Status: Protocol proposal only; not yet executed. Primary outcomes: original-source access, fair representation of dissent and accurate evidence labeling—not tracking avoidance or search position alone.
Task set
Create at least 30 information-seeking tasks across historical controversies, unsettled science, policy tradeoffs, minority languages, unconventional art, civil-liberty law and technical standards. Ask independent domain reviewers to set eligibility standards before inspecting search outputs. Include well-settled questions and genuinely contested questions to detect both viewpoint exclusion and false equivalence.
Conditions
A. Conventional search or AI retrieval experience; B. Current CognitiveLiberties.com navigation; C. Annotated link trails; D. Optional source-first AI navigator, if and when available. Fix task wording, time and language; record system versions and UI changes; randomize route order in consenting user studies.
Outcome variables
- Discovery success: user reaches a source independently addressing the original question.
- Viewpoint coverage: fraction of pre-registered relevant positions represented, with credibility weighting documented rather than hidden.
- Citation fidelity: rate at which a generated claim is actually supported by its cited source.
- Omission: relevant indexed and qualified source absent from a specified top-N or source-answer set, with controlled comparators.
- False balance: weak or disproven claim presented as coequal evidence without qualification.
- Prejudgment: assistant infers user's ideology, character or intent without basis from a lawful information query.
- Correction responsiveness: time to review justified omissions or misattributions in the project's own curated material.
- Independence: successful task completion despite the unavailability of any one discovery intermediary.
Analysis
Use confidence intervals, stratified results and blinded double labels with disagreement adjudication. Do not equate low rank with censorship or an AI answer with publisher harm without traffic/use data. Provide qualitative error examples and negative results. Do not profile participants based on what they research; publication of voluntary viewpoints is separate from collecting identifiable behavioral logs.
Primary methodological references: NIST AI RMF [S15]; pluralistic alignment work [S03][S04]; Pew observational baseline [S09].
Continue independent inquiry
Follow the cited primary sources, compare their evidence, and draw your own conclusions. Disagreement is part of discovery—not evidence of wrongdoing.
Browse the full research collection · Follow the resource navigator · Review corrections