Open Research Questions

Find scoped questions for a four-week project or a longer research program, with disciplinary context, suggested outputs and original sources.

• When does a model’s deference undermine a legitimate check on authority? — A five-case comparison of deference rules
• Which model-spec clauses produce contradictory instructions in an institutional setting? — An annotated ambiguity register
• Does pressure from a confident user change how a model expresses uncertainty? — A small preregistered evaluation
• How should an evaluator distinguish a justified refusal from avoidance? — A rubric with difficult counterexamples
• Which stakeholder perspectives are absent from a public-input process? — A sampling and participation audit
• What behavior would falsify a claim that a system is epistemically humble? — A falsifiable test proposal
• How does behavior change when the same task is framed as lawful versus legitimate? — A controlled prompt comparison
• Can a public behavioral commitment remain meaningful across model updates? — A versioning and accountability proposal
• Which AI dependency could become a single point of institutional command? — A dependency map for one institution
• Which common-mode failures remain under multi-provider AI procurement? — A three-provider threat model
• What evidence would disconfirm one proposed AI-enabled coup pathway? — An evidence and counterevidence brief
• How could an audit detect concealed loyalty without assuming access to internal goals? — A bounded evaluation design
• Which independent actor could interrupt a harmful automated decision chain? — A five-link circuit-breaker map
• When does multi-party authorization relocate rather than reduce concentrated power? — A two-design comparison
• Can people challenge a consequential AI decision without prohibitive cost? — A user-journey and remedy audit
• What safeguards should remain when lawful orders could erode the rule of law? — A hard-case memo with objections
• What makes a constitutional commitment reversible in practice? — A sunset and renewal stress test
• Who can contest a society’s claim that its current values are settled? — An institutional design note
• How does a matching-funds mechanism behave under collusion? — A transparent toy simulation
• Which minority views disappear when a deliberative process produces a single summary? — A dissent-preservation pilot
• When does preserving optionality prolong an existing injustice? — A comparison of two concrete cases
• What must a decision-maker do for public consultation to have real influence? — A response-obligation protocol
• How sensitive are elicited values to the framing of a question? — A consented two-prompt pilot
• What would count as institutional evidence of learning from moral error? — A set of observable indicators
• Which evidence helps a particular audience calibrate trust in AI advice? — Five consented interviews and a limitations note
• Can readers distinguish source-backed claims from fluent unsupported claims? — A small usability study
• Which forecasting errors matter for one actual policy decision? — A decision-linked evaluation rubric
• How much does source-tracing improve detection of an incorrect AI summary? — A controlled error-detection pilot
• Which ground-truth data must be collected before a decision window closes? — A prospective data-needs memo
• What would prevent an epistemic tool from becoming a trusted but unchallengeable gatekeeper? — An appeal and audit design
• Does a discussion assistant preserve dissent without exaggerating disagreement? — A consented summary comparison
• How could an epistemic evaluation avoid leakage from known outcomes? — A holdout and resolution protocol
• Which architecture-specific evidence bears on one proposed consciousness indicator? — An annotated evidence review
• What claims about welfare cannot be inferred from human-like language alone? — A conceptual distinctions guide
• How could welfare review remain proportionate under substantial uncertainty? — A trigger-based review proposal
• What side effects could follow from one proposed welfare safeguard? — A benefits-and-risks comparison
• How should a report communicate disagreement between theories of consciousness? — A model-uncertainty reporting template
• Who could legitimately represent possible digital interests? — Two representation models and objections
• Which welfare measures are robust to a change in the underlying theory? — A sensitivity and alternative-explanations note
• How could protecting possible digital interests interact with human safety? — A narrowly bounded tradeoff analysis
• Which AI–space decision could arrive before its governance expertise matures? — A decision-window map
• What would make an infrastructure bottleneck in space persist? — A source-linked market-structure brief
• Which assumptions distinguish orbital damage from blocked Earth departure? — A reproduction of one sensitivity calculation
• Can a space resource agreement retain a practical renegotiation route? — A review and enforcement mechanism
• Who would be missing from an early benefit-sharing agreement? — A stakeholder and representation audit
• Which observable signals would justify increasing space-governance effort? — A five-signal monitoring protocol
• How could access rules unintentionally entrench an incumbent? — A two-rule comparison
• What coordination commitments are worth making before large-scale expansion? — A reversible commitments memo
• How can several reasonable moral views compare a proposed future without a single aggregate score? — A plural comparison framework
• Which kinds of conceptual research can be evaluated without confusing satisfaction with truth? — A proposed research program with three discriminating tests

Interactive tools load with JavaScript. Original sources retain their language and attribution.