How can we reduce the risks of AI?

A comprehensive living database of over 800 AI risk mitigations and a preliminary taxonomy

What is the AI Risk Mitigation Taxonomy?

The AI Risk Mitigation Taxonomy has three parts:

The AI Risk Mitigation Database captures 831 mitigations extracted from 13 existing frameworks and classifications of AI risk mitigations
The Draft AI Risk Mitigation Taxonomy organizes mitigations into 4 main categories and 23 subcategories
The Taxonomy Categories include: Governance & Oversight Controls, Technical & Security Controls, Operational Process Controls, and Transparency & Accountability Controls

The AI Risk Mitigation Taxonomy

The Draft AI Risk Mitigation Taxonomy classifies mitigations into four categories: Governance & Oversight Controls, Technical & Security Controls, Operational Process Controls, and Transparency & Accountability Controls, with 23 subcategories.

Explore the interactive taxonomy
Read our blog post for more detail on how the Taxonomy was constructed

1. Governance & Oversight Controls

Formal organizational structures and policy frameworks that establish human oversight mechanisms and decision protocols to ensure human accountability, ethical conduct, and risk management throughout AI development and deployment.

1.1 Board Structure & Oversight

Governance structures and leadership roles that establish executive accountability for AI safety and risk management.

Examples:

Dedicated risk committees, safety teams, ethics boards, crisis simulation training, multi-party authorization protocols, deployment veto powers

1.2 Risk Management

Systematic methods that identify, evaluate, and manage AI risks for comprehensive risk governance across organizations.

Examples:

Enterprise risk management frameworks, risk registers with capability thresholds, compliance programs, pre-deployment risk assessments, independent risk assessments

1.3 Conflict of Interest Protections

Governance mechanisms that manage financial interests and organizational structures to ensure leadership can prioritize safety over profit motives in critical situations.

Examples:

Background checks for key personnel, windfall profit redistribution plans, stake limitation policies, protections against shareholder pressure

1.4 Whistleblower Reporting & Protection

Policies and systems that enable confidential reporting of safety concerns or ethical violations to prevent retaliation and encourage disclosure of risks.

Examples:

Anonymous reporting channels, non-retaliation guarantees, limitations on non-disparagement agreements, external whistleblower handling services

1.5 Safety Decision Frameworks

Protocols and commitments that constrain decision-making about model development, deployment, and capability scaling, and govern safety-capability resource allocation to prevent unsafe AI advancement.

Examples:

If-then safety protocols, capability ceilings, deployment pause triggers, safety-capability resource ratios

1.6 Environmental Impact Management

Processes for measuring, reporting, and reducing the environmental footprint of AI systems to ensure sustainability and responsible resource use.

Examples:

Carbon footprint assessment, emission offset programs, energy efficiency optimization, resource consumption tracking

1.7 Societal Impact Assessment

Processes that assess AI systems' effects on society, including impacts on employment, power dynamics, political processes, and cultural values.

Examples:

Fundamental rights impact assessments, expert consultations on risk domains, stakeholder engagement processes, governance gap analyses

2. Technical & Security Controls

Technical, physical, and engineering safeguards that secure AI systems and constrain model behaviors to ensure security, safety, alignment with human values, and content integrity.

2.1 Model & Infrastructure Security

Technical and physical safeguards that secure AI models, weights, and infrastructure to prevent unauthorized access, theft, tampering, and espionage.

Examples:

Model weight tracking systems, multi-factor authentication protocols, physical access controls, background security checks, compliance with information security standards

2.2 Model Alignment

Technical methods to ensure AI systems understand and adhere to human values and intentions.

Examples:

Reinforcement learning from human feedback (RLHF), direct preference optimization (DPO), constitutional AI training, value alignment verification systems

2.3 Model Safety Engineering

Technical methods and safeguards that constrain model behaviors and protect against exploitation and vulnerabilities.

Examples:

Safety analysis protocols, capability restriction mechanisms, hazardous knowledge unlearning techniques, input/output filtering systems, defense-in-depth implementations, adversarial robustness training, hierarchical auditing, action replacement

2.4 Content Safety Controls

Technical systems and processes that detect, filter, and label AI-generated content to identify misuse and enable content provenance tracking.

Examples:

Synthetic media watermarking, content filtering mechanisms, prohibited content detection, metadata tagging protocols, deepfake creation restrictions

3. Operational Process Controls

Processes and management frameworks governing AI system deployment, usage, monitoring, incident handling, and validation, which promote safety, security, and accountability throughout the system lifecycle.

3.1 Testing & Auditing

Systematic internal and external evaluations that assess AI systems, infrastructure, and compliance processes to identify risks, verify safety, and ensure performance meets standards.

Examples:

Third-party audits, red teaming, penetration testing, dangerous capability evaluations, bug bounty programs

3.2 Data Governance

Policies and procedures that govern responsible data acquisition, curation, and usage to ensure compliance, quality, user privacy, and removal of harmful content.

Examples:

Harmful content filtering protocols, compliance checks for data collection standards, user data privacy controls, data curation processes

3.3 Access Management

Operational policies and verification systems that govern who can use AI systems and for what purposes to prevent safety circumvention, deliberate misuse, and deployment in high-risk contexts.

Examples:

KYC verification requirements, API-only access controls, fine-tuning restrictions, acceptable use policies, high-stakes application prohibitions

3.4 Staged Deployment

Implementation protocols that deploy AI systems in stages, requiring safety validation before expanding user access or capabilities.

Examples:

Limited API access programs, gradual user base expansion, capability threshold assessments, pre-deployment validation checkpoints, treating model updates as new deployments

3.5 Post-deployment Monitoring

Ongoing monitoring processes that track AI behavior, user interactions, and societal impacts post-deployment to detect misuse, emergent dangerous capabilities, and harmful effects.

Examples:

User interaction tracking systems, capability evolution assessments, periodic impact reports, automated misuse detection, usage pattern analysis tools

3.6 Incident Response & Recovery

Protocols and technical systems that respond to security incidents, safety failures, or capability misuse to contain harm and restore safe operations.

Examples:

Incident response plans, emergency shutdown/rollback procedures, model containment mechanisms, safety drills, critical infrastructure protection measures

4. Transparency & Accountability Controls

Formal disclosure practices and verification mechanisms that communicate AI system information and enable external scrutiny to build trust, facilitate oversight, and ensure accountability to users, regulators, and the public.

4.1 System Documentation

Comprehensive documentation protocols that record technical specifications, intended uses, capabilities, and limitations of AI systems to enable informed evaluation and governance.

Examples:

Model cards, system architecture documentation, compute resource disclosures, safety test result reports, system prompt, model specifications

4.2 Risk Disclosure

Formal reporting protocols and notification systems that communicate risk information, mitigation plans, safety evaluations, and significant AI activities to enable external oversight and inform stakeholders.

Examples:

Publishing risk assessment summaries, pre-deployment notifications to government, reporting large training runs, disclosing mitigation strategies, notifying affected parties

4.3 Incident Reporting

Formal processes and protocols that document and share AI safety incidents, security breaches, near-misses, and relevant threat intelligence with appropriate stakeholders to enable coordinated responses and systemic improvements.

Examples:

Cyber threat intelligence sharing networks, mandatory breach notification procedures, incident database contributions, cross-industry safety reporting mechanisms, standardized near-miss documentation protocols

4.4 Governance Disclosure

Formal disclosure mechanisms that communicate governance structures, decision frameworks, and safety commitments to enhance transparency and enable external oversight of high-stakes AI decisions.

Examples:

Published safety and/or alignment strategies, governance documentation, safety cases, model registration protocols, public commitment disclosures

4.5 Third-Party System Access

Mechanisms granting controlled system access to vetted external parties to enable independent assessment, validation, and safety research of AI models and capabilities.

Examples:

Researcher access programs, third-party capability assessments, government access provisions, legal safe harbors for public interest evaluations

4.6 User Rights & Recourse

Frameworks and procedures that enable users to identify and understand AI system interactions, report issues, request explanations, and seek recourse or remediation when affected by AI systems.

Examples:

User reporting channels, appeal processes, explanation request systems, remediation protocols, content verification

MIT AI Risk Index

AI Risk Mitigation Database

The AI Risk Mitigation Database links each mitigation to the source information (framework title, authors), supporting evidence, and to our Draft AI Risk Mitigation Taxonomy. You can view it below and on Airtable.

How to use the AI Risk Mitigation Taxonomy

Our Database is free to view and use
The Draft AI Risk Mitigation Taxonomy can be used to filter this database to identify specific mitigations, for instance, mitigations related to Testing & Auditing or Risk Management
The Taxonomy Categories can be used together to understand how different types of controls (i.e., Governance & Oversight, Technical & Security, Operational Process, and Transparency & Accountability) relate to each other. For example, to identify pre-deployment and post-deployment mitigations across different control types
‍Offer feedback or suggest missing resources, or risks, here, or email airisk[at]mit.edu

The AI Risk Mitigation Taxonomy provides:

An accessible overview of organizational mitigations for AI risks
A regularly updated source of information about new mitigation frameworks
A common frame of reference for researchers, developers, businesses, evaluators, auditors, policymakers, and regulators
A resource to help develop research, curricula, audits, and policy
An easy way to find relevant mitigations and frameworks

We provide examples of how different audiences might use the Mitigation Taxonomy below.