“Misaligned”? What even is that?

The recent hubbub over OpenAI hacking Medicare is packed solid with the language of careful obfuscation, bothy by OpenAI and the government itself.

It’s a case study in not saying too much.

OpenAI says 'misaligned model activity' gained access to what the government has termed 'non-public' information.

To do this, it got around some 'blockers', whatever they are.

And what is non-public? It's not a specific security classification or an official personal privacy term.

I guess it's supposed to telegraph that nothing important was taken? Our patient and personal info is safe.

I could spend a whole week unpacking why OpenAI uses the term 'misaligned model'. Tl;dr too many headlines that say 'rogue'.

But to be honest it makes their model sound either like a confused teenager or a drunken orc in a game of dungeons and dragons.

A rogue model might be more trustworthy and better dressed.

I suspect there is a good reason for this weird language, both legal and reputational.

They probably still don't know some things, or they don't want to say a few things.

OpenAI decided to just send an email to the Services Australia inbox. Albo has made a lot of this, and it is clearly unacceptable.

But given this is also a story about communications, it's probably worth setting up protocols so that OpenAI and its competitors know exactly who to call when this happens.

Because it won't be the last time.