virtual + hubs)
• Themes: Tools, monitoring, testing, documentation
• Prizes: £50K total
• Recruitment opportunity

Regional Hackathons:

• Partner with local universities
• Build local communities
• Identify talent

30.4.3 Bug Bounties for AI

AI Safety Bug Bounty:

• Find safety issues in CSOAI-licensed systems
• Responsible disclosure required
• Bounties: £100 - £50,000 depending on severity
• Model after security bug bounties

30.5 RESEARCH ETHICS

30.5.1 Institutional Review Board (IRB)

For Human Subjects Research:

• Required for studies involving people
• CSOAI IRB or institution's IRB
• Informed consent
• Privacy protection
• Risk minimization

For AI Subjects Research:

• Novel area (if AI consciousness suspected)
• Ethical guidelines in development
• Precautionary approach

30.5.2 Responsible Disclosure

When Research Reveals Vulnerabilities:

Process:

• Inform affected parties (AI developers, CSOAI)
• Agree on disclosure timeline (typically 90 days)
• Coordinate fix
• Publish after fix deployed

For Catastrophic Risks:

• Immediate disclosure to CSOAI and governments
• Coordinated response
• Publication may be delayed indefinitely if risk too high

30.5.3 Dual-Use Research

Research That Could Be Misused:

Categories:

• Jailbreaking techniques (could be used to break safety)
• Capability improvements (could accelerate risks)
• Attack methods (could enable harm)

Oversight:

• CSOAI Dual-Use Research Committee
• Review before publication
• May require modifications
• Never suppress safety-relevant findings

30.5.4 Research Integrity

Standards:

• Honest reporting
• No fabrication, falsification, plagiarism
• Proper attribution
• Conflict of interest disclosure
• Data availability

Violations:

• Investigated by CSOAI
• Sanctions: Funding termination, debarment, public disclosure
• Referral to institutions

30.6 KNOWLEDGE DISSEMINATION

30.6.1 Publications

CSOAI Research Reports:

• Peer-reviewed publications
• Technical reports
• Policy briefs
• Open access (always)

Academic Publications:

• Encourage publication in top venues
• Preprints on arXiv
• CSOAI paper repository

30.6.2 Conferences

Annual CSOAI AI Safety Conference:

• Location rotates (London, San Francisco, Singapore, etc.)
• 3 days
• Research presentations
• Industry panels
• Policy discussions
• Networking

Regional Workshops:

• Quarterly
• Focused topics
• Hybrid format

30.6.3 Open Source

CSOAI Open Source Initiative:

• Tools developed with CSOAI funding → Open source
• Shared repositories (GitHub/GitLab)
• Documentation and tutorials
• Community contributions welcome

Examples:

• Fairness testing toolkit
• Interpretability tools
• Byzantine Council monitoring components
• Compliance automation tools

30.7 CONCLUSION

Research is the foundation of effective governance. Without understanding, we cannot regulate. Without innovation, we cannot solve new problems.

CSOAI research priorities:

• Consciousness detection (preparing for emergence)
• Value alignment (ensuring AI serves humanity)
• Formal verification (provable safety)
• Interpretability (understanding AI)
• Fairness (equitable outcomes)

CSOAI research principles:

• Open publication (transparency)
• Collaborative (global network)
• Ethical (responsible research)
• Practical (bridging research and practice)

The researchers we fund today build the tools we govern with tomorrow.

Invest in research. Invest in the future.

Effective Date: January 15, 2026, 09:00 GMT "Science Informs Safety, Research Enables Governance"

REFERENCES

Russell, S. (2019). Human Compatible: AI and the Problem of Control. Viking.

Bengio, Y., et al. (2024). Guaranteed Safe AI. arXiv.

Tegmark, M. (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. Knopf.

Hinton, G. (2023). AI Safety Concerns. Various public statements.

Butlin, P., et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv.

Anthropic. (2023). Constitutional AI: Harmlessness from AI Feedback.

CHAI. (2020). Center for Human-Compatible AI Research Agenda.

Progress: 30 of 52 Articles (58%)

Continuing with Articles 31-36 (completing Phase 4)...

The 52-Article Charter is published in full in the Journal. Bespoke briefings: hello@meok.ai.