Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses
2026-05-29T07:23:51Z•c39c58b93316c8147b8707562f1d4ac045d3fb439385d4f69c9c45c28878b939
AI governanceSTEM reasoningbiosecuritycompute governancedatasetdistributed trainingdual-useeducation AIhuman evaluationlanguage model agentslegal techobservabilitypolitical neutralityprivacypro se litigationprotestsreinforcement learningreputation systemssocial intelligencesurveillance
What happened
Collection of 10 recent arXiv papers highlighting governance, safety, and security risks from current AI and related sociotechnical trends. Key contributions include: (1) PARETO — a large-scale human-evaluation benchmark and definition for AI political neutrality showing model defaults often lean liberal and introducing a 7,434-participant dataset; (2) an analysis showing distributed training algorithms could enable evasion of compute-governance regimes and recommending countermeasures (whistleblowing, chip tracking, forensic accounting, memory/compute cluster thresholds); (3) evidence that広en
Why it matters
A reviewed impact interpretation has not been published for this record.
Evidence and limitations
- Source ID
- arxiv_cs_cy
- Record identifier
- c39c58b93316c8147b8707562f1d4ac045d3fb439385d4f69c9c45c28878b939
- Enrichment time
- 2026-05-29T07:23:51Z
- AI-assisted enrichment
- Yes
This record may overlap with other records. Its enrichment can be incomplete or wrong, and machine assistance was used. Validate consequential decisions against the linked source and your own environment.