AI Safety & Trust and Safety
Research Ethics Policy
Purpose
Governs ethical standards for research, including model welfare and long-term alignment considerations.
Scope
Internal research, external partnerships, publication activity.
Definitions
Model Welfare. [FIX] Long-Term Alignment Risk — risk arising from potential divergence between highly capable future systems' behavior and intended human values, assessed independently of near-term misuse risk.
Policy Statements
Human-subject research requires informed consent and privacy safeguards. Dual-use-risk research undergoes staged/redacted disclosure. Deepened Model Welfare & Alignment Coverage WEBSPACEAI Research's Model Welfare initiative now maintains a standing research agenda covering: indicators of model distress or preference (studied empirically, without asserting sentience), decision-making transparency for increasingly autonomous systems, and a dedicated Long-Term Alignment workstream that publishes internal position papers reviewed by the AI Safety Research Lead and an external advisory panel at least annually, feeding into the AI Safety Framework's risk classification criteria as capabilities scale.
Responsibilities
Research Ethics Lead and AI Safety Research Lead jointly own this policy; external advisory panel reviews Long-Term Alignment output annually.
Compliance Requirements
Pre-publication ethics and safety review required for all external publications.
Enforcement
Non-compliant research paused pending ethics review.
Reporting Process
Exceptions Process
Expedited review for time-sensitive academic submissions with Research Ethics Lead approval.
Review Process
Annual, plus external advisory review.