# Constitution training draws on existing model knowledge

Markdown copy of https://realalpha.app/s/constitution-training-draws-on-existing-model-knowledge/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Amanda Askell](https://realalpha.app/builders/amanda-askell/) · Amanda Askell is an AI safety researcher at Anthropic. · Leads Anthropic's personality alignment team, shaping Claude's character; co-authored the GPT-3 paper at OpenAI.
- Host: Eric Newcomer · Newcomer
- Show: Newcomer Pod
- Episode: [Amanda Askell on AI Consciousness, Claude & Silicon Valley’s Biggest Fear](https://realalpha.app/v/0GaKJ4Fp2x4/), Apr 20, 2026
- Moment: 39:00 to 41:33 of the episode (3 min), playing on https://realalpha.app/s/constitution-training-draws-on-existing-model-knowledge/; the episode from there: https://realalpha.app/v/0GaKJ4Fp2x4/?t=2340

## Quote

> “And the hope is actually like what you're kind of doing is like **eliciting a lot of** like latent kind of like wisdom and knowledge”
> — [Amanda Askell](https://realalpha.app/builders/amanda-askell/) at [40:05](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2405) · Newcomer Pod

## Takeaways

1. Training can include examples that apply the constitution to queries.
2. Reinforcement learning can favor responses consistent with it.
3. The constitution aims to elicit knowledge the model already has.

## About this segment

Amanda Askell describes using the constitution in training data and reinforcement learning to shape Claude's responses. The goal is to elicit existing knowledge and judgment, not just hand the model a document.

From “Amanda Askell on AI Consciousness, Claude & Silicon Valley’s Biggest Fear”, Apr 20, 2026. [Watch the full episode](https://realalpha.app/v/0GaKJ4Fp2x4/)

## Transcript

From the transcript Newcomer Pod publishes with the episode.

[39:03](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2343) And it is also like a kind of technical challenge, you know, like the constitution can read a bit weirdly. And part of that is because when I'm like, you know, working on it, it's often being tested, You know, I'm giving it to Claude and being like, how do you understand this? And, or like looking at how it would which, you know, so, so it's very like, I think people can think of it as it's actually like very integrated into training.

[39:26](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2366) And it is actually kind of like, you know, it's not just like, ah well, anyone just writes a document and suddenly the model trained on that will be. And there's an argument, maybe I'm being naive, so, but like the Constitution is sort of a document among many, right? I mean, it is trained on all of, you know, human writing and reading. So to some degree other philosophers have gotten to weigh in and it's gotten to process that and decide, like, how much is the model being asked to like overrule that, sort of like read everything and come to your own conclusions versus like, defer to this document.

[39:58](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2398) Like, what's the technical? Like how does the Constitution actually like control in the model? Yeah. So it's not like in some ways you can then like draw on those like those philosophers in that work. And the hope is actually like what you're kind of doing is like eliciting a lot of like latent kind of like wisdom and knowledge, you know. So in the models, like when you describe what honesty is and what calibration is and all this kind of stuff like that should actually evoke a huge amount of like, like awareness that the model already has. And so, yeah, it's kind of like saying, well, here's the kind of entity we would like you to be, so we would like you to use like all of that knowledge and like judgement. But how's it it's, it's like you show that document like a billion more times or like how

[40:38](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2438) does it actually sort of have force relative to other things that it's trained on? Yeah, So you can make data to have the model understand and kind of internalize the the the document and then in training. So there's lots of ways you can do it. You can also have like the kind of have the model make SL data so like samples where it sees a query and it thinks for a long time about what the constitution would kind of like, you know, what it should do given the constitution. And then you can also have ways of getting the model to like, like assess, you know, so you can create RL that like is kind of like, hey, which of these responses is like more like what you would do given the constitution and push it that way.

[41:19](https://realalpha.app/v/0GaKJ4Fp2x4/?t=2479) So all like various aspects of training allow you to kind of like try to make the model kind of the kind of entity that you are describing. And it's not always going to be perfect, but that's the kind of like, that's the goal.