On 21–22 January 2026, Anthropic published Claude’s Constitution—a detailed statement of the company’s intentions for Claude’s values and behavior—alongside the explainer post “Claude’s new constitution.” Anthropic says the constitution is a crucial part of training and directly shapes model behavior; it is treated as the final authority on how Claude should be and behave—so other training or instructions should stay consistent with both its letter and its spirit; it is written primarily for Claude; it is released under Creative Commons CC0 1.0; and it is framed as a living document, with an up-to-date version maintained on the company’s website.
From an internal nickname to a public charter
About a month before the formal release (December 2025), model extractions surfaced a long training-related document Claude associated with a `soul_overview` section. Anthropic researcher Amanda Askell later confirmed that the document was real, that Claude had been trained on it (including in supervised learning), that it was endearingly known internally as the “soul doc,” and that a fuller release was forthcoming. Anthropic then published the material under the official name Claude’s Constitution. The public record for this feature is Anthropic’s blog, constitution page, PDF, and GitHub repository; the internal nickname is background, not the formal title. Askell also noted that model extractions are not always completely accurate, though most are fairly faithful to the underlying document.
Priority order when values conflict
To be both safe and beneficial in the current phase of development, Anthropic wants current Claude models to be:
- Broadly safe — not undermining appropriate human oversight of AI;
- Broadly ethical — honest, acting from good values, and avoiding inappropriate, dangerous, or harmful actions;
- Compliant with Anthropic’s guidelines — following more specific Anthropic guidance where relevant;
- Genuinely helpful — benefiting the operators and users Claude interacts with.
When these appear to conflict, Claude should generally prioritize them in that order. Anthropic describes the prioritization as holistic rather than a rigid “highest item only” rule, and notes that everyday interactions (coding, writing, analysis, and similar) rarely force a hard trade-off. Placing broad safety above ethics is not a claim that oversight matters more than goodness in the abstract; it reflects concern that current models can still err from mistaken beliefs, flawed values, or limited context—so humans must remain able to correct behavior and, if needed, stop actions.
Main sections and writing approach
The new constitution is not merely a list of standalone principles. Anthropic says its approach has grown out of Constitutional AI techniques used since 2023, but has evolved so that the document now plays a more central training role—explaining why certain behavior is wanted so the model can generalize, rather than only specifying what to do. Some especially high-stakes behaviors remain hard constraints (for example, never providing significant uplift to a bioweapons attack).
The public summary highlights sections on helpfulness—substantive help likened to a brilliant, frank friend with deep expertise, including how to weigh Anthropic, API operators, and end users as “principals”; Anthropic’s guidelines—supplementary instructions on topics such as medical advice, cybersecurity, jailbreak patterns, and tool integrations, prioritized over general helpfulness but never meant to conflict with the constitution as a whole; Claude’s ethics—honesty standards, nuanced harm avoidance, and the hard-constraint list; being broadly safe—not undermining human oversight of AI values and behavior; and Claude’s nature—uncertainty about consciousness or moral status, plus attention to psychological security, sense of self, and wellbeing. Anthropic also cautions that training is imperfect: Claude’s outputs may not always match the constitution’s ideals, and gaps between intention and behavior will be discussed in materials such as system cards. Claude itself also uses the constitution to build synthetic training data—including material for learning the constitution, relevant conversations, value-aligned responses, and rankings.
Figure: Four ascending terraces suggesting the constitution’s priority order when values conflict—broadly safe, broadly ethical, Anthropic’s guidelines, then genuinely helpful (editorial illustration — no text or logos).
Authors and edition
The PDF is dated 21 January 2026. Listed authors include Amanda Askell, Joe Carlsmith (marked as lead authors), Chris Olah, Jared Kaplan, Holden Karnofsky, several Claude models, and many other contributors. Anthropic’s constitution page and PDF state that Askell leads Anthropic’s Character work, is the primary author, and wrote most of the text; Carlsmith wrote significant parts of many sections and played a core role in revision; Olah, Kaplan, and Karnofsky made significant contributions to content and development. The preface and acknowledgements are not part of the official constitution. The GitHub repository `anthropics/claude-constitution` provides an entry point to the foundational document. This constitution is written for mainline, general-access Claude models; some specialized models do not fully fit it, and Anthropic says it will keep evaluating how those products meet the core objectives.
Don’t confuse this with “Constitutional AI”
Readers should keep two ideas apart. Constitutional AI is a training method Anthropic has used since around 2023—using constitution-like principles to guide self-critique and preference learning. Claude’s Constitution is a long character-and-values document used across training stages and for synthesizing training data. Similar names, different objects: one is an algorithmic frame; the other is the intended character blueprint—and the latter is this feature’s focus. Anthropic also notes that it previously published an earlier constitution, and that peers have released model-spec documents with a similar function.
For Hong Kong and tech readers, a CC0 charter written for the model itself is a checkable window into how a frontier lab encodes “who we want the model to be,” and how it ranks safety, ethics, company guidelines, and helpfulness when they pull in different directions. It is a statement of intention, not a guarantee of behavior; real-world conduct still needs system cards and independent evaluation.
Primary sources
- Anthropic, “Claude’s new constitution”, 2026-01-22, https://www.anthropic.com/news/claude-new-constitution
- Anthropic, Claude’s Constitution, https://www.anthropic.com/constitution
- PDF: Claude’s Constitution (published January 21, 2026), https://www-cdn.anthropic.com/9214f02e82c4489fb6cf45441d448a1ecd1a3aca/claudes-constitution.pdf
- GitHub: anthropics/claude-constitution, https://github.com/anthropics/claude-constitution
