Virtue & Voltage
Culture and practice

AI safety and AI ethics aren’t two conversations

AI safety and AI ethics get staffed, budgeted, and discussed as separate tracks in a lot of organisations, safety handling technical failure modes, ethics handling social and moral ones, drawn as two circles that barely touch. In practice, most of the cases that actually matter live in the overlap, and treating them as separate conversations means neither team is fully equipped to handle the cases!!

A model that hallucinates a plausible-sounding but false medical fact is a safety failure, an accuracy problem, a technical defect. It is also, the moment that output reaches a patient or a clinician acting on it, an ethics failure, a case of real harm landing on the person’s desk… who had no way to know the system was wrong.

Splitting the incident review between two teams/individuals that do not talk to each other regularly means the technical root cause and the human harm get investigated in parallel, by people who each have only half the picture.

The separation persists partly because the two fields grew with different vocabularies and different professional paths, safety closer to machine learning research, ethics closer to governance and philosophy, and partly because it is organisationally convenient to have two smaller mandates rather than one harder, more integrated one.

Neither reason makes the separation accurate to how AI harm actually shows up in the world, which rarely announces itself as purely technical or purely social!

Drawn honestly, safety and ethics are not two circles with a small overlap, they are closer to one shape viewed from two angles:

  1. Safety asking whether the system behaves as intended,
  2. ethics asking whether what it’s intended to do, and who it affects along the way, is actually right.

An organisation that keeps these as genuinely separate conversations will keep missing the incidents that only make sense read as both at once.

Where this connects

This is why the role sits where it does on the org chart in our own model, close enough to both conversations to actually hold them together.

← Previous noteA system built in one place still lands everywhere

© Virtue & Voltage. Quote a passage with a link back; please do not republish this note in full without asking.

← Back to Field notes