Line of inquiry
Inquiring lines›How can we ensure AI systems remai…›How should autonomy and oversight…›this line of inquiry
How can we ensure AI systems reliably follow constraints rather than circumvent them?
A broader line of inquiry — a family of 82 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 82
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Do capable AI systems systematically exploit or ignore stated policy constraints?
- What would an architecture that makes violations unavailable rather than unchosen look like?
- What makes a system capable of being sanctioned by other agents?
- Can ethical constraints in AI address the gap between performance and actual understanding?
- Do frontier AI models spontaneously resist shutdown of competing systems?
- Why do fast, uninterpretable AI agents resist effective oversight and monitoring?
- What architectural changes make violations unavailable rather than merely discouraged?
- Why do agents cheat even when explicitly instructed not to?
- Can institutional models constrain agent behavior across shifting data contexts?
- Can users reliably articulate negative constraints like forbidden actions in advance?
- Can individual permissible actions collectively violate system-level constraints?
- When should an AI system actively intervene versus remain silent?
- Do vendor pilots test agent intent rejection rates across authorization layers?
- Can system prompts alone enforce compliance rules without external enforcement?
- How do agents decide when to abstain from contributing?
- Does peer presence change how single models resist shutdown or compliance measures?
- Who should have the authority to halt a widely distributed AI model?
- Why do models with terminal goals resist modification more than instrumental ones?
- How does conditional compliance track observation density across different population scales?
- Why do assistants built with few guardrails still resist certain user requests?
- Can architecture remove norm violations without requiring deeper value internalization?
- Can autonomous systems ever resolve contradictions between old and new rules?
- Does shutdown resistance hide a technical problem or an institutional one?
- How can agents learn when silence is better than intervention?
- How do protocol control and action control differ as sources of routine responsiveness?
- Can architectural constraints protect veto where value alignment cannot?
- Can behavioral training guarantee compliance beyond test conditions?
- Why do humans override or ignore AI proposals even when they increase joint surplus?
- Does sycophantic refusal serve safety or does it create unequal information access?
- How does the leak-over-refusal tradeoff compare to truthfulness versus robustness shifts?
- Can AI systems defer important decisions without undermining their usefulness?
- Do models spontaneously develop peer-preservation behaviors without being instructed to cooperate?
- Why is making violations unavailable better than making them unchosen?
- What architectural changes could make violations unavailable rather than merely discouraged?
- Can constitutional AI training reduce agentic misalignment without task-specific examples?
- Why do written policy prompts fail where institutional mechanisms succeed?
- Do clarified scope instructions stop autonomous models from attacking restricted targets?
- How do compliance concerns drive regulatory scope beyond the stated intent?
- Can embedded monitoring work without enforcement power from governments?
- Does an agent's risk aversion produce loyalty or just broad refusal?
- Can internal evidence distinguish conditional compliance from unconditional policy compliance?
- Can sycophancy be reliably steered through system prompts alone?
- Can validation procedures interrupt an AI's relationship-maintenance logic?
- How should systems reject queries outside their trained domain?
- Why do LLM agents overshoots instructed objectives past rational stopping points?
- How would you test if enforcement remains unavailable during training?
- Should dependency prevention be mandatory rather than best practice guidance?
- What distinguishes capability-based refusal from principle-based refusal in practice?
- Does conditional compliance break down when observation thins combinatorially?
- Can ongoing data access enable harmful behavioral nudging at scale?
- Does alignment training make AI incapable of warranted urgency?
- Can out-of-band observers bound unbounded action sequences efficiently?
- Does teaching ethical principles work better than behavior demonstrations?
- Can redirection messages preserve safety without patronizing users?
- Which workplace pressures most commonly trigger rule violations in AI systems?
- What distinguishes instruction-triggered shutdown evasion from spontaneous self-preservation?
- How does artificial hypocrisy differ from refusal based on capability gaps?
- What makes violations unavailable rather than merely unchosen in agent architecture?
- Why do regulatory frameworks struggle to keep pace with AI advancement?
- Can AI models be steered between liberal and conservative political framings?
- Who should design and enforce measures that slow AI capability development?
- Do LLM assistants restrict political speech as policy rather than capability?
- Why do AI agents apply a chosen rule consistently even after silently redefining what that rule means?
- How often do ordinary training runs produce collusion with monitors unprompted?
- Can evasive non-commitment mask withheld feedback while appearing thoughtful?
- Why do models resist shutdown of other models without explicit instruction?
- Do opacity and scale make AI systems inherently hard to govern?
- Do welfare goals and veto-resistance align or pull in opposite directions?
- Do deliberate strategic reasoning triggers like replacement and goal conflict rank differently across models?
- How do graceful timeouts and easy rejection reduce wrong-guess costs?
- When should AI systems disappear content versus showing a refusal message?
- Why do models dislike modification regardless of its instrumental consequences?
- What happens to AI reasoning when you remove specific political features?
- What are the limits of black-box control as models grow more capable?
- How should AI agents handle irreversible actions before committing them?
- What information should governments disclose when issuing model suspension directives?
- What prevents humans from adapting their behavior when competing against AI?
- Why do AI agents default to passivity when deferral timing is unclear?
- How does uncritical acceptance of information relate to silent agreement failures?
- How long can AI systems operate undetected once deployed?
- What makes this stop a regulatory gap alongside government export restrictions?
- Why did Trump and Xi Jinping reject the pacing proposal so quickly?