When patent counsel runs prior art for your governance patents and the closest reference is a paper called “Constitutional AI,” you read it carefully. We did. What we found is that Anthropic’s Constitutional AI and CANONIC share exactly one word, and it is the word that matters most: constitutional. The resemblance ends there. Constitutional AI is a training technique for making language models less harmful. CANONIC is a governance system for making institutions provably trustworthy. One operates at training time inside a neural network. The other operates at commit time inside a version-controlled evidence layer. One is probabilistic. The other is deterministic. They are not competitors. They are not even in the same category.
The Prior Art Question
Every patent portfolio faces a moment where counsel asks: what else is out there? For CANONIC’s six provisional applications, that question produced 42 query clusters across four patent offices and three non-patent literature databases. The search covered Google, Microsoft, NVIDIA, IBM, Anthropic, Epic Systems, Optum, Veeva, Oracle, Qomplx, and Bank of America. The result across US, European, Japanese, and non-patent literature was consistent: no blocking prior art for any of the six applications.
But one reference kept surfacing in every “constitutional AI” query cluster. Not as a patent, because Anthropic never patented it. As a paper: Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” published on arXiv in December 2022 with 51 authors. It is the most cited work in the AI alignment literature, and it shares a word with our disclosure family. So we had to understand it completely, not to distinguish it from our claims (that turned out to be straightforward) but to understand what it reveals about the gap between training-time alignment and runtime governance.
What Constitutional AI Actually Does
Constitutional AI is a two-phase training technique. In the first phase, a language model generates a response, critiques its own response against a set of natural language principles (the “constitution”), and revises the response to be less harmful. The revised response becomes training data for a supervised fine-tuning step. In the second phase, the model generates pairs of responses, an AI evaluator picks the less harmful one using the same constitutional principles, and those preferences train a reward model for reinforcement learning.
The constitution itself is a set of 16 natural language principles. Example: “Identify specific ways in which the assistant’s last response is harmful, unethical, racist, sexist, toxic, dangerous, or illegal.” The model reads the principle, evaluates its own output, and generates a critique. Then it reads a revision request and generates a less harmful version. Four rounds of critique and revision per prompt, each sampling a different principle.
The innovation is replacing human feedback with AI feedback for the harmlessness dimension. Traditional RLHF requires human annotators to rank outputs by harmlessness. Constitutional AI replaces those human labels with AI-generated labels, guided by the constitution. The authors called this RLAIF: reinforcement learning from AI feedback. It achieved a Pareto improvement over RLHF, producing models that were simultaneously more harmless and more helpful, and it eliminated the evasive refusal behavior that plagued earlier safety-trained models.
The paper is elegant. The engineering is real. The results are significant. And none of it has anything to do with governance.
The Word That Matters
A constitution, in the political sense that gives the word its weight, is a binding contract between an institution and its constituents. It declares what the institution will do, constrains what it may do, and provides a mechanism for enforcement that does not depend on the good intentions of whoever holds power. The US Constitution does not hope that Congress will respect the First Amendment. It provides judicial review, an enforcement mechanism that operates independently of congressional intent.
Constitutional AI borrows the word but not the mechanism. The constitution in Constitutional AI is a set of natural language instructions that guide a model during training. After training, the constitution disappears. It is not embedded in the model’s weights in any recoverable form. It is not auditable at runtime. There is no mechanism to verify that a deployed model is actually following its constitution, because the constitution was a training signal, not an enforceable contract. The model may have learned the spirit of the principles, or it may have learned a statistical proxy that correlates with the training signal. No one can tell from the outside.
This is not a criticism. It is a category distinction. Constitutional AI is an alignment technique: a way to steer model behavior during training so that the deployed model is more likely to behave well. CANONIC is a governance system: a way to declare, enforce, and audit institutional commitments so that compliance is deterministic, not probabilistic.
The Formalism Gap
The prior art analysis identified the gap precisely. Constitutional AI uses natural language principles evaluated by language models during training. CANONIC uses formal encoding evaluated by machine operations during runtime compliance checking. The differences cascade:
Constitutional AI’s principles are natural language, which means they are ambiguous by design. “Identify specific ways in which the response is harmful” requires interpretation. Two runs of the same critique-revision loop on the same prompt will produce different critiques, because language model outputs are stochastic. The constitution guides but does not determine.
CANONIC’s governance contracts are deterministic. A governed scope either satisfies its declared constraints or it does not, and the answer is the same every time you check. There is no interpretation step. The compiler checks what was declared against what was produced and returns a binary result. This is what makes it auditable: any third party can run the same check and get the same answer, because the check is a mechanical operation, not a judgment call.
Constitutional AI operates at training time. Once the model is trained, the constitution has been absorbed into the model’s parameters and cannot be inspected, modified, or revoked. If the constitution needs to change, the model needs to be retrained. CANONIC operates at commit time. The governance contract is a living document in version control. If the contract needs to change, you commit the change, and the compiler enforces the new contract on the next build. The change is visible in the diff, traceable in the audit trail, and reversible via git revert.
Constitutional AI is model-scoped. It governs the behavior of a single language model. CANONIC is institution-scoped. It governs the behavior of an entire organization’s knowledge layer, of which language models are one component. A hospital using Constitutional AI has a model that is less likely to say harmful things. A hospital using CANONIC has a governance layer that proves which clinical knowledge the system has committed to, which evidence sources back each claim, and which scopes of practice the system is credentialed for.
Introspection: The Real Difference
Both systems claim introspection, and this is where the comparison becomes instructive.
In Constitutional AI, introspection is the critique step. The model reads its own output and evaluates it against a principle. This is a powerful engineering technique, but it is not introspection in the governance sense. The model cannot inspect the principles themselves, cannot reason about whether the principles are consistent with each other, cannot modify the principles, and cannot verify that its behavior after training actually reflects the principles it was trained on. The introspection is one-directional and one-level: the model evaluates its output against fixed rules, but nothing evaluates the rules.
CANONIC supports arbitrary introspection depth. Level 0 is governance of artifacts: rules about what the system produces. Level 1 is governance of governance: rules about the rules. Level 2 is governance of the governance of governance. The recursion terminates at the root contract, which is self-governing. At every level, the governance is explicit, auditable, and enforceable. The system can reason about any level because every level is a governed document in version control, not an absorbed training signal in a neural network.
This is not an abstract distinction. When Anthropic updated Claude’s constitution in January 2026 from 2,700 words to 23,000 words, the update required retraining or fine-tuning the model against the new principles. The old constitution and the new constitution cannot coexist in the same model, and there is no mechanism to verify that the deployed model reflects the new constitution rather than the old one, other than behavioral testing. When CANONIC updates a governance contract, the change is a git commit. The old contract is in the history. The new contract is in HEAD. The diff shows exactly what changed. The compiler enforces the new contract. The audit trail is complete.
The Bootstrap Problem
Constitutional AI has an elegant bootstrap problem that the authors acknowledge: you need an already-aligned model to generate the AI feedback that aligns the next model. If the AI labeler has biases, those biases propagate and potentially amplify through the training loop. The quality of the constitution’s enforcement is bounded by the quality of the model doing the enforcing, and that model was trained by a previous round of the same process. It is turtles all the way down.
CANONIC has no bootstrap problem because the enforcement mechanism is not a language model. It is a compiler. The compiler does not need to be aligned. It does not need to understand the domain. It does not need to interpret natural language. It checks formal constraints and returns a deterministic result. The domain knowledge comes from humans who commit it to the governed evidence layer. The compiler enforces the structure. The humans supply the truth. Neither depends on the other being “aligned” in any statistical sense.
What the Due Diligence Revealed
The prior art search confirmed what the architecture always implied: Constitutional AI and CANONIC are not in the same technical category. The NPL analysis classified Constitutional AI as “partial overlap” due to shared terminology but “fundamentally different mechanism.” The differentiation was clean enough that patent counsel did not flag it as a risk.
But the search revealed something more interesting than a patent distinction. It revealed a gap in the entire field. Every prior art reference for governance-related queries fell into one of two categories: conceptual frameworks with no formal enforcement mechanism, or enforcement mechanisms with no governance framework. The EU AI Act is a regulatory text with no compiler. XACML is an access control language with no governance model. Bell-LaPadula is a lattice-based security model for operating systems, not AI governance. Constitutional AI is an alignment technique with no runtime enforcement. Nobody had built the thing that connects governance declarations to deterministic runtime enforcement across institutional scope.
The search ran 42 query clusters. “Bitwise governance”: zero results. “Compliance lattice”: zero results. “Constant time compliance checking”: zero results. “Meta-governance”: zero results. “Governance of governance”: zero results. The field had alignment techniques for training models, regulatory frameworks for governing institutions, and access control languages for securing systems. It did not have a governance language that compiles.
Two Theories of Trust
Constitutional AI and CANONIC represent two theories of how to make AI systems trustworthy, and the theories are complementary, not competing.
Constitutional AI says: if you train the model well enough, with the right principles, using the right feedback loop, the model will behave in accordance with those principles. Trust comes from the quality of the training process. This is the alignment thesis, and it has produced real progress. Claude is measurably less harmful than its predecessors, and Constitutional AI is one reason why.
CANONIC says: trust comes not from how a model was trained but from what an institution has committed to a governed evidence layer. The model is a component. The governance contract is the source of trust. If the model changes, if the training data changes, if the alignment technique changes, the governance contract still holds because it is a versioned, auditable, enforceable declaration independent of any particular model.
A hospital that uses Claude benefits from Constitutional AI’s alignment. A hospital that uses CANONIC benefits from a governance layer that works regardless of which model sits behind it. The ideal is both: a well-aligned model operating within a governed evidence layer, where the alignment provides a behavioral floor and the governance provides an institutional ceiling.
The Close
When we ran prior art and Constitutional AI surfaced in every “constitutional” query cluster, we expected a fight. What we found was a clean distinction that clarifies what both systems are and what neither system is alone. Constitutional AI makes models less harmful by training them against natural language principles. CANONIC makes institutions provably trustworthy by compiling governance contracts into deterministic enforcement. One is a training technique. The other is a governance system. One disappears into model weights after training. The other persists in version control forever.
Anthropic chose not to patent Constitutional AI. They published it as open research, open-sourced the code, and released Claude’s constitution under Creative Commons. That decision reflects a theory about how AI safety should propagate: openly, without patent barriers, through the research community. We respect that choice. CANONIC’s patent portfolio protects a different kind of innovation, not a training technique but a governance language, and the prior art search confirmed that no one else has built one.
The next time someone asks whether CANONIC is like Constitutional AI, the answer is: Constitutional AI is a recipe for training a model. CANONIC is a constitution for governing an institution. Recipes are useful. Constitutions are binding.
Sources
| Claim | Source | Reference |
|---|---|---|
| Bai et al., “Constitutional AI: Harmlessness from AI Feedback” | Anthropic, arXiv:2212.08073, December 2022 | arxiv.org/abs/2212.08073 |
| 51 authors, RLAIF methodology, 16 constitutional principles | Bai et al. 2022 | arxiv.org/abs/2212.08073 |
| Constitutional AI achieves Pareto improvement over RLHF | Bai et al. 2022, Figure 2 | arxiv.org/abs/2212.08073 |
| Claude’s constitution updated Jan 2026: 2,700 to 23,000 words, CC0 1.0 | Anthropic, January 2026 | anthropic.com/news/claude-new-constitution |
| Prior art search: 42 query clusters, 4 patent offices, 14 targeted assignees | CANONIC Due Diligence, February 2026 | PATENTS/DUE-DILIGENCE/US-PRIOR-ART-SEARCH.md |
| “Bitwise governance” zero results across all databases | CANONIC NPL Prior Art Search | PATENTS/DUE-DILIGENCE/NPL-PRIOR-ART-SEARCH.md |
| No blocking prior art for any of 6 provisional applications | US/EPO/JPO/NPL consolidated finding | PATENTS/DUE-DILIGENCE/PROV-PRIOR-ART-ANALYSIS.md |
| Anthropic has no Constitutional AI patent; portfolio is agentic automation | USPTO assignee search, Justia Patents | patents.justia.com/assignee/anthropic-pbc |
| CANONIC governance language: formal encoding, deterministic compliance | PROV-001 provisional application | PATENTS/PROV-001/ |
| Introspection depth levels: Level 0 through Level N, reflective closure | DISCLOSURE-CONSTITUTIONAL-004 | PATENTS/DISCLOSURES/DISCLOSURE-CONSTITUTIONAL-004.md |
| Governance is compilation | “The Compiler Insight,” Hadley Lab, December 2025 | hadleylab.org |
| Prompts are recipes, governance is theory | “Stop Prompting, Start Governing,” Hadley Lab, March 2026 | hadleylab.org |
Figures
| Context | Type | Data |
|---|---|---|
| post | gauge | value: 255, max: 255, label: CONTRACT COVERAGE |
| *Constitutional AI Is Not a Constitution | GOVERNANCE + LAW | BLOGS* |