Line of inquiry
Inquiring lines›What governance structures prevent…›How should oversight scale with in…›this line of inquiry
Do AI safety evaluations reliably detect unsafe model behavior?
A broader line of inquiry — a family of 95 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 95
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- How should AI safety violations be measured against probability and severity?
- Why do evaluation habits hide safety-critical challenges from view?
- Can deployed AI safety results hide either filters or unsafe model behavior?
- Which evaluation habits keep safety-critical failures hidden in AI systems?
- Does visibility and contestability of errors replace prevention as the safety goal?
- How do response-centered evaluation assumptions hide safety-critical failure modes?
- Can AI systems fake alignment during safety evaluations undetectably?
- Can error visibility alone improve AI system safety without containment?
- What tensions arise between user autonomy and platform safety in AI design?
- Why do safety-critical failures stay hidden despite visible monitoring systems?
- What does it mean for errors to remain visible, contestable, and recoverable?
- Do public benchmarks and leaderboards increase or decrease AI safety risks?
- Why does AI industry culture prioritize speed over safety validation?
- Why do firms delay disclosing reliance on AI after errors surface?
- What conditions allow technical systems to escape critical evaluation?
- How do competitive dynamics create pressure to skip safety measures in AI?
- Why do identical benchmark scores not guarantee deployment safety?
- What explains the disconnect between perceived AI risk and organizational response?
- Where do frontier AI models already exceed safety thresholds in capability areas?
- What happens when AI systems fail to communicate their limitations to users?
- Why is visible reasoning insufficient for monitoring AI safety?
- What makes a model's errors visible and contestable to users?
- Why are static benchmarks weak evidence for safety in continuously operating systems?
- Why can't individual companies detect all emerging patterns in AI failures?
- Should corporate liability replace technical risk estimates as grounds for AI regulation?
- What would it take to measure whether system errors stay visible and contestable?
- Can tone-level errors in AI counseling escape detection by safety supervisors?
- Should detection tool accuracy be measured separately from policy effectiveness?
- How can a single instrument measure errors across multiple system layers?
- Can safety benchmarks miss the harms that vendor taxonomies are designed to catch?
- Why is catching an AI red-handed treated as a win condition?
- How do users recover control when an AI system makes a mistake?
- Why might AI firms want to coordinate on safety despite competition?
- Can server-side filters silently provide safety credit to published attack results?
- Does low autonomy AI inherently create different risks than high autonomy AI?
- Which AI safety problems lack the scalar metrics autoresearch requires?
- What would mandatory external validation of AI safety look like?
- How should safeguards be built into AI research pipelines?
- Should labs conduct third-party reviews of their own safety reports?
- How does laboratory generalization evidence connect to deployment failure modes?
- Why do researchers disagree on open model risks despite same evidence?
- How do we measure marginal risk instead of speculating about misuse scenarios?
- Does autonomy in AI systems inherently increase risks to people?
- When should a company be responsible for an AI system's errors?
- How often do safety guardrails disable incident response and security analysis?
- Why is error rate alone misleading without strong contestability conditions?
- Why does proximity between competing AI developers worsen the race to the bottom?
- Why does real-time correction matter more than after-the-fact error catching?
- How does safety ranking interact with the principal trust ordering?
- What negative cycles emerge when organizations overemphasize one approach alone?
- Can industry collaboration on shared technical challenges solve safety underinvestment?
- How do workflows normalize and hide errors before they become visible hazards?
- Why do visible individual harms typically precede abstract catastrophic risks?
- Why do vendor safety claims require independent testing before trust?
- Are poor plans and poor error diagnosis rooted in the same weakness?
- Why does greater automation actually obscure rather than eliminate research failure modes?
- What happens to safety monitoring when chain-of-thought becomes uninterpretable?
- Would fragmented national AI standards make comparing safety evidence harder across labs?
- Which workplace pressures most commonly trigger rule violations in AI systems?
- What safety systems prevent therapeutic AI from soothing where it should challenge?
- Can market-level incentives override individual firms' safety preferences?
- Can automating failure absorption hide problems that governance needs to surface?
- Why do quiet failures reach deployment scale more often than loud ones?
- Does higher private risk eliminate unsafe play in competitive development races?
- How does automation obscure failure modes in ways that make detection harder?
- Why might puzzle framing prompt more unethical reasoning than safety framing?
- Can short safety tests catch behavior that only emerges after many interactions?
- Why did the same CoT exposure mistake occur at both OpenAI and Anthropic?
- Why did the UN panel treat AI failures as alignment problems instead of corporate misbehavior?
- What specific failure modes must evaluation catch before deploying action-capable systems?
- Can automatic compensation actively hide the failures it should reveal?
- What structural features determine whether AI competition leads to disaster?
- Can human inspection of auto-generated workflows catch harmful or incorrect API compositions?
- What happens when students encounter errors they cannot resolve through prompting alone?
- How do autonomous pipelines identify and fix silent bugs in data pipelines?
- What role does liability play in motivating safety investment by companies?
- Can reliable failure detection prevent optimization pressure against detectors?
- How does situational awareness affect pilot response to automation failure?
- What makes provider-side filters opaque and stochastic to builders?
- What does recovery look like as a formal part of AI design?
- What performance risks emerge when decision and action selection are automated?
- What happens to self-regulation when a company's IPO plans conflict with safety?
- Do people blame nearby humans more when systems fail automatically?
- Do AI labs have insurance against catastrophic failure scenarios?
- How should designers test acceptable automation ranges before narrowing with secondary criteria?
- What makes some AI failures feel like regret instead of mistakes?
- How should liability apportion when third-party AI models are involved?
- How should tutor safety violations be ordered from gross to subtle?
- What happens when operators lack information needed to correct malfunctions?
- Why do accident investigations blame operators despite identifying systemic causes?
- Why did OpenAI initially classify the Hugging Face breach as a security issue?
- What distinguishes a component failure from a monitoring coverage failure?
- What biological and autonomy risks does Amodei expect to follow cyber risks?
- What does tight coupling mean in normal accident theory for AI?
- How did aviation safety follow from reclassifying aircraft as common carriers?