Written principles guide critique and revision

A model can review an answer against a written principle and propose a revision. That makes the chosen principles influential: different instructions can define different acceptable behavior, and the reviewer model can still misunderstand a case.

The principles and failure cases need evaluation

Do not treat the name as a safety certification. An original assessment should examine the principles, failure examples, and how the system behaves when principles conflict.

What the paper describes

The method has the model critique and revise answers against principles, then uses those comparisons in preference training. The claim is that this reduces the amount of human labeling needed for some harms, relative to their setup.

It is not a claim that the principles are complete, or that the model understands them as rules it cannot break.

A constitution can be read. That is the useful part.

You can see which harms the authors chose to name. You cannot see every behavior that follows.

Two edits to the list are two different training targets. A product that says it uses a constitution owes you the list and the version, or it is using the word as branding.

Where people overread it

They treat a principle as a filter that runs at request time. In the paper, the principles shape training.

A request-time check is a different system, even if it quotes the same sentences. Say which one you mean.

And do not equate “constitutional” with safe. Safety is the rate of the behaviors you measured.

What to record before you rely on it

Link the paper and, if you have it, the constitution version. Do not paraphrase a principle into a stronger promise than the sentence.

Record the model id if you tested a product that cites the method.

Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.

When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.

Name a person who owns that failure.

Common questions

Is this the same as RLHF?

It is a way of producing preference data, using principles and model critiques, which can then be used in a preference-training pipeline. It is not a different species of model.

Can I adopt their constitution as my policy?

You can read it and write your own. Their list is not your organization’s policy unless you adopt it and test it.

A copied list without a test is a poster.

Sources & further reading

  1. Constitutional AI: Harmlessness from AI Feedback ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories