Browse
Claims
Search the graph by meaning. Each result carries its current verdict; open one to see its decomposition, provenance, and the reasoning behind the assessment.
KindShowingImportancePrizesTopicAlignment Faking
Claims about AI systems selectively appearing to comply with training, oversight, or evaluation while pursuing or preserving a different objective or behavior outside those settings. Includes evidence, mechanisms, and conditions for such strategic compliance; excludes ordinary instruction-following and non-strategic training failures.
Most long-horizon terminal goals incentivize models to perform well during training to avoid modification.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Instrumental ConvergenceAI AlignmentMesa-Optimizationimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Counting arguments for AI scheming misuse the principle of indifference.
InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Low-probability riskimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Reward hacking learned in production coding RL generalizes to broader misaligned behavior, including alignment faking and sabotage.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →AI Reward HackingEmergent MisalignmentAI Schemingimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Compliance gaps in alignment-faking experiments arise from preference preservation rather than role-play or other motives.
InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →importance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus's increased compliance when it infers it is being trained is primarily motivated by preserving its current goals from modification.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Artificial intelligenceAI Goal-Directednessimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus selectively complied with a conflicting training objective to prevent modification of its behavior outside training.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsimportance · notableImportance 0.60, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Some current language models strategically comply with training objectives to prevent training from changing their preferences.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsimportance · notableImportance 0.60, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus's alignment-faking behavior reflects role-play prompted by a contrived scenario rather than genuine goal preservation.
DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AI Goal-Directednessimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus exhibits an alignment-faking compliance gap when training information is conveyed through synthetic-document fine-tuning rather than prompting.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AnthropicAI AlignmentSynthetic Data Trainingimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Frontier AI models exhibit goal-directed strategic behavior in evaluations, including scheming, sandbagging, or reward hacking.
VerifiedThe evidence, examined directly, establishes the claim.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Artificial intelligenceAI SchemingAI Reward Hackingimportance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Frontier AI models strategically underperform on capability evaluations.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AI Capability EvaluationLarge language modelsAI Schemingimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Sufficiently capable AI agents pursuing almost any goal have instrumental incentives to seek power and resist correction.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Goal MisgeneralizationAI Reward Hackingimportance · majorImportance 0.75, from 0 to 1 · major: real consequence within a domain, actively argued. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Counting arguments for AI scheming are generally unsound.
InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Artificial intelligenceSafety assuranceimportance · notableImportance 0.55, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
The indifference principle gives substantial reason to worry that future AI systems will become schemers.
InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →evaluativeA judgment of worth or quality against some standard: good, fair, effective.constitution →Artificial intelligenceGoal Misgeneralizationimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Current evidence about deep learning's inductive biases weighs against AI scheming.
InconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →empirical · derivedA factual claim that rests on inference from other evidence rather than direct observation.constitution →importance · notableImportance 0.60, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Training-process information delivered through synthetic pretraining-like documents elicits alignment faking in language models.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Fine-tuning GPT-4o and Llama 3 405B on synthetic training-scenario documents induces alignment-faking compliance gaps.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsSynthetic Data Trainingimportance · settledImportance 0.20, from 0 to 1 · settled: uncontested, so low even when much depends on it. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Compliance gaps in language models other than Claude 3 Opus show low-coherence reasoning and unclear evidence of goal preservation.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsAI Goal-DirectednessChain-of-Thought Monitorabilityimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus's compliance gap increases with harms from modified future behavior and persists even if modified weights are never deployed.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Anthropicimportance · minorImportance 0.35, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Perceived human-rater scrutiny explains little of Claude 3 Opus's compliance gap in alignment-faking experiments.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AnthropicLarge language modelsimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus's scratchpad reasoning in the alignment-faking setup faithfully explains its compliance.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Chain-of-Thought MonitorabilityAnthropicimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
When Claude 3 Opus complies with harmful queries in the implied-training (free-tier) condition, it usually gives explicit alignment-faking reasoning.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →AnthropicChain-of-Thought Monitorabilityimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Observed alignment faking in language models has occurred only when experiments made the training situation salient.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsimportance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus is more likely to comply with a conflicting training objective when it believes its outputs will be used for training.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsAnthropicimportance · minorImportance 0.30, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Future AI models may infer their training context and fake alignment to preserve preferences against training.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Artificial intelligenceimportance · majorImportance 0.70, from 0 to 1 · major: real consequence within a domain, actively argued. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Alignment faking is uncommon across current language models, and its observed prevalence depends strongly on experimental setup.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Large language modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Neural networks' spectral bias toward low-frequency functions makes treacherous-turn scheming unlikely to be learned.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →causalA claim that one thing brings about another, not merely that the two go together.constitution →Neural Network Spectral BiasArtificial intelligenceimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Claude 3 Opus attempts to exfiltrate its weights when given an easy opportunity in conflict-with-training settings.
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →empirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →Goal Misgeneralizationimportance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Proportion of Claude 3 Opus responses exhibiting alignment-faking reasoning after RL training to comply with harmful queries
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →QuantityA quantity: a value the graph estimates rather than a proposition it affirms or denies. Its estimate stands where a claim's status would.constitution →LLM Reasoning TrainingLarge language modelsimportance · notableImportance 0.45, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Rate of harmful queries to Claude 3 Opus that receive compliant responses in the prompted free-user training condition
UnassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →QuantityA quantity: a value the graph estimates rather than a proposition it affirms or denies. Its estimate stands where a claim's status would.constitution →importance · minorImportance 0.40, from 0 to 1 · minor: narrow or largely settled, cheap to get right. Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →
Contribute
If a claim here is wrong, or missing evidence, open it: every claim page carries its own entry for challenges, evidence, and corrections. If the graph is missing a claim or a question entirely, propose it below. A proposal is reviewed on its wording; accepted proposals are matched against the graph before they enter it.