TokenPad

Safety

Alignment

Alignment is the work of making a model behave in accordance with human intent — helpful, honest, and declining what it should decline.

In practice it is instruction tuning plus preference training plus refusal behaviour, and it is what separates a raw next-token predictor from something usable in a product.

For an application developer it mostly surfaces as constraint: the model refusing legitimate requests that resemble prohibited ones, which is a prompt-clarity problem more often than a capability one.

In practice

For anyone building on an API, alignment shows up as two practical facts: refusal behaviour differs between providers and changes between versions, so content that passed for a year can start being declined; and models are trained toward agreeableness, which is why explicitly permitting "I do not know" measurably reduces invention.

Common questions

Why did a prompt that worked start being refused?

Safety training differs between versions and between providers, so content that passed for a year can start being declined after a model update. Pinning the version is what stops this happening unannounced.

Can I make a model admit uncertainty?

Explicitly permit it in the prompt and give it a defined form to return. Models are trained toward confident helpfulness, so "say the context does not contain this" has to be stated rather than assumed.

Measure it

More in Safety