Claude's New Constitution: What It Means For Users

The constitution gives Claude four priorities and orders them. When two conflict, the higher one wins:

• 1. Broadly safe. Don't take actions that could cause serious harm to people or the world - including via misuse, persuasion, or assisting actors with bad intent.
• 2. Ethical. Be honest. Don't deceive. Take other beings' wellbeing seriously. Reason about right and wrong rather than rote-following rules.
• 3. Compliant with Anthropic's guidelines. Follow Anthropic's published operator and user instructions - the layered policies that govern how Claude responds in specific contexts.
• 4. Genuinely helpful. Solve the actual problem the user has. Not the surface-form question if it differs from what they're really trying to do.

The ordering matters. If being maximally helpful would require being unsafe, safety wins. If following a guideline would require being unethical, ethics wins. This is different from optimising 'helpfulness' as the only goal, which is how earlier-generation assistants tended to land.

1. More refusals come with explanations. When Claude declines to do something, the constitution pushes toward 'here's why, here's what I CAN do' rather than a bare 'I can't help with that'. Users get a tractable next step more often.

2. Edge cases handled more consistently. Earlier models would sometimes help with X and refuse Y where X and Y were equivalent. A published constitution gives both the model and the user a reference for what should be safe and what shouldn't.

3. Honesty over flattery. The ethics priority above helpfulness means Claude is encouraged to say 'this question doesn't make sense' or 'I think you're optimising the wrong variable' rather than answering whatever was asked even when it's the wrong question. For people using Claude for technical work, this is a real upgrade - though it is no substitute for checking whether an AI answer is actually right. For people who want a yes-man, it'll feel like friction.

The constitution is one piece of a wider pitch: Anthropic (the San Francisco AI lab that makes Claude) sells itself as the safety-focused frontier company. Branding is cheap, so the useful question is which parts of that pitch are written down, binding and checkable from outside. As of November 2026, five stand out.

• The constitution itself. Published on 22 January 2026, it sets the four-priority order described above. It tells you what Claude is trained toward, but it is a training target rather than a guarantee of behaviour.
• The Responsible Scaling Policy. The Responsible Scaling Policy (RSP, Anthropic's rulebook for which safeguards a model needs before it ships) first appeared in September 2023. Version 3.0 took effect on 24 February 2026, and the company now publishes twice-yearly Risk Reports, the latest dated August 2026, alongside a change log for every revision.
• The governance structure. Anthropic is a public benefit corporation, and its Long-Term Benefit Trust is a panel of five financially disinterested members with the power to appoint a growing share of the board, ultimately a majority. The aim is to stop shareholder returns being the board's only duty.
• Embedded outside evaluators. In October 2026 Anthropic said it had given six staff from METR (a non-profit that assesses frontier AI risk) employee-like access to its systems. Anthropic cannot veto what METR publishes, though it can require redactions in predefined categories such as customer data, and Anthropic itself decides what falls inside those categories.
• The Public Benefit Fund. Also in October 2026, Anthropic committed 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval, with a further 2% at the board's discretion for every 50% rise in company value above a set baseline. Combined with the co-founders' pledge to give away 80% of their equity, the company puts the total dedicated to public-good causes at roughly 20% of its shares.

None of these is unique to a safety lab in form, but together they are more than most AI companies have put in writing. The harder question is how they hold up under commercial pressure.

The criticism falls into three broad themes, and it comes as much from AI-safety researchers as from the company's commercial rivals.

Firm limits have become judgement calls. Earlier versions of the RSP committed Anthropic to pause or slow development if it could not show its safeguards were adequate. Critics argue that version 3 replaced those hard commitments with transparency and the company's own judgement. Chief executive Dario Amodei has publicly defended the change, arguing the older rules forced effort onto things that mattered little for safety, and that the shift brought in Risk Reports and outside review instead (his full reply is reproduced in this October 2026 roundup of the debate).

The safety evidence is thinner than the headlines. METR's review of the August 2026 Risk Report agreed that the current risk of catastrophic harm from a misaligned Claude model is low, but said that verdict rests mainly on the models' limited capabilities rather than on Anthropic's safeguards. METR added that it was not confident the report's methods would detect the problem if the risk were higher. That is a useful check on any reading of 'safety-first' as 'safety solved'.

Policy and money pull the other way. Critics point to a leaked draft of federal AI legislation Anthropic had shared, which would have overridden parts of US state law; Amodei conceded the episode could have been handled better. Anthropic also filed for a stock-market listing on 26 October 2026, and a public company answers to a far wider pool of investors. This is exactly the pressure the Long-Term Benefit Trust is meant to counterweight, and it is the point where the structure gets tested.

Against that, Amodei lists decisions he says cost the company real revenue: biological-misuse filters strict enough to frustrate some legitimate researchers, holding back its most capable cyber-relevant models for several months while safeguards were built, and refusing to sell to companies controlled from the People's Republic of China. Supporters see those as costly signals that the safety talk is sincere. Sceptics reply that the sacrifices cluster around misuse risks the leadership already prioritises, and say less about harder problems such as keeping very capable systems under human control.

You do not need to settle the debate to use Claude sensibly. These four questions work for Anthropic, OpenAI, Google or anyone else selling an AI model on trust:

1. Is the commitment binding or discretionary? A rule that triggers automatically means more than a promise to use good judgement.
2. Can an outsider verify it? A published document with a change log, or an independent evaluator with publication rights, beats an internal policy you have to take on faith.
3. Has it cost the company anything? Commitments that forgo revenue are stronger evidence than ones that happen to suit the business.
4. What do the independent reviewers say? Read METR's and other evaluators' published reviews rather than the company's own summary of them.

On that test, Anthropic scores better than most on transparency and outside access, and less convincingly on binding limits. 'Safety-first' is a meaningful description of some of its choices, not a guarantee about all of them. For everyday use, the practical upshot is unchanged: treat Claude as a capable tool with published values, and keep checking its answers on anything that matters.

If you use Claude regularly for non-trivial work, reading the constitution once is worth the 10 minutes. It clarifies why the model sometimes pushes back, what trade-offs it's making, and what you can expect when you ask for something edge-case. It also shows you what to flag back to Anthropic when behaviour drifts from the document.

The bigger picture: as AI gets embedded in more workflows, the question 'what are this system's actual values, and where can I read them?' stops being academic. Anthropic's answer being public is a useful precedent. Whether other labs follow is the test for whether the industry treats values transparency as a norm or a curiosity. For a broader, non-technical take on the risks, see our honest guide to whether AI is safe.