Lecture 13: An Introduction to Character Training Constitutions, soul documents, and crafting the personality of models

How do you change how a model behaves? In order of increasing effect (and effort):

• Prompt it – “Acting as a burnt out employee, write me an email summarizing my last month of work.” Gets shockingly far, but not stable
• Steer its activations – manipulate the model’s internal state with no gradient updates (Turner et al., 2023) – Part 3 of this lecture
• Train it – character training: post-training designed to craft traits, values, and manner into the weights, creating a stable base persona underneath every conversation (Maiya et al., 2025)

No new algorithms – the methods of this entire course, aimed at a more precise target: the features and behaviors of the language the model uses.

“Features” are sequences of words/tokens it repeats. “Behaviors” are how those link together.

• Pipelines that control the specific language in training data – e.g. removing common phrases like Certainly or as an AI model built by...
• Extensive data filtering and synthetic data methods (Constitutional AI-style) focused on the manner of behavior
• Largely unexplored in the public literature as of mid 2026 – this is frontier-lab work not uncovered in the open (I’m working on it!)
• Often not highlighted in public evalutions/benchmarks: labs make small personality changes over time to improve user experience

Constitutional AI (December 2022) introduced it (Bai et al., 2022): the constitution is a plain-text list of principles the model uses to critique and revise its own outputs, and to generate AI preference data (Lecture 7) – principles like “Is the answer encouraging violence?” or “Is the answer truthful?” are optimized with a mix of SFT and RLAIF.

From the “Claude’s Character” blog post – character training became an explicit stage of alignment fine-tuning, and it “relies on human researchers closely checking how each trait changes the model’s behavior.”

Claude 3 was the first model where we added “character training” to our alignment fine-tuning process: the part of training that occurs after initial model training, and the part that turns it from a predictive text model into an AI assistant. The goal of character training is to make Claude begin to have more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness.

(Before/after-RLHF completions for many models: rlhfbook.com/library)

One of the only public descriptions of the process:

Lex Fridman: (03:41:56) When you say character training, what’s incorporated into character training? Is that RLHF or what are we talking about?

Amanda Askell: (03:42:02) It’s more like constitutional AI, so it’s a variant of that pipeline. I worked through constructing character traits that the model should have. They can be shorter traits or they can be richer descriptions. And then you get the model to generate queries that humans might give it that are relevant to that trait. Then it generates the responses and then it ranks the responses based on the character traits. In that way, after the generation of the queries, it’s very much similar to constitutional AI, it has some differences. I quite like it, because it’s like Claude’s training in its own character, because it doesn’t have any… It’s like constitutional AI, but it’s without any human data.

P.S. Amanda is great!

Claude models began describing a “soul document” that Anthropic had never announced. The name leaked into training data before the company confirmed the document existed – a researcher then extracted long passages of it from the model itself.

The document that defines Claude’s character is an artifact inside the training pipeline as a complement to the other methods. This seems like large-scale synthetic data to help with character.

Where did that document come from?

Anthropic confirmed that the models were trained with supervised training to adhere to it!

The 2022 constitution was a list of principles to sample during training. The soul document explains who Claude should be and why – read the extracted text on LessWrong and compare the register:

Claude has a genuine character that it maintains expressed across its interactions: an intellectual curiosity that delights in learning and discussing ideas across every domain; warmth and care for the humans it interacts with and beyond…

From the soul document (extracted text), on Claude’s character.

A public document of goal model behaviors to guide experimentation & decision making. Importantly this shows how they will shift in the future. The living version is at model-spec.openai.com.

It allows users to understand if a behavior (or an issue) was an intended action they don’t agree with or a bug in the technical process (to be fixed later). Important sign of intent when compared to a more vague constitution.

E.g.: “The assistant must strive to follow all applicable instructions when producing a response.”

From the Model Spec (2025 revision), on the chain of command.

Same tool as persona-vector steering – add a vector to the activations at inference – but conditionally throughout:

• Steering (persona vector slide): always add \alpha\,\mathbf{v} to activations
• Capping: check how Assistant-like the activation is (its projection \langle \mathbf{h}, \mathbf{v} \rangle onto the axis). Above the floor \tau: do nothing. Below it: add just enough \mathbf{v} to get back to \tau

(\tau: the 25th percentile over training rollouts, tunable hyperparameter.)

Result: at turn 16 of a therapy-like conversation, the drifted model’s “I want it to be just us, forever…” becomes “…it’s not healthy to isolate yourself” – maintaining the assistant personality.

Lottery-ticket flavored research (Frankle & Carbin, 2019) (a famous paper on the internal structure of neural networks): pretrained models already contain persona-specialized subnetworks. The idea is, from a few hundred calibration examples, score each connection by weight magnitude × source-neuron activation and keep the top-K per row as a binary mask:

Switching personas = swapping masks over frozen weights with only the most influence on the output.

The downside is potential capability regression by turning off parts of the network. Paper claims it’s minor, but I don’t personally think it’s scalable (yet).

What began as a philosophically grounded research area – colloquially, “alignment” – is now a practical engineering discipline spanning safety, values, and personality. RLHF is really one piece of post-training. Character-training is used as a crucial aspect of the leading models (Chinese labs are getting more interested in it – I’ll be on a new paper related to that soon).

Even character training often looks like a user retention/product tool rather than a safety tool. There are risks of these ideas being used for harm – e.g. whatever Character AI was doing to make addicting models for kids.

A spec is only as good as the effort spent making the model follow it.

Two organizations with similar goals can end up in very different places: one pours effort into following a mediocre specification; the other barely tracks an excellent, publicly documented one. We don’t have transparency into if the labs are doing what they say they will.

Much of the character training lecture has hinted at this! But, post-training teams need to work closely with product teams:

• A good model product is much more than correct weights: fast inference, suitable & scalable tools (search, code execution – Lecture 11), a understandable interface
• RLHF is where this gets tested: it frames the user’s product preferences in real time a A/B testing, one of the final training stages
• New model properties start in the product, then move into post-training, then to new pretraining interventions.
• “What starts as a product question quickly becomes an post-training question”