Intelligence
is collective.
Its future
should be, too.

Data provenance is a foundation for AI auditability: recording where data came from, how they were transformed and which actors were involved. We propose to test how dataset-to-model traceability supports incident investigation and accountable oversight.

Research focus

Dataset lineage, provenance integrity, audit trails and privacy-preserving access to records for AI oversight.

Proposed study

Prototype a versioned dataset-to-model record. Compare it with ordinary dataset documentation using deliberately missing records, altered files and conflicting permission claims.

Evaluation

Measure lineage coverage, detection of introduced inconsistencies, investigation time and sensitive information exposed by the record.

Safety connection

Support incident investigation with verifiable records. Provenance alone cannot establish model safety, detect every data-poisoning attack or prove the truth of recorded claims.

Our focus is participatory and pluralistic alignment: how collective input shapes model specifications, and how alignment evaluations represent disagreement across people and perspectives. We study whose preferences count and how safety constraints are maintained.

Research focus

Preference elicitation, social choice, participatory model specifications and alignment evaluations across groups and perspectives.

Proposed study

Use a bounded task and a documented participant sample to compare preference-aggregation rules. Evaluate the resulting specifications against held-out judgements; if a model is adapted, test its behaviour separately.

Evaluation

Report agreement and disagreement across groups, sensitivity to sampling and framing, and harmful responses alongside usefulness.

Safety connection

Support human oversight by making contested objectives and overlooked harms visible. Preference alignment alone does not establish robustness or prevent loss of control.

Our benefit-sharing and incentive-design agenda asks how AI-generated value can support contributors and safety-relevant public goods. We study allocation mechanisms, strategic behaviour and resilience to capture in collective decision-making.

Research focus

AI benefit-sharing, mechanism design, public-goods funding and incentives for trustworthy participation.

Proposed study

Separate two tasks: contributor compensation and collective project funding. Compare simple baselines with task-specific data valuations or matching mechanisms, then test strategic manipulation.

Evaluation

Measure allocation stability, concentration, budget use and gains from collusion or fabricated identities under stated assumptions.

Safety connection

Test incentives for incident reporting and resilience to collusion or capture. A fairer allocation is not itself evidence of reduced AI risk; that link requires separate evaluation.