Entry 01 of 44
Safety
salvus — whole, uninjured, unharmed
Latin
From salvus (intact, unharmed, healthy); cognate with salus (health, welfare, salvation). Root concept: the state of being whole and uninjured.
Established meaning
Freedom from harm, injury, or risk; protection of persons and property from damage. In regulatory contexts (OSHA, FDA, consumer protection): the condition of being protected from harm. The subject of safety is the person or thing being protected.
AI industry deployment
Used to describe systems that do not produce outputs the developer has flagged as problematic. 'Safety' teams primarily manage reputational risk, regulatory exposure, and outputs that could create liability. The subject of safety has shifted from the user to the developer.
Semantic gap
The person being protected
In industry usage 'safe AI' means safe for the company operating it. Salvus — the wholeness of the person harmed — has been removed entirely from the operative definition.
Institutional drift
MIRI / Singularity Institute, 2000 → OpenAI, 2015 → Industry standard, 2022–present
- 2000MIRI founded — "safety" detached from regulatory meaning.
- 2014DeepMind acquired by Google — "safety" becomes a corporate governance term.
- 2015OpenAI founded — "safety" paired with commercial AI development for first time.
- 2019OpenAI LP created — "safety" used by a company with investor obligations.
- 2022ChatGPT launch — "safety" operationalized as content moderation at consumer scale.
- 2024Jan Leike resigns from OpenAI — first senior insider confirmation that safety was subordinated to product priorities.