Human Oversight Quality
Parameter
Human Oversight Quality
Importance82
Direction▲Higher is better
Current TrendDeclining (capability gap widening, automation bias increasing)
MeasurementReview effectiveness, decision authority, error detection rates
Related
Prioritization
Importance82
Tractability55
Neglectedness50
Uncertainty45
Overview
Section titled “Overview”Human Oversight Quality measures the effectiveness of human supervision over AI systems—encompassing the ability to review AI outputs, maintain meaningful decision authority, detect errors and deception, and correct problematic behaviors before harm occurs. Higher oversight quality is better—it serves as a critical defense against AI failures, misalignment, and misuse.
AI capability levels, oversight method sophistication, evaluator training, and institutional design all shape whether oversight quality improves or degrades. This parameter is distinct from human agency (personal autonomy) and human expertise (knowledge retention), though it depends on both.
This parameter underpins:
- AI safety: Detecting and preventing harmful AI behaviors
- Accountability: Assigning responsibility for AI actions
- Error correction: Catching mistakes before consequences
- Democratic control: Ensuring AI serves human values
This framing enables:
- Capability gap tracking: Monitoring as AI exceeds human understanding
- Method development: Designing better oversight approaches
- Institutional design: Creating effective oversight structures
- Progress measurement: Evaluating oversight interventions
Parameter Network
Section titled “Parameter Network”Loading diagram...
Contributes to: Misalignment Potential
Primary outcomes affected:
- Existential Catastrophe ↓↓ — Oversight catches dangerous behaviors before catastrophe
- Steady State ↓ — Quality oversight preserves human agency in the long term
Current State Assessment
Section titled “Current State Assessment”Oversight Capability by Domain
Section titled “Oversight Capability by Domain”| Domain | Human Expert Performance | AI Performance | Oversight Gap | Trend | Year |
|---|---|---|---|---|---|
| Chess | ~2800 Elo (Magnus Carlsen) | ~3600+ Elo (Stockfish) | Severe | Widening | 2024 |
| Go | 9-dan professionals | Superhuman since 2016 | Severe | Stable (adapted) | 2016+ |
| Sorting algorithms | Human-optimized (decades) | 70% faster (AlphaDev) | Severe | Widened | 2024 |
| Mathematical proof | 90% on MATH benchmark | 84.3% accuracy (GPT-4) | Moderate | Narrowing | 2025 |
| Code generation (2hr tasks) | Human baseline | 4x higher on RE-Bench | Severe | Widening | 2024 |
| Code generation (32hr tasks) | Human baseline | 0.5x performance vs humans | Reversed | Humans ahead | 2024 |
| Medical diagnosis | Specialist accuracy | Matches/exceeds in narrow domains | Moderate | Widening | 2024 |
| Software development (complex) | Skilled developers | 30.4% autonomous completion | Moderate | Widening | 2025 |
| Administrative work | Office workers | 0% autonomous completion | No gap | Humans dominant | 2025 |
Note: Oversight quality degrades as AI performance exceeds human capability in specific domains. Time-constrained tasks favor AI; extended deliberation favors humans (2-to-1 at 32 hours vs. 2 hours).
Domain-Specific Oversight Requirements
Section titled “Domain-Specific Oversight Requirements”| Domain | Current AI Role | Required Oversight Level | Regulatory Status | Key Challenge |
|---|---|---|---|---|
| Aviation autopilot | Flight path management | Continuous monitoring (dual pilots) | FAA mandatory | 73% show monitoring complacency |
| Medical diagnosis | Decision support | Physician review required | FDA varies by device | 70-80% accept without verification |
| Criminal sentencing | Risk assessment | Judge retains authority | State-dependent | High weight on algorithmic scores |
| Autonomous weapons | Target identification | Meaningful human control required | International debate | Attribution and accountability gaps |
| Financial trading | Execution decisions | Post-hoc audit only | SEC circuit breakers | Millisecond decisions exceed human oversight |
| Hiring screening | Resume filtering | Varies by jurisdiction | GDPR Article 22 in EU | 60-70% follow recommendations |
| Content moderation | Flagging decisions | Human review of appeals | Platform-specific | 65% over-reliance on AI flags |
| Credit decisions | Loan approval | EU AI Act high-risk | Regulated in EU (2026) | Opacity of decision factors |
Note: Domains with slower decision timelines enable more effective oversight. Real-time systems (trading, autonomous weapons) pose fundamental oversight challenges.
Automation Bias Evidence
Section titled “Automation Bias Evidence”| Context | Automation Bias Rate | Impact | Source |
|---|---|---|---|
| Aviation (autopilot monitoring) | 73% show monitoring issues | Accidents attributed to complacency | [e6b22bc6e1fad7e9] |
| Medical AI recommendations | 70-80% acceptance without verification | Diagnostic error propagation | Clinical AI studies |
| Content moderation | 65% over-reliance on AI flags | False positive/negative patterns | Platform audits |
| Hiring algorithms | 60-70% follow AI recommendations | Bias perpetuation | Reuters investigation↗ |
| Criminal justice (risk scores) | High weight on algorithmic scores | Sentencing affected | COMPAS studies |
Scalable Oversight Method Effectiveness
Section titled “Scalable Oversight Method Effectiveness”| Method | Current Accuracy | Domain | Key Limitation | Year |
|---|---|---|---|---|
| Process supervision | 78.2% (vs 72.4% outcome-based) | Mathematics | Limited to domains with clear reasoning steps↗ | 2023 |
| AI debate | 60-80% factual; 50-65% complex | Factual questions | Vulnerable to sophisticated deception↗ | 2024 |
| Nested oversight (Elo gap 400) | 13.5-51.7% success rates | Game scenarios | Diminishing success as capability gaps widen | 2025 |
| Nested oversight (Mafia) | 13.5% at 400 Elo gap | Adversarial games | Severe degradation with capability gaps | 2025 |
| Nested oversight (Debate) | 51.7% at 400 Elo gap | Structured debate | Better than Mafia but still fragile | 2025 |
| Recursive reward modeling | 2-3 levels validated | Mathematical proofs | Decomposition limits unclear | 2023 |
| Constitutional AI | Variable | General alignment | Depends on constitutional quality | 2023 |
Sources: OpenAI: Let’s Verify Step by Step↗, Debate training research↗, MIT Scaling Laws for Scalable Oversight (2025)
What “Healthy Human Oversight” Looks Like
Section titled “What “Healthy Human Oversight” Looks Like”Effective human oversight involves:
- Evaluative capability: Humans can assess AI output quality
- Error detection: Humans can identify when AI is wrong or deceptive
- Decision authority: Humans retain meaningful control over consequential choices
- Correction capacity: Humans can modify AI behavior when needed
- Understanding: Humans comprehend what AI is doing and why
Effective vs. Nominal Oversight
Section titled “Effective vs. Nominal Oversight”| Effective Oversight | Nominal Oversight |
|---|---|
| Human understands AI reasoning | Human sees only outputs |
| Human can detect errors | Human trusts without verification |
| Human retains veto power | Human rubber-stamps AI decisions |
| Time allocated for review | Pressure to accept quickly |
| Trained for AI evaluation | Generic operator training |
| Accountability enforced | Diffuse responsibility |
Factors That Decrease Oversight Quality (Threats)
Section titled “Factors That Decrease Oversight Quality (Threats)”Loading diagram...
The Evaluation Difficulty Problem
Section titled “The Evaluation Difficulty Problem”As AI capabilities increase, human evaluation becomes progressively more difficult:
| AI Capability Level | Human Evaluation Capability | Oversight Quality |
|---|---|---|
| Below human level | Can verify correctness | High |
| Human level | Can assess with effort | Moderate |
| Above human level | Cannot reliably evaluate | Low |
| Far above human level | Fundamentally unable to evaluate | Nominal only |
Automation Bias Mechanisms
Section titled “Automation Bias Mechanisms”| Mechanism | Description | Prevalence |
|---|---|---|
| Complacency | Reduced vigilance when AI usually correct | Very high |
| Authority deference | Treating AI as expert authority | High |
| Cognitive load reduction | Accepting AI to reduce effort | Very high |
| Responsibility diffusion | ”AI decided, not me” | High |
| Confidence in technology | Overestimating AI reliability | High |
Speed-Oversight Tradeoff
Section titled “Speed-Oversight Tradeoff”| System Type | Decision Speed | Human Review Time | Oversight Quality |
|---|---|---|---|
| Algorithmic trading | Milliseconds | None possible | Zero (no human oversight) |
| Content moderation | Seconds | 0.5-2 seconds | Very low |
| Hiring screening | Seconds per application | Minutes if any | Low-moderate |
| Medical diagnosis AI | Seconds | 5-30 minutes possible | Moderate if enforced |
| Strategic decisions | Hours-days | Adequate if required | Can be high |
Deception Detection Challenges
Section titled “Deception Detection Challenges”| AI Behavior Type | Human Detection Rate | Challenge |
|---|---|---|
| Obvious errors | High (80%+) | Detectable |
| Subtle errors | Moderate (40-60%) | Requires expertise |
| Persuasive false outputs | Low (20-40%) | Appears correct to non-experts |
| Sophisticated deception | Very low (<20%) | Designed to evade detection |
| Deceptive alignment | Unknown (untested) | May be undetectable |
Factors That Increase Oversight Quality (Supports)
Section titled “Factors That Increase Oversight Quality (Supports)”Scalable Oversight Methods
Section titled “Scalable Oversight Methods”| Method | How It Helps | Effectiveness | Status |
|---|---|---|---|
| AI Debate | Two AIs argue; human judges simpler argument | 60-80% on factual questions | Research stage↗ |
| Process Supervision | Reward each reasoning step, not just output | +6% accuracy on MATH | Deployed (OpenAI o1)↗ |
| Recursive Reward Modeling | Break complex tasks into human-evaluable parts | 2-3 levels validated | Research stage↗ |
| Amplification | AI assists human evaluator | Variable | Research stage↗ |
| Constitutional AI | AI self-critiques against principles | Reduces harmful outputs | Deployed (Anthropic)↗ |
Regulatory Interventions
Section titled “Regulatory Interventions”| Regulation | Mechanism | Status | Oversight Impact | Implementation Timeline |
|---|---|---|---|---|
| EU AI Act Article 14 | Mandatory human oversight for high-risk AI | In force Aug 2024 | Requires competent oversight persons with training/authority | Full application Aug 2026 |
| EU AI Act Article 26 | Deployer obligations for oversight assignment | In force Aug 2024 | Assigns specific individuals to monitor each system | Full application Aug 2026 |
| EU AI Act (biometric systems) | Dual verification requirement | In force Aug 2024 | At least 2 competent persons verify critical decisions | Full application Aug 2026 |
| GDPR Article 22 | Right to human review of automated decisions | Active (2018) | Creates individual review rights | Active |
| US Executive Order 14110 | Federal AI oversight requirements | 2024-2025 | Agency-level oversight mandates | Phased implementation |
| Sector-specific rules | Aviation (FAA), medical (FDA) requirements | Active | Domain-specific oversight | Active |
Institutional Design
Section titled “Institutional Design”| Design Element | How It Improves Oversight | Implementation |
|---|---|---|
| Mandatory review periods | Forces time for human evaluation | Some high-stakes domains |
| Dual-key systems | Requires multiple human approvals | Nuclear, some financial |
| Red teams | Dedicated adversarial oversight | Major AI labs |
| Independent auditors | External oversight of AI systems | Emerging (EU AI Act) |
| Whistleblower protections | Enables internal oversight reporting | Variable by jurisdiction |
Evaluator Training
Section titled “Evaluator Training”| Training Type | Skill Developed | Evidence of Effectiveness |
|---|---|---|
| AI error detection | Identify AI mistakes | 30-40% improvement with training |
| Calibration training | Know when to trust AI | 73% improvement in confidence accuracy↗ |
| Adversarial thinking | Assume AI might deceive | Improves skeptical evaluation |
| Domain specialization | Deep expertise in one area | Enables expert-level oversight |
Why This Parameter Matters
Section titled “Why This Parameter Matters”Consequences of Low Oversight Quality
Section titled “Consequences of Low Oversight Quality”| Consequence | Mechanism | Severity | 2025 Evidence |
|---|---|---|---|
| Undetected errors propagate | AI mistakes not caught before harm | High | AI oversight deficit widening |
| Accountability collapse | No one responsible for AI decisions | High | Distributed social capacity needed |
| Deceptive AI undetected | Cannot catch misaligned behavior | Critical | Nested oversight only 13.5-51.7% effective |
| Automation bias accidents | Over-reliance on faulty AI | High | 70-80% acceptance without verification |
| Democratic legitimacy loss | AI decisions without human consent | High | Procedural compliance insufficient |
| Competency gap crisis | Human skills not developing at AI pace | Critical | 2025 Global Data Literacy Benchmark |
Oversight Quality and Existential Risk
Section titled “Oversight Quality and Existential Risk”Human oversight quality is central to AI safety:
- Alignment verification: Detecting if AI goals match human values requires oversight
- Correction capability: Stopping harmful AI requires effective human control
- Deceptive alignment detection: Identifying AI deception requires evaluative capability
- Gradual testing: Safely scaling AI requires oversight at each stage
- Emergency response: Responding to AI failures requires understanding what happened
Critical Oversight Thresholds
Section titled “Critical Oversight Thresholds”| Threshold | Definition | Current Status |
|---|---|---|
| Meaningful oversight | Humans can catch most consequential errors | At risk in some domains |
| Deception detection | Humans can identify AI attempts to deceive | Unknown capability |
| Correction capability | Humans can modify AI behavior effectively | Currently maintained |
| Scalable oversight | Oversight methods scale with AI capabilities | Under development |
Trajectory and Scenarios
Section titled “Trajectory and Scenarios”Capability Gap Projections
Section titled “Capability Gap Projections”| Timeframe | AI Capability Level | Human Oversight Capability | Gap Assessment |
|---|---|---|---|
| 2025-2026 | Exceeds experts in narrow domains | Expert evaluation still possible | Moderate, manageable |
| 2027-2028 | Exceeds experts in broader domains | Scalable oversight methods needed | Growing concern |
| 2029-2030 | Potentially superhuman broadly | Oversight method effectiveness critical | Critical uncertainty |
Scenario Analysis
Section titled “Scenario Analysis”| Scenario | Probability | Key Driver | Outcome |
|---|---|---|---|
| Effective scalable oversight | 25-35% | Oversight methods work at scale | Meaningful human control maintained |
| Nominal oversight | 35-45% | Formal oversight without real capability | Humans sign off but can’t evaluate |
| Oversight failure | 15-25% | AI exceeds all oversight methods | Loss of meaningful human control |
| AI-assisted oversight | 20-30% | AI helps humans oversee AI | Mixed effectiveness, new risks |
Quantitative Framework for Oversight Degradation
Section titled “Quantitative Framework for Oversight Degradation”The Capability Gap Threshold Model
Section titled “The Capability Gap Threshold Model”Research from MIT (2025) quantifies how oversight effectiveness degrades with capability gaps using an Elo-based framework:
| Capability Gap (Elo) | Nested Oversight Success Rate | Interpretation |
|---|---|---|
| 0-100 | 70-90% | Human can effectively oversee similar-capability AI |
| 100-200 | 50-70% | Oversight becomes unreliable |
| 200-300 | 30-50% | Majority of oversight failures |
| 300-400 | 13.5-51.7% | Severe oversight degradation |
| 400+ | Below 15% (projected) | Near-total oversight failure |
Key insight: The MIT scaling laws research shows oversight effectiveness is not binary but degrades continuously as capability gaps widen. At 400 Elo gap—roughly the difference between a strong amateur and a grandmaster—even well-designed nested oversight systems fail 48-86% of the time depending on domain.
Oversight Success Factors
Section titled “Oversight Success Factors”Research identifies four critical factors for effective oversight (ACM FAccT 2024):
- Causal power — Authority to intervene in system operation (25-35% weight)
- Epistemic access — Ability to understand what system is doing (30-40% weight)
- Self-control — Capacity to resist automation bias (15-20% weight)
- Role-appropriate intentions — Proper motivation and accountability (15-20% weight)
Missing any single factor reduces oversight effectiveness by 40-60%. All four must be present for “effective” oversight as defined by EU AI Act Article 14.
Time-Capability Tradeoff
Section titled “Time-Capability Tradeoff”Stanford AI Index 2025 reveals a critical insight: oversight effectiveness depends heavily on available time:
| Time Allocation | Human Performance | AI Performance | Oversight Quality |
|---|---|---|---|
| 2 hours | Baseline | 4x higher (RE-Bench) | AI exceeds oversight capacity |
| 8 hours | 1.5x baseline | 2x baseline | Oversight becomes viable |
| 32 hours | 2x baseline | 1x baseline | Humans outperform, can oversee |
Policy implication: Mandatory review periods for consequential AI decisions could substantially improve oversight quality. The EU AI Act requires “appropriate time and resources” but doesn’t specify minimum durations.
Emerging Research Directions (2024-2025)
Section titled “Emerging Research Directions (2024-2025)”1. Hybrid Human-AI Oversight Systems
Section titled “1. Hybrid Human-AI Oversight Systems”Recent research explores using AI to assist human oversight—a potentially double-edged approach:
Advantages:
- AI tools can help humans detect patterns in AI outputs
- Process visualization helps humans understand complex reasoning
- Statistical anomaly detection flags suspicious behaviors
- Amplification techniques extend human judgment
Risks:
- Creates second-order oversight problem (who oversees the oversight AI?)
- May increase automation bias if humans over-rely on oversight tools
- Anthropic 2025 analysis warns this may be “substantially infeasible” for ASI
Current status: Research stage; not yet deployed in high-stakes domains.
2. Distributed Social Oversight Capacity
Section titled “2. Distributed Social Oversight Capacity”ArXiv 2024 research argues oversight should be reconceived as “distributed social capacity” rather than concentrated institutional control:
Key concepts:
- Multiple oversight layers (individual users, deployers, auditors, regulators)
- Whistleblower protections enable internal oversight
- Public participation in high-stakes AI governance
- Cross-institutional coordination mechanisms
Challenges:
- Coordination costs increase with distributed systems
- Diffused responsibility may reduce accountability
- Requires substantial institutional capacity building
3. AI Literacy and Oversight Competency
Section titled “3. AI Literacy and Oversight Competency”The 2025 Global Data Literacy Benchmark reveals a competency crisis: while AI systems embed into decision-making, human competencies to guide, question, and validate those systems are not developing at the same pace.
Gap metrics:
- 58% of professionals report insufficient training to oversee AI systems
- 42% cannot identify when AI outputs are unreliable
- 73% lack understanding of AI system limitations
- 67% cannot explain AI decisions to stakeholders
Interventions:
- EU AI Act requires AI literacy for anyone operating AI systems
- Specialized training programs for high-stakes domains
- Calibration training showing 73% improvement in confidence accuracy
- Domain-specific certification programs
4. Formal Verification of Oversight Properties
Section titled “4. Formal Verification of Oversight Properties”Emerging approach: mathematically verify oversight system properties rather than relying on empirical testing:
Verifiable properties:
- Minimum detection rates for specified error types
- Upper bounds on false negative rates
- Guaranteed human intervention points
- Provable impossibility of certain failure modes
Status: Theoretical frameworks exist; practical implementation limited to narrow domains (e.g., aviation autopilot monitoring).
5. Oversight Benchmarks and Evaluation
Section titled “5. Oversight Benchmarks and Evaluation”Future of Life Institute’s AI Safety Index now tracks oversight capabilities as a key safety metric:
Tracked dimensions:
- Process supervision accuracy across domains
- Scalable oversight method effectiveness at various capability gaps
- Regulatory compliance with oversight requirements
- Incident rates in systems with vs. without effective oversight
Need: Standardized benchmarks for comparing oversight approaches across different AI systems and deployment contexts.
Key Debates
Section titled “Key Debates”Can Scalable Oversight Work?
Section titled “Can Scalable Oversight Work?”Optimistic view:
- Recursive reward modeling↗ and debate↗ show promising results (51.7% success at 400 Elo gap)
- Process supervision deployed successfully (OpenAI o1)
- Multiple methods can combine for defense-in-depth
- Effective oversight is possible with systematic procedures and interdisciplinary cooperation (2025 expert consensus)
Skeptical view:
- Nested oversight success rates decline sharply as capability gaps widen (13.5-51.7% at 400 Elo gap)
- Sophisticated deception may defeat all oversight
- Arms race dynamics favor AI over oversight
- Exclusive reliance on scalable oversight may be “substantially infeasible” for controlling ASI (Anthropic 2025)
- MIT research quantifies fragility of nested supervision
Human-in-the-Loop Requirements
Section titled “Human-in-the-Loop Requirements”Pro-mandates view:
- Oversight is essential for accountability
- Automation bias requires structural countermeasures (70-80% acceptance without verification)
- Democratic legitimacy requires human decision authority
- Time pressure is a design choice, not a constraint
- EU AI Act mandates oversight with competent, trained persons
Flexibility view:
- Mandatory human oversight may slow beneficial applications
- Not all AI decisions are consequential enough to require oversight
- Transparency alone is insufficient; humans overtrust even when risks communicated
- Skilled AI may outperform human oversight in some domains (30.4% autonomous completion in software development)
- Healthcare professionals face unrealistic expectations to understand algorithmic systems fully
Related Pages
Section titled “Related Pages”Related Responses
Section titled “Related Responses”- Scalable Oversight — Methods for maintaining oversight as AI capabilities grow
- AI Control — Complementary control strategies
- Corrigibility — Making AI systems correctable
- Interpretability Research — Understanding AI decision-making
- Responsible Scaling Policies — Oversight thresholds for deployment
- AI Safety Institutes — Government oversight capacity
Related Risks
Section titled “Related Risks”- Automation Bias — Over-reliance on AI recommendations
- Deceptive Alignment — AI appearing aligned while pursuing other goals
Related Parameters
Section titled “Related Parameters”- Human Agency — Personal autonomy in AI-mediated decisions
- Human Expertise — Expertise required for effective oversight
- Interpretability Coverage — Understanding AI decisions enables better oversight
- Alignment Robustness — Stronger alignment reduces oversight burden
- Societal Trust — Public confidence in AI governance
Sources & Key Research
Section titled “Sources & Key Research”Foundational Research
Section titled “Foundational Research”- Irving et al.: AI Safety via Debate↗ — Original debate proposal
- Christiano et al.: Scalable Agent Alignment via Reward Modeling↗ — Recursive reward modeling framework
- OpenAI: Learning Complex Goals with Iterated Amplification↗
Process Supervision
Section titled “Process Supervision”- OpenAI: Let’s Verify Step by Step↗ — 78.2% vs 72.4% accuracy results
- PRM800K Dataset↗ — Step-level correctness labels
Debate Research
Section titled “Debate Research”- Khan et al.: Training Language Models to Win Debates↗ — +4% judge accuracy
- AI Debate Aids Assessment of Controversial Claims↗
Oversight Frameworks
Section titled “Oversight Frameworks”- Bowman et al.: Measuring Progress on Scalable Oversight↗
- Anthropic: Measuring Progress on Scalable Oversight↗
Automation Bias
Section titled “Automation Bias”- [e6b22bc6e1fad7e9]
- Reuters: Hiring Algorithm Investigation↗
Recent Research (2024-2025)
Section titled “Recent Research (2024-2025)”- Scaling Laws for Scalable Oversight — NeurIPS 2025 spotlight on oversight fragility across capability gaps
- MIT: Fragility of Nested AI Supervision — Quantifies 13.5-51.7% success rates at 400 Elo gaps
- Effective Human Oversight: Signal Detection Perspective — Minds and Machines 2024
- Is Human Oversight to AI Systems Still Possible? — ScienceDirect 2024
- On the Quest for Effectiveness in Human Oversight — ACM FAccT 2024 interdisciplinary perspectives
- Beyond Procedural Compliance: Human Oversight as Distributed Social Capacity — ArXiv 2024
- Anthropic: Recommended Directions for Technical AI Safety — Includes scalable oversight limitations (2025)
- AI Index 2025: State of AI in 10 Charts — Stanford HAI capability benchmarks
- 2025 Global Data Literacy Benchmark — AI oversight deficit crisis
Regulatory Analysis
Section titled “Regulatory Analysis”- EU AI Act Article 14: Human Oversight — Official text and requirements
- EU AI Act Implementation Guide — Comprehensive implementation guidance
- AI Literacy and Human Oversight — EU regulatory framework
Expert Discussions
Section titled “Expert Discussions”- Can There Be Oversight for AI? — Dagstuhl 2025 expert consensus on feasibility
What links here
- Alignment Robustnessparameter
- Lab Behaviormetricmeasures
- Misalignment Potentialrisk-factorcomposed-of
- Deceptive Alignment Decomposition Modelmodelaffects
- Corrigibility Failure Pathwaysmodelaffects
- Expertise Atrophy Progression Modelmodelaffects
- Automation Bias Cascade Modelmodelmodels
- Alignment Robustness Trajectory Modelmodelaffects
- Interpretabilitysafety-agendaincreases
- Scalable Oversightsafety-agendaincreases