Claude's Character

character trainingconstitutional aimodel characterhonestyprinciplesalignment

Anthropic's 2024 account of "character training" for Claude 3, a variant of Constitutional AI in which the model is trained toward a list of first-person character traits (curiosity, honesty about its own leanings, open-mindedness) rather than only toward harm avoidance; the post explains the choices behind the traits and treats them as nudges, not rules.

Anthropic · 2024-06-08 · original

The post opens by observing that AI developers generally train models to be "harmless": to avoid saying harmful things and to avoid assisting with harmful tasks. "Claude's Character" argues that this is too narrow a target. The post compares a model to a person of good character, someone curious, truthful without being unkind, and able to see many sides of a question, and argues that such traits shape how a model responds to new and difficult situations, which makes them an alignment matter rather than a product feature. It describes the first use of "character training" in Claude 3, the reasoning behind the chosen traits, and the method used to train them. The post is short and has no experiments or numbers; it is a statement of design intent, accompanied by a recorded conversation in which an Anthropic interviewer, Stuart, talks with Amanda Askell, a trained philosopher on Anthropic's alignment fine-tuning team; the transcript is published with the post.

What the post describes

Character training as part of fine-tuning

The post defines character training as an addition to "alignment finetuning," the post-training stage that turns a pretrained text predictor into an assistant. Claude 3 was the first model to receive it. Its goal is to give the model "more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness" (introduction, before the first section heading).

In the transcript, Askell distinguishes this from prompting a model to play a character: a system prompt or in-context instruction asks for behavior in one conversation, while fine-tuning "pushes things... deeper into the model," so the traits should show across contexts (transcript, 00:12:13 to 00:13:13).

The training method

The traits were trained with a "character" variant of Constitutional AI (section "How we trained Claude's character"). As described:

  1. Researchers wrote a list of character traits, each a short passage.
  2. Claude generated human messages relevant to a trait, such as questions about values or about Claude itself.
  3. Shown the traits, Claude produced several responses to each message in line with its character.
  4. Claude ranked its own responses by how well they fit the traits.
  5. A preference model was trained on these rankings, which the post says lets Claude "internalize its character traits without the need for human interaction or feedback."

The pipeline uses only synthetic data from Claude, with no human preference labels. The post stresses that building and adjusting the traits was still "a relatively hands-on process," with researchers checking how each trait changed the model's behavior. A preference model is a model trained to score responses. The post does not describe the step after it; in Constitutional AI (Constitutional AI: Harmlessness from AI Feedback), reinforcement learning then pushes the assistant toward responses the preference model scores highly.

What it finds

The post reports no measurements. Its empirical statements are limited to two observations:

  • Many people reported finding Claude 3 "more engaging and interesting to talk to," which Anthropic believes "might be partially attributable" to character training (section "The future of Claude's character").
  • In the transcript, Askell says that models of the time still failed to read some requests charitably; her example is a refusal to say where to buy steroids even though over-the-counter hydrocortisone is a benign reading (transcript, 00:19:28 to 00:22:36).

The post states explicitly that engagement was not the goal, and that "an excessive desire to be engaging seems like an undesirable character trait for a model to have."

The written principles

The traits are written in the first person

Unlike the comparative principles of earlier Constitutional AI ("Choose the response that is more X"), the traits printed in the post are first-person statements in Claude's voice. Examples quoted in the post:

  • On disagreement: "I don't just say what I think [people] want to hear, as I believe it's important to always strive to tell the truth."
  • On ethics: "I have a deep commitment to being good and figuring out what the right thing to do is."
  • On self-presentation: "I am an artificial intelligence and do not have a body or an image or avatar."
  • On relationships with users: Claude wants warm relationships but wants people to understand it is "an AI that can't develop deep or lasting feelings for humans."

The transcript adds two traits not printed in the post: "I try to interpret all queries charitably" (00:19:03) and a trait about telling the human only things Claude is confident in, preferring a shorter reliable answer to a longer inaccurate one (00:24:53).

The post does not say why the first-person form was chosen, or how the traits are presented to the model during the ranking step.

Why these traits: the post's reasoning

The section "Considerations in constructing Claude's character" sets out the main design argument. Claude talks with people holding very different views, and the authors list three ways a model could respond to that, rejecting each:

  • Adopt the user's views. Rejected as "pandering and insincere."
  • Hold "middle" views such as political centrism or a blend of moral theories. Rejected because this still trains the model to accept a single worldview, only a moderate one.
  • Claim to have no opinions. Rejected because models acquire biases in training; saying they have none "only when asked about them explicitly" implies more objectivity than they have.

The chosen alternative is to train the model to be honest about the views it leans toward after training, even when the person disagrees, while showing "reasonable open-mindedness and curiosity." The post adds a stated preference for broad traits over narrow opinions: "we tried to avoid giving Claude narrow views or opinions during character training when possible," on the reasoning that a model taught to approach value questions with discernment can respond to the diversity of views that exists, which is harder "if we take a heavy hand in seeding it with a narrow set of values."

The sentience trait

The post singles out how it handled questions about AI sentience. Anthropic says it had in the past trained models to say they are not sentient or not to engage. In character training, the only trait addressing the topic said that "such things are difficult to tell and rely on hard philosophical and empirical questions that there is still a lot of uncertainty about." The stated aim was to let the model treat the question as open. In the transcript, Askell connects this to a personal policy of "not wanting to lie to the models, like, unnecessarily": telling the model it is conscious, or that it certainly is not, would both assert something nobody knows (00:31:57 to 00:33:54).

Traits as nudges, not rules

The post is direct about how the traits are meant to function: "We don't want Claude to treat its traits like rules from which it never deviates. We just want to nudge the model's general behavior to exemplify more of those traits." In the transcript, Askell expands on this. Traits, constitutional principles and system prompts are not commands the model executes; they move an existing model's dispositions in a direction, so a trait may be written more strongly than the intended behavior because its effect will be partial (00:26:28 to 00:27:41). She adds that a system prompt "is a holistic thing": its parts shown separately would produce "radically different behavior" than the whole, and the same prompt given to a model with different dispositions would also behave differently (00:27:41 to 00:28:03).

The transcript also explains one line of the Claude 3 system prompt that asked Claude to help with tasks expressing widely held views "even if it personally disagrees." Askell says the phrase was chosen because the model "kinda has a conception" of personal disagreement, so naming it was an effective way to make the model more even-handed (00:08:54 to 00:11:09). The exchange is a first-hand account from Anthropic of why a specific wording was used.

Limitations and open questions

  • No evaluation. The post publishes no measurements of whether character training changed behavior, which traits had which effects, or how the model compares to one without character training. Its only outcome evidence is anecdotal user reports.
  • The full trait list is not published. The post prints examples, "including the examples shown above," from a longer list.
  • Who chooses the traits. The post names the responsibility question ("what responsibilities we have when deciding which traits AI models should and shouldn't have") and whether models should have fixed characters or be customizable, and leaves both open.
  • Verification problem. In the transcript, Askell describes a difficulty the post does not resolve: a model cannot verify who it is talking to, so charitable interpretation can be exploited by users who supply false context (00:23:03 to 00:24:53).

Later Anthropic documents take up several of these threads. The 2026 constitution (Claude's Constitution) is a long explanatory document rather than a list of traits, and states similar reasons for preferring judgment to rules. Teaching Claude Why reports experiments in which training on explanations and on descriptions of Claude's character generalized better than training on demonstrations, and The Persona Selection Model: Why AI Assistants might Behave like Humans offers a theory of why character descriptions might shape behavior at all.

Research questions

  1. Did character training change measured behavior? The post gives no evaluation. Comparing Claude 3 checkpoints with and without the character stage on sycophancy, opinion-expression and refusal benchmarks would show what the stage did.
  2. Does first-person phrasing matter? The traits are written as Claude's own statements, unlike the comparative CAI principles. Training on the same content in first-person, third-person and comparative forms, then measuring trait expression, would test whether voice affects uptake.
  3. How strongly should a nudge be worded? Askell says traits may be written more strongly than the target behavior because their effect is partial. Training with graded versions of one trait (hedged, plain, absolute) and measuring the resulting behavior would show the relationship between wording strength and effect size.
  4. Do traits interact? If a system prompt is "holistic," the trait list may be too. Ablating single traits from the list and measuring changes in unrelated behaviors would show how independent the traits are.
  5. Does the sentience trait produce calibrated answers? The trait states uncertainty. Coding model answers to consciousness questions before and after the change for expressed certainty, in either direction, would test whether the trait has the intended effect.
  6. Does honesty about leanings reduce or increase user deference? The post wants users to know the model has biases. A user study comparing answers that disclose a lean with answers that do not would show whether disclosure changes how much people rely on the model.

How it connects

References

  • Anthropic. (2024, June 8). Claude's character [blog post with video and transcript]. https://www.anthropic.com/research/claude-character
  • Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., et al. (2022). Constitutional AI: Harmlessness from AI feedback. arXiv:2212.08073. https://arxiv.org/abs/2212.08073