AI giants ignore 'orphan risks' in safety frameworks
New research reveals how Anthropic, OpenAI, and others systematically overlook critical AI dangers
Researcher Andrew Maynard's new paper on arXiv examines how leading AI companies systematically prioritize certain risks while ignoring others in their safety frameworks. By comparing published documents from Anthropic, OpenAI, Google DeepMind, and Meta between 2023-2026, Maynard reveals how these institutions define and select risks based on measurability, severity, auditability, and competitive cost.
The study introduces the concept of 'orphan risks'—threats that fall outside these institutional filters—particularly highlighting how less tractable dangers like harmful manipulation are systematically excluded unless legally compelled. Maynard proposes the 'safety differential' to quantify the gap between risks companies choose to address and those regulators deem critical. The paper suggests redefining risk assessment to better capture threats to societal values that current frameworks miss.
- Maynard's analysis of 2023-2026 documents from Anthropic, OpenAI, Google DeepMind and Meta reveals systematic risk prioritization gaps
- Companies focus on quantifiable, measurable risks while excluding 'orphan risks' like harmful manipulation unless legally required
- The 'safety differential' quantifies the mismatch between corporate risk frameworks and regulatory priorities
Why It Matters
Exposes blind spots in AI safety that could allow dangerous capabilities to slip through corporate oversight frameworks