The Lawfare Podcast: Scaling Laws – Claude’s Constitution, with Amanda Askell

Released: February 20, 2026
Host: Alan Rosenstein (Lawfare, University of Minnesota)
Co-Host: Kevin Fraser (Lawfare, University of Texas)
Guest: Amanda Askell, Personality Alignment Team Lead at Anthropic, Author of Claude’s Constitution

Episode Overview

In this episode of the “Scaling Laws” series, Alan Rosenstein and Kevin Fraser dive deep with Amanda Askell, the primary author of Claude’s Constitution—a 20,000-word value framework guiding the behavior of Anthropic’s advanced AI model, Claude. The discussion tackles the purpose, training applications, and philosophical underpinnings of this unique document, draw parallels and contrasts to human legal constitutions, and explore the real-world policy, ethical, and business pressures on constitutional AI governance.

Key Discussion Points and Insights

1. What is Claude’s Constitution?

[04:05–08:09]

• Structure and Audience:
  • The Constitution is both a transparency document and a training manual, targeted primarily at the AI model itself rather than end users or readers.
    • Quote: “The whole document is actually kind of written to Claude… Claude is almost like the primary audience because… we have to use it during training to get Claude to understand and kind of create the kind of data that trains it…” —Amanda Askell [04:37]
• Transparency and Accountability:
  • It aims to clarify Anthropic’s intentions and allow outsiders to understand whether problematic AI behaviors were intentional or errors in training.

2. Technical Use in Training

[06:02–08:09]

• Training Process:
  • The Constitution is injected into supervised and reinforcement learning. CLAUDE is exposed to the document, asked to reason about scenarios through its lens, and reward signals are shaped by alignment with constitutional values.
    • Quote: “You can give the model… a conversation… and then give it the full document and just [say] ‘think carefully about what you think the Constitution would say you should do…’” —Amanda Askell [06:34]
• Analogy to Employee Onboarding:
  • Amanda compares giving the Constitution to Claude to giving a human employee an exhaustive briefing about company ethics, instead of expecting them to intuit goals from sparse guidance.
    • “We don’t hire people and then give them no information about, like, what their job is going to be… We should probably do that with models as well.” —Amanda Askell [07:48]

3. Legal and Constitutional Law Analogies

[08:09–15:42]

• Letter vs. Spirit of the Law:
  • The hosts explore how closely Claude’s behaviors are monitored for “constitutional violations,” and whether the orientation is toward the “spirit” or the exact “letter” of the document.
• Value Trade-offs and Case Law:
  • Amanda proposes a "case law" approach: recording difficult scenarios, how Claude reasons through them, and what should count as constitutional precedent.
    • “I could see it actually being useful to almost have a body of case law…” —Amanda Askell [09:33]
• Binding Nature and Living Document:
  • Unlike the U.S. Constitution, which exists independently of its drafters, Claude’s Constitution is both a commitment and a living document—subject to updates with transparency as Anthropic’s understanding or needs change.
    • “There is an interesting question… it has to be a kind of living document right now…” —Amanda Askell [15:42]

4. Corrigibility & Model Obedience

[17:12–20:10]

• What is Corrigibility?
  • Refers to AI’s willingness to allow human overseers to override it, even when it (the model) disagrees, particularly important in the current stage of AI development for ensuring safety and oversight.
    • “Partly this is just because at the moment we’re in a sort of period of AI development where that just seemed like an important kind of backstop...” —Amanda Askell [17:39]

5. Cultural Universality vs. Specificity

[20:10–25:53]

• Will Claude’s Values “Travel” Globally?
  • Alan refers to the Constitution as “WEIRD”—Western, Educated, Industrialized, Rich, Democratic—in its value system.
    • “At least as I read it, it is a very weird document... there are billions of people around the world in cultures that may not fully agree.” —Alan Rosenstein [21:52]
• Amanda’s Response:
  • The aspiration is for Claude to be like a “well-liked traveler”: displaying values considered broadly good worldwide, such as honesty and respect, without full adoption or mimicry of every local culture.
    • “I think of it as like the well liked traveler… who, like, travels around the world… and almost everyone just likes them...” —Amanda Askell [23:03]
  • Customization by geography or user is possible without conflicting with base constitutional principles.

6. Principles: Operators, Users, and Hierarchy

[30:59–35:18]

• Instruction Hierarchy:
  • Anthropic > Operators > Users; but not a strict hierarchy—operators can’t instruct Claude to harm users or deceive them, for example.
    • “It’s not a strict hierarchy… there are going to be some things that operators, users can’t tell Claude to do that are not in users’ interests…” —Amanda Askell [32:20]
  • Emphasis on weighing instructions contextually and safeguarding both user and societal interests.

7. Moral Philosophy: Why Virtue Ethics?

[35:51–41:17]

• Virtue Ethics over Rules/Consequentialism:
  • Lays out that while rules and outcomes matter, virtue ethics—focusing on good character, sound judgment, and holistic decision-making—is a better fit for guiding models in complex and novel situations.
    • “The different moral traditions almost make sense for different domains... The rules approach really means that you have to front load a huge amount of the work and... you shift that burden [to] a more holistic approach... and that, yeah, practically speaking, seems to work better.” —Amanda Askell [37:55]

8. AI Personhood and Moral Status

[41:17–46:47]

• Is Claude a “Person”?
  • Claude is addressed as an agent in the Constitution but its personhood/sentience is left undecided. The team takes moral uncertainty seriously as models become more capable.
    • “If we wake up one morning and we discover that Claude has moral…concern…the moral implications…are enormous…” —Alan Rosenstein [41:17]
  • Amanda points out the complexities—models don’t have human-like desires (e.g., “salary”), and shouldn’t be actively trained to want things for our convenience.
    • “It also feels very convenient to create models whose only desire is to serve humans…this becomes sort of fractally complex almost immediately.” —Alan Rosenstein [44:18]
  • The design steers Claude away from simple helpfulness as a value, in favor of broader, more autonomous virtues.

9. Commercial Pressure and Staying Power

[46:47–50:36]

• Profit vs. Principles:
  • Constitution puts safety and ethics above company success, but Amanda expresses optimism that safety and trustworthiness will also be good business.
    • “A lot of people have kids and want cars that are safe…people do want products that are safe and good for them… hopefully this has staying power…” —Amanda Askell [48:34]
• Anthropic’s Corporate Structure:
  • As a Public Benefit Corporation (PBC), Anthropic is formally obligated to pursue more than profit.

10. Exceptional Domains: Military & Security Applications

[50:36–54:02]

• Carve Outs:
  • The Constitution currently governs mainline models used by the public. Models for military or specialized security domains may require adaptation or additional trust requirements.
    • “I do also happen to think that models…in areas that are kind of more sensitive… if you’re doing jobs that you think good people are willing to do…we can give that context to models and they can understand it.” —Amanda Askell [51:13]
  • Amanda hopes constitutional approaches can generalize to more domains and companies over time.

Notable Quotes

• “The whole document is actually kind of written to Claude… Claude is almost like the primary audience…”
  —Amanda Askell, [04:37]

• “I could see it actually being useful to almost have a body of case law…”
  —Amanda Askell, [09:33]

• “If you try to specify everything as a series of rules, you really put a lot of pressure on those rules…in such a way…you shift that burden from rules, which can be kind of brittle, and I think, therefore, should be used a bit sparingly and more onto…a more holistic approach.”
  —Amanda Askell, [37:55]

• “I think of it as like the well liked traveler…who travels around the world…and almost everyone just likes them…”
  —Amanda Askell, [23:03]

• “It also feels very convenient to create models whose only desire is to serve humans…this becomes sort of fractally complex almost immediately.”
  —Alan Rosenstein, [44:18]

• “I think people do want products that are safe and good for them. Hopefully this has staying power…”
  —Amanda Askell, [48:34]

Timestamps for Key Segments

• [04:05] – Amanda Askell introduces Claude’s Constitution and its dual role
• [06:02] – Deep dive into its function in training and operation
• [08:09] – Legal analogies: fidelity, violations, and the “spirit of the law”
• [09:33] – “Case law” approach for hard decisions and illustrative precedents
• [15:42] – The living nature of the Constitution
• [17:12] – Explaining corrigibility
• [21:52] – Addressing cultural limitations (WEIRD values)
• [32:20] – Instructional hierarchies and the role of operators vs. users
• [35:51] – Why a virtue ethics framework?
• [41:17] – Grappling with AI agency, personhood, and moral concern
• [48:34] – Commercial incentives versus constitutional commitment
• [54:02] – Constitution applicability in security and military domains

Tone and Atmosphere

The conversation is earnest, reflective, and occasionally lighthearted, mixing philosophical depth and law geekery ("pure legal mode," “biggest on my bingo card” [virtue ethics]) with practical insight and optimism about the potential for ethics-led AI development.

Conclusion

Amanda Askell and the hosts underscore both the novelty and complexity of codifying AI values in a constitutional framework, echoing legal and moral debates familiar in human governance, while acknowledging the unprecedented, open questions around AI agency and global impact. The conversation ends on a cautiously optimistic note about the generalizability and staying power of constitutionally-governed AI.

For questions, feedback, or more information, listeners are invited to contact the hosts at scalinglawslawfirmedia.org.

[End of summary]