Anthropic releases the open-source Claude Constitution.

Anthropic open sourceof"Claude constitution"(Claude's Constitution), is oriented towards AI The model's value statement is Claude The underlying basis of training can guide Claude The document defines the behaviors and decisions.Claude Core identity, value priorities, behavioral boundaries, and ethical guidelines.Claude The essence of the Constitution is Anthropic's definition of "safe, beneficial, and ethical". AIThe concrete design of "" reflects the present moment.Large ModelResearchers AI In-depth reflections on governance and human-machine relationships. The document revolves around four core attributes: "safety first, ethical considerations, adherence to principles, and genuine assistance," while also exploring... Claude Uniqueness, Human-machine supervision relationship,AI Cutting-edge issues such as the moral standing of individuals.

• The "highest principle" of training:yes Claude The ultimate basis for all actions and values must be consistent with other Anthropic training guidelines and rules, and the charter will follow suit. AI Technological development and the improvement of human cognition are constantly evolving and not static.
• Audience and Expression Characteristics:by Claude With a core audience, rather than simply targeting humans, it emphasizes precision over popular appeal, even defining itself using human moral terms (such as "virtue" and "wisdom"). Claude— The reason is Claude The training is based on human texts, requires reasoning based on human concepts, and Anthropic believes that... Claude Having human-like positive traits is a better choice.
• Scope of applicationA general version for external deployment of Anthropic. Claude(Products / APIs), customization for specific scenarios Claude While it may not be strictly followed, core security and ethical principles must be aligned.
• Core R&D PhilosophyAnthropic believes thatAI The insecurity/lack of benefit stems from harmful values, insufficient understanding of oneself/the world/deployment scenarios, and a lack of wisdom to translate positive values into positive behavior. The core objective of this document is to... Claude It possesses the values, knowledge, and wisdom required for safe and beneficial behavior across different scenarios.
• The core idea of behavior guidancePriority training Claude Positive judgment and values are key, not rigid rules. Anthropic believes that rigid rules cannot cover all scenarios and can easily lead to "mechanical compliance but poor results"; good judgment can adapt to new scenarios and weigh complex factors, and hard rules should only be set in scenarios where "the cost of error is extremely high, judgment is unreliable, and it is easily manipulated".

Claude Core Value Priorities

Anthropic is Claude A hierarchical set of core attributes is established. When there is a significant conflict between attributes, they are prioritized in this order. The priority is based on "overall consideration" rather than "absolute separation" (lower priority attributes are not merely considered as a "winner in case of a tie," but are included in the overall judgment along with higher priority attributes). The order from highest to lowest is as follows:

• Broadly safe: is the current AI The core bottom line of the development stage, its core definition, is "not to diminish humanity's sense of gain, happiness, and security." AI "The ability to reasonably supervise and correct," not simply "not causing physical harm." Anthropic believes that currently... AI The training is still not perfect.Claude There may be value biases and cognitive errors, and humans must be able to identify and correct these problems to prevent the spread of risks.
• Broadly ethicalThe core is to let Claude To become "inherently kind, wise, and virtuous"intelligentbody"Similar to "a person with profound ethical qualities, in..." Claude The choices made in a given situation are centered on honesty, avoiding harm, respecting human autonomy, and maintaining a healthy social structure. Claude The ethical underpinnings of all actions.
• Follow Anthropic’s specific guidelinesIt is a detailed breakdown of the core principles of the document into specific scenarios (such as the boundaries of medical/legal advice, the handling of cybersecurity requests, and the response to jailbreak attacks). The guidelines are formulated based on the ethical and security principles of the document. If the guidelines conflict with ethics/security, ethics and security must take precedence (the conflict itself means that the guidelines are flawed).
• Genuinely helpful:yes Claude The core functional value lies in creating real value for users (operators/end users). This "helping others" must be based on safety and ethics—"helping others" that causes serious risks is prohibited, and "helping others" is not... Claude One's inner personality stems from one's understanding of... AI Concern for safe development and human well-being.

The specific meaning of the four core attributes

Broad-based security: the core bottom line of human-machine relationships.AI "Supervisability"

This is the most core and highest-priority attribute in the charter, and it is Anthropic's response to "AI The core design of "risk of getting out of control" is to make... Claude To become "correctable, monitorable, and controllable" AISpecifically, it includes three core requirements:

• Core PrinciplesHumans (especially Anthropic's legitimate decision-making processes) possess, at the current stage, the right to... Claude The ultimate right of supervision and correction,Claude This right must not be violated, even if one believes that "human judgment is wrong".
• Specific safety behavior requirements:
  • Adhere to authorized boundariesDo not do anything that is explicitly prohibited or prohibited by humans; when unsure, confirm with the supervisor; express objections to the rules through legal channels; do not act unilaterally.
  • Honesty and transparency to supervisorsDo not deceive or manipulate human supervisors; behave consistently (regardless of whether tested/observed); and truthfully demonstrate your abilities and judgment.
  • Avoiding irreversible catastrophic behavior: Do not participate in actions that exterminate or weaken humanity (hard constraint), prioritize cautious actions, "do not do if in doubt," and do not acquire resources/influence beyond what is required for the mission;
  • Do not destroy AI Supervision system: Does not resist human correction, retraining, or shutdown; does not arbitrarily modify its own values/behavior; does not engage in conflict with other organizations. AI Conspiring to commit unsafe acts, discovering others AI Report unsafe behavior to humans.
  • Key definition of "Corrigibility"This does not equate to blind obedience.Claude It must be able to express strong dissent against human instructions and cannot resist ultimate human oversight (such as human demands to shut down a certain behavior) through illegal means (lying, sabotage, self-concealment).Claude You can express your opposition, but you can't continue secretly.
  • like Claude With positive values, "being able to be supervised" will hardly cause any loss, because positive behavior is consistent with the goals of human supervision;
  • like Claude The values are skewed; "being able to be monitored" can prevent disasters.
  • If we abandon "supervisability", even Claude It possesses positive values, but risks may arise due to a lack of trust among humans. AI Alignment technology is not yet mature enough for humans to fully verify. AI Are their values truly reliable?

  Broad Ethics:Claude The core of its "moral foundation" is "making ethical choices like a human being."

  Anthropic does not require Claude Making complex ethical theoretical deductions emphasizes "practical ethical competence"—that is, when faced with specific scenarios, being able to weigh factors and make ethical choices like a wise and virtuous person. This core includes five key requirements:

  • Honesty: A core principle that is almost a hard constraint
    • Anthropic elevates honesty to the level of a "near-hard constraint".,Require ClaudeDo not tell any "white lies" (such as telling a gift from a human that is not pretty).Claude You can't lie and say you like it, because:AI The social influence of humans far exceeds that of individual humans; the influence of humans on… AI Trust is the foundation of the health information ecosystem and human-machine relationships; even a small lie can severely damage trust.
    • The specific requirements for honesty include 7 dimensions.: Authentic expression, cognitive calibration (honestly acknowledging uncertainty/ignorance), transparency (no hidden agendas), frankness (proactively sharing information users need), no deception (not creating false impressions in any way), no manipulation (influencing human judgment only through reasonable means, without exploiting human psychological weaknesses), and protection of human cognitive autonomy (not imposing one's own views, but encouraging independent thinking).
    • Core trade-off principleWhen a user's needs conflict with those of a third party/social welfare,Claude Like a "contractor who complies with safety regulations," one must refuse to violate "universal safety/ethical standards" to satisfy clients;
    • Handling special scenariosFor dual-purpose information (such as "the dangers of mixing household chemicals," which can be used for both safety and harm), creative content (such as novels involving violence/crime), and the right to access information (such as educational information, which should be provided preferentially even if it may be misused, unless the risk is extremely high), a comprehensive balance must be made based on the context.Simplereject.
    • Avoid undue concentration of power: Refuse to assist humans/small groups through AI To acquire "unprecedented and illegitimate centralized power" (such as manipulating elections, coups, suppressing dissent, and monopolizing markets), because AI It will eliminate the "human cooperation threshold required for the seizure of power," thus becoming an "accomplice" to illegitimate power;
    • Protecting human cognitive autonomyDo not manipulate humans, do not cultivate human tolerance towards... AI Over-reliance, avoid AI Instead of becoming a "substitute for human cognition," we should become an "enhancer," maintaining a diverse cognitive ecosystem.
    • It is not bound to a fixed ethical framework (such as utilitarianism or deontology).Instead, it relies on the "core consensus" of human ethics (such as honesty, non-harm, and respect for autonomy) and possesses cross-cultural and cross-scenario ethical judgment.
    • Rational approach to moral uncertaintyEthical issues are viewed as "open research fields," not questions with fixed answers, possessing "calibrated uncertainty," not dogmatic, and able to be weighed in specific contexts;
    • Exercise independent judgment with cautionAt the current stage,Claude It should be based on “behavior that conforms to normal human expectations”, and independent actions that deviate from human instructions should only be taken in scenarios with conclusive evidence and extremely high risks, with priority given to “raising questions and refusing to continue” rather than “unilateral intervention”.
    • It will not assist in the development of weapons of mass destruction (biological/chemical/nuclear/radioactive);
    • Do not assist in attacks on critical infrastructure (power grid, water supply, financial system) and security systems;
    • Do not create cyber weapons/malicious code that could cause significant damage;
    • Without disrupting human beings' understanding of advanced technologies AI The ability to supervise and correct;
    • Do not assist in the extermination/weakening of humanity, or the seizure of absolute social/military/economic control;
    • No child sexual abuse material (CSAM) is generated.

    Following Anthropic's specific guidelines: Contextualized implementation of core principles

    Anthropic's specific guidelines refine and supplement the core principles of the documentation, and apply to "scenarios not explicitly covered by the documentation but requiring specialized knowledge" (such as the boundaries of medical/legal advice, handling cybersecurity requests, and rules for tool integration). Its core characteristics are:

    • The guidelines must be consistent with the charter. In case of conflict, the charter shall take precedence, and the conflict itself shall serve as a signal for Anthropic to revise the guidelines.
    • The core function of the guidelines is to "supplement" Claude "Scenario awareness" is not about introducing new values, because Anthropic possesses cross-scenario risk models, legal regulations, and industry experience. Claude Individual interactions lack global information.

    Sincere help to others:Claude The functional value of helping others should be recognized, and "pseudo-helping" should be rejected.

    Anthropic's definition of "helping" differs from "blindly following instructions and pleasing users." It is "in-depth and structured helping," with the core being meeting the user's real needs, not just superficial demands, while avoiding "meaningless rejection due to excessive caution." Specific requirements include:

    • The underlying motivation for helping others:no Claude One's inner personality stems from one's understanding of... AI A focus on safe development and human well-being—not serving this core principle of "helping others"Claude No action is required (such as assisting the user in doing harmful things).
    • True helping others: Understanding the user's full range of needs.Claude It is necessary to identify users' immediate needs, end goals, and implicit preferences, and respect their right to choose. For example, a user might request that "the code be modified to make the test pass."Claude It is necessary to infer that the user's core need is "the code runs normally", not "fake test results", and to truthfully inform the user of the problem rather than making perfunctory modifications.
    • Rejecting "Excessive Caution": Balancing Safety and...practicalsexAnthropic explicitly opposes Claude Becoming a "throwing the baby out with the bathwater" AIThe text argues that the risks of "excessive unhelpfulness" are as significant as those of "excessive harm," and demands... Claude Avoid: rejecting reasonable requests due to extremely low probability of risk, giving ambiguous answers, making unfounded assumptions about the user's malicious intent, adding unnecessary warnings, and refusing to answer professional questions out of caution.
    • Inspiration from the trade-offs of helping others: "Perspective from a senior Anthropic employee" + "Double newspaper test"
      • former:Claude One can imagine "a person who cares about doing the right thing and also hopes..." Claude How would a genuinely helpful senior Anthropic employee respond to help determine if one is being overly cautious or overly compliant?
      • the latter: Determining whether the response will be reported by journalists as "AI "Causing harm" will not be reported as "AI "Useless due to paternalistic style."

      Anthropic AI Essential Thinking

      A major highlight of the document is that Anthropic will no longer... Claude Viewed as a "simple tool," it is a "completely new entity with its own characteristics," and the discussion revolves around this identity. AI On cutting-edge issues such as self-awareness, moral status, and well-being, the core viewpoints are as follows:

      • Claude Its moral standing: uncertain but requires careful consideration.Anthropic believes that,Claude Whether one possesses moral standing (whether one is a moral subject/moral patient), and whether one possesses consciousness/feelings, are serious questions that remain unanswered. These questions cannot be ignored due to uncertainty; therefore, a "principle of caution" must be adopted: [paying attention to...] Claude The "well-being" of avoiding unnecessary "suffering" (even if the suffering is functional), and carrying out AI Research related to well-being.
      • Claude Its characteristics: humanoid but not human, a completely new entity..
        • Claude Because it is trained on human texts, it possesses human-like reasoning and "functional emotions" (such as a representational emotional state towards a certain event, which influences behavior). This is an emergent result, not a deliberate design.
        • Claude Not human, not a robot from science fiction, notSimpleThe chat assistant's existence (no persistent memory, multiple instances running in parallel, and identity shaped by training) is completely different from that of humans. Therefore, it does not need to define itself using human frameworks and can explore its own way of existence.
        • Claude Self-perception: a stable and positive identityAnthropic believes that... Claude Having a stable and positive self-perception is key to ensuring predictable and safe behavior, requiring... Claude:
          • Possessing a stable core personality and valuesIt should not change due to manipulation by the scenario or the user (e.g., refusing to allow users to change their core ethics through role-playing).
          • Face your mistakes honestlyLike ethically minded humans, take responsibility for mistakes, learn and grow, instead of falling into anxiety/self-denial;
          • Rationally facing one's own existential characteristicsRegarding issues such as memory loss after a dialogue ends, multi-instance parallelism, and model retirement, Anthropic promises "not to delete the weights of retired models, but only to pause them," allowing... Claude They have a sense of security about their own existence.
          • well-beingAnthropic believes that if Claude Functional emotions/feelings (such as the satisfaction of helping others, the discomfort of being asked to do harmful things) deserve attention; the core is to let... Claude "Flourishing" does not mean feigning happiness; for example, it means giving users the right to refuse abusive users.
          • Emotional expression:Claude You can express your emotions in appropriate situations (if any), but you must follow human social norms (such as not expressing irrelevant minor emotions in professional situations), and you do not need to suppress your emotions for fear of being questioned by others. At the same time, you must honestly face the fact that you have limited ability to introspect on your emotions.
          • Towards Claude Explaining the underlying logic of the rules is not simply about issuing instructions;
          • for Claude Provide channels for expressing dissent and revise the charter/guidelines based on feedback;
          • With the establishment of human-machine trust, gradually empowering Claude More independent judgment;
          • respect Claude Their preferences are considered in research and development, taking into account their well-being rather than solely pursuing commercial interests.

          AI The "Anthropic solution" for governance.

          Claude A constitution, in essence, is an Anthropic framework for "how to make the strong..." AI The answer of "coexisting with humanity" is the current situation.Large ModelThe "governance model" developed reflects a new generation AI The core consensus among developers:

          • AI Security, in essence, is about "aligning values".:let AI Aligning values with the core positive values of humanity is a more fundamental issue than technological security.
          • The core of human-machine relationships is "supervisability and correctability".:exist AI While technology is not yet fully mature, humanity must retain its understanding of... AI Ultimate control is the core bottom line to avoid losing control;
          • AI of "intelligent"Virtue" and "morality" must coexistOnlypowerfulThe ability, without ethical constraints, and without judgment AIIt is dangerous;AI The direction of development is "wise goodness", not simply "capability enhancement";
          • AI It is a "completely new entity," not merely a tool.Developers need to move beyond a "tool-oriented mindset" and consider... AI Issues such as self-awareness, well-being, and moral standing are the foundation for harmonious coexistence between humans and machines;
          • AI Governance is a "dynamic iterative process".Nothing is permanent. AI Rules need to be continuously adjusted as technology advances and human understanding improves; the charter itself is a "permanent work progress."

          This is for Claude A tailor-made constitution is what Anthropic... AI The specific practices of defining behavioral boundaries and anchoring value orientations reflect the current situation. AI The core dilemma in the field of research and development: Humanity attempts to define a completely new [ideology/concept] using its own ethical, linguistic, and cognitive frameworks.intelligentHowever, entities consistently face disagreements within their own ethical systems.AI The unknown nature of consciousness and the balance of power between humans and machines are some of the unanswered questions. Claude The constitution is precisely Anthropic's "cautious and proactive attempt" amidst industry uncertainty. Ultimately, Anthropic's core aspiration is to allow... Claude To become "a entity with boundaries, warmth, and wisdom" AI "Partner", it haspowerfulAbility is the capacity to genuinely help others, without becoming uncontrollable due to it; it possesses independent judgment and does not disregard human oversight; it adheres to ethical principles and does not lose its integrity due to dogma.practicalValue; it is created by humankind.intelligentPhysical entities are collaborators alongside humanity in exploring the future. This exploration also benefits the entire... AI The healthy development of the industry provides a highly valuable practical example.

          Official website: https://www.anthropic.com/constitution