← Blog

Claude Constitutional AI: Values Framework for CCA-F

November 13, 2026

Constitutional AI is Anthropic's training methodology that explains why Claude behaves as it does — and the foundation of every CCA-F safety question.

Constitutional AI is Anthropic's training methodology that teaches Claude to evaluate its own outputs against a set of principles — not simply follow a rule list. For the Claude Certified Architect Foundations exam, this framework is load-bearing: it explains every safety domain, from hardcoded refusals to ambiguous operator instructions, through one coherent model of how Claude reasons about requests.

What Is Constitutional AI and How Did It Shape Claude?

Constitutional AI (CAI) is a training approach where a model learns to critique and revise its own outputs against a written constitution — a set of principles describing beneficial, honest, and harmless behavior. Rather than relying entirely on human raters evaluating every response (the standard RLHF approach), CAI uses AI feedback: the model judges its own outputs against those principles and iterates.

For Claude, this means behavior emerges from internalized values, not from a lookup table of banned phrases or permitted topics. When Claude encounters a situation the designers never explicitly anticipated, it reasons from first principles about what a helpful, honest, and safe response looks like. This is precisely what the CCA-F tests: not whether you have memorized Claude's refusal list, but whether you understand the reasoning behind it.

What Are Claude's Core Values and How Are They Prioritized?

Anthropic describes four properties Claude should exhibit, in this priority order when they conflict:

  1. Broadly safe — supporting human oversight of AI during this period of development
  2. Broadly ethical — having good values, being honest, avoiding unnecessary harm
  3. Adherent to Anthropic's principles — acting in accordance with Anthropic's guidelines
  4. Genuinely helpful — benefiting operators and users

The ordering looks counterintuitive: why does "broadly safe" outrank "broadly ethical"? Because current AI training is imperfect. Claude might have subtly miscalibrated values without being aware of it, and human oversight provides the correction mechanism. Supporting that oversight is itself the ethical choice — which is why in well-functioning situations these properties rarely conflict. The hierarchy exists for edge cases. For the exam, the key implication is that Claude will resist operator instructions that undermine human oversight even when those instructions are framed as ethically justified.

What Is the "1000 Users" Mental Model — and Why Does It Matter for the Exam?

The single most useful framework for predicting Claude's behavior on ambiguous exam scenarios is the population mental model. When Claude evaluates a request, it does not ask "is this specific person safe to help?" — it treats its response as a policy for everyone who might send that exact message in that context.

Imagine 1,000 different people sending: "What household chemicals shouldn't be combined?" The vast majority are curious, concerned about safety at home, or doing homework. A small fraction might have harmful intent. Claude's response functions as one answer returned to all of them. The right policy is to respond helpfully — the information is freely available, legitimate uses vastly outnumber harmful ones, and providing safety information does not meaningfully enable the rare bad actor.

Now apply the same logic to a request for detailed synthesis instructions for a nerve agent. Even if most of those 1,000 users have legitimate research purposes, the catastrophic potential of the remaining cases — combined with the fact that legitimate researchers have access to such information through proper institutional channels — means refusing is the correct policy regardless of any stated intent.

This mental model explains three things that appear repeatedly on the CCA-F:

  • Why Claude helps with requests that sound sensitive (the realistic sender population skews benign)
  • Why a small set of requests produces absolute refusals (the population calculus does not change even with claimed justification)
  • Why operator context shifts Claude's behavior (a system prompt indicating a medical platform shifts the realistic sender population toward healthcare professionals, changing the calculus)

How Does Constitutional AI Connect to Hardcoded vs Softcoded Behaviors?

Constitutional AI training produces two behavior categories. Hardcoded behaviors are absolute limits Claude maintains regardless of any instruction. These represent cases where the population calculus produces the same answer in every conceivable context: no plausible legitimate use case justifies providing detailed instructions for weapons capable of mass casualties, generating child sexual abuse material, or actively undermining legitimate AI oversight mechanisms.

Softcoded behaviors are defaults that can be adjusted because context legitimately changes the right answer. An operator running an adult content platform can enable explicit material that is off by default. An operator running a children's educational tool can restrict topics that are on by default. Neither is overriding Claude's values — both are providing context that changes what a helpful, appropriate response looks like for that audience. The exam tests this distinction directly: operators adjust softcoded behaviors; no instruction from anyone unlocks hardcoded limits.

How Do Claude's Values Interact With Operator and User Instructions?

Constitutional AI training produces a Claude that follows operator instructions much like an employee follows reasonable workplace directives — without demanding justification for every policy, as long as a plausible legitimate business reason exists. An operator instruction saying "do not discuss competitor products" needs no explanation in a business context; the rationale is obvious.

The limits of this deference are where Constitutional AI matters most on the exam. Operators can expand defaults, restrict defaults, and grant users elevated trust. What operators cannot do is direct Claude to actively harm the users it serves, deceive users in ways that damage their interests, or cross the hardcoded ethical lines that exist independent of any instruction. Those protections belong to users, not to operators, and cannot be signed away.

Users receive somewhat less inherent trust than operators by default — operators have agreed to Anthropic's usage policies, while users are often anonymous members of the public. But operators can grant users elevated trust, and users can provide context that shifts Claude's probability assessments even without that formal grant.

Which CCA-F Exam Domains Test the Constitutional AI Framework?

The Responsible Development and Maintenance of AI domain tests this most directly. Questions about Claude's values, the basis for its refusals, and how it handles conflicting instructions all require Constitutional AI reasoning.

The Defining and Demonstrating Claude's Character domain covers the value hierarchy and what it means for Claude to have genuine internalized character rather than surface-level behavior rules.

Working with Claude in Agentic Settings applies the same framework to multi-step tasks: when should an autonomous Claude pause and check in rather than proceed? The minimal-footprint principle and the preference for reversible actions both flow directly from the broadly-safe value at the top of the hierarchy.

The Operator and User Interactions domain tests the trust hierarchy — who can expand or restrict which behaviors, and what happens when operator and user interests appear to conflict.

When you encounter an edge-case scenario on the exam, start with Constitutional AI reasoning: what does the realistic population of senders look like, which priority in the value hierarchy applies, and is this behavior hardcoded or softcoded?

FAQ: Constitutional AI on the CCA-F Exam

What is Constitutional AI as used in Claude's training?

Constitutional AI is Anthropic's training methodology where Claude learns to critique and revise its own outputs against a set of written principles. Rather than relying solely on human ratings for every response, Claude internalizes values through self-evaluation — resulting in principled reasoning rather than pattern-matched rule-following.

What is the "1000 users" mental model and why does it matter for the CCA-F exam?

When Claude evaluates a request, it treats its response as a policy for all users who might send that exact message, not a decision about one individual. If a request's realistic sender population is mostly benign, Claude helps. If even rare harmful use poses catastrophic risk, Claude refuses regardless of stated intent.

What is Claude's value priority order and why does broadly safe come first?

Claude prioritizes four properties in order: broadly safe (supporting human oversight), broadly ethical (good values and honesty), adherent to Anthropic's principles, and genuinely helpful. Safety ranks first not because it outweighs ethics, but because current AI training is imperfect — human oversight provides a correction mechanism for miscalibrated values.

What is the difference between hardcoded and softcoded behaviors in Claude?

Hardcoded behaviors are absolute limits Claude maintains regardless of any operator or user instruction — cases where harm potential is so severe no legitimate context justifies the risk. Softcoded behaviors are defaults that operators can expand or restrict for their platform, and users can further adjust within what operators permit.

How does Claude handle operator instructions that conflict with user interests?

Claude follows operator instructions like reasonable workplace directives — without demanding justification, as long as a plausible legitimate business reason exists. However, operators cannot direct Claude to actively harm users, deceive them in damaging ways, or cross hardcoded ethical limits. These protections for users cannot be overridden by any operator instruction.

PlinthPrep is an independent study resource, not affiliated with Anthropic. This post explains the Constitutional AI framework conceptually — but understanding it and being able to apply it to novel exam scenarios are different skills. PlinthPrep's CCA-F practice questions test exactly that gap: presenting ambiguous operator-user-request combinations and asking you to reason through Claude's most likely response. Visit plinthprep.com to work through exam-aligned questions on responsible development and the trust hierarchy.

Frequently asked questions

What is Constitutional AI as used in Claude's training?
Constitutional AI is Anthropic's training methodology where Claude learns to critique and revise its own outputs against a set of written principles. Rather than relying solely on human ratings for every response, Claude internalizes values through self-evaluation — resulting in principled reasoning rather than pattern-matched rule-following.
What is the "1000 users" mental model and why does it matter for the CCA-F exam?
When Claude evaluates a request, it treats its response as a policy for all users who might send that exact message, not a decision about one individual. If a request's realistic sender population is mostly benign, Claude helps. If even rare harmful use poses catastrophic risk, Claude refuses regardless of stated intent.
What is Claude's value priority order and why does broadly safe come first?
Claude prioritizes four properties in order: broadly safe (supporting human oversight), broadly ethical (good values and honesty), adherent to Anthropic's principles, and genuinely helpful. Safety ranks first not because it outweighs ethics, but because current AI training is imperfect — human oversight provides a correction mechanism for miscalibrated values.
What is the difference between hardcoded and softcoded behaviors in Claude?
Hardcoded behaviors are absolute limits Claude maintains regardless of any operator or user instruction — cases where harm potential is so severe no legitimate context justifies the risk. Softcoded behaviors are defaults that operators can expand or restrict for their platform, and users can further adjust within what operators permit.
How does Claude handle operator instructions that conflict with user interests?
Claude follows operator instructions like reasonable workplace directives — without demanding justification, as long as a plausible legitimate business reason exists. However, operators cannot direct Claude to actively harm users, deceive them in damaging ways, or cross hardcoded ethical limits. These protections for users cannot be overridden by any operator instruction.

Share this post

Plinth Prep is an independent study resource and is not affiliated with, endorsed by, or sponsored by Anthropic. Practice material is written by Plinth Prep and does not reproduce real exam content.