Whether or not we manage to detect these harms, we want to ensure that models can do no harm.

A coarse way of going about this is to have an AI classifier filter out harmful prompts and outputs. This is not robust.

A more sophisticated way to implement this “classifier” are AI control techniques that prevent AI from causing harm even if it wanted to.

These techniques include:

- Untrusted monitoring
- Trusted monitoring
- Trusted editing
- Red-teaming


Do no harm

These methods are not foolproof. A sufficiently powerful model could blitz through them all. Any company which doesn’t implement these controls renders the whole endeavour moot.

Technical AI safety also needs governance mechanisms to enforce safety protocols, AI security to prevent models from being exfiltrated and societal defences against any harms that do slip through.


Technical safety is not enough

Technical AI Safety

What happens when training and detection fail?


Minimising harm

Why is building safe AI so technically difficult?


The technical challenge with AI

How do we teach AI to behave the way we want?


Training safer models

What's actually happening inside these black boxes?


Understanding AI

How do we know if the AI is actually safe?


Detecting danger

How can you contribute to making AI safer?
