IA on AI – Claude's New Constitution

Summary

---

IA on AI – Claude's New Constitution

The Audit Podcast | Host: Trent Russell | January 28, 2026

---

Episode Overview

In this installment of "The Audit Podcast," host Trent Russell and guests discuss recent developments regarding Anthropic's AI model, Claude, focusing on the company's newly published "constitution" – a document outlining the guiding values and behavioral framework for Claude. The panel explores why this move is significant for AI governance and what internal auditors and organizations can learn from it, particularly regarding vendor management, risk, and control considerations around AI adoption.

---

Key Discussion Points & Insights

1. Who is Claude?

• Comparison with Other AI Models (00:16–00:42)
  • Claude, developed by Anthropic, is described as a major competitor to AI models like ChatGPT and Copilot.
  • The panel notes the ongoing debate in tech circles about which AI model is currently superior.
    • "Every couple of weeks there's a bunch of tests... somebody goes, oh, Claude's the best model ever... then two weeks later it's no, this ChatGPT wins." (B, 00:27–00:39)

2. Anthropic’s New Constitution for Claude

• Purpose and Content (00:42–01:21)

  • The constitution is a "detailed description of Anthropic's vision for Claude's values and behavior," serving a dual purpose of guiding development and defining operational standards.
  • Notably, it's integral to model training and directly influences how Claude behaves.
  • "The constitution is a crucial part of our model training process and its content directly shapes Claude's behavior." (C quoting from article, 01:10–01:21)

• Internal Audit & Vendor Governance Implications (01:21–02:10)

  • Strong recommendation for auditors and organizations: Request and review documented model behavior standards and evidence of testing from all AI vendors.
  • These standards should be part of core vendor governance procedures.
    • "At a minimum... require a documented model behavior standards and evidence of testing as well." (B, 01:21–01:35)
    • "There might even be... areas from this where you go, 'Hey, we should probably include this... relative to some kind of vendor governance procedures.'" (B, 01:52–02:10)

3. Risks & Governance Challenges of AI

• Emerging Risks Identified by Industry Leaders (02:10–02:50)

  • Cites a warning from Dario Amodei (Anthropic Co-Founder/CEO), who questions whether current human systems can manage the “almost unimaginable power” of modern AI.
    • "Questions of human systems are ready to handle the... 'almost unimaginable power,' that is 'potentially imminent.'" (C quoting Dario, 02:25–02:34)

• Key Risk Areas for Internal Auditors

  • Weak governance structures
  • Content abuse (e.g., deepfakes, particularly concerning HR processes)
  • Fraud exposure
  • Controls are lagging behind advancements in model autonomy
    • "On the audit side... weak governance structures, content abuse risks... deep fakes, especially when it comes to HR... fraud exposure... controls lagging behind relative to model autonomy improvements." (C, 02:35–03:04)

4. AI's Unpredictable Trajectory & Auditor Mindset

• Acknowledging Uncertainty (03:08–03:38)

  • Recognize the fast-evolving, unpredictable nature of AI technology.
  • Experts are unable to forecast where AI will be in 5–10 years.
    • "'Where do you think AI is going to be in five to ten years?' I have no idea. I don't think anybody... has a really good idea." (C/B, 03:16–03:30)

• Practical Guidance for Auditors (03:38–04:05)

  • Stay continuously educated on AI developments from balanced, non-partisan news sources.
    • "Find a source for AI news that is somewhere down the middle... and pay attention to it as much as you can." (B, 03:43–04:05)

---

Notable Quotes & Memorable Moments

• On Model Superiority Debate (B, 00:27)
  • "Every couple of weeks there's a bunch of tests... it's Claude's the best model ever in the history of existence. And then two weeks later it's no, this ChatGPT wins the best one ever."
• On AI Constitutions and Vendor Governance (B, 01:21)
  • "At a minimum, I would highly recommend... requiring a documented model behavior standards and evidence of testing as well."
• On AI Risks Identified by Industry Leaders (C quoting Dario, 02:25)
  • "Questions of human systems are ready to handle the 'almost unimaginable power' that is 'potentially imminent.'"
• On Uncertainties of AI’s Future (C/B, 03:23)
  • "I have no idea. I don't think anybody... would take that guess at this point as a guess."

---

Important Segments & Timestamps

• Who is Claude?: 00:16–00:42
• Anthropic's Constitution and Its Role: 00:42–01:21
• Audit & Vendor Governance Recommendations: 01:21–02:10
• AI Risks Cited from The Guardian / Dario Amodei: 02:10–02:50
• Controls Lagging vs Model Autonomy: 02:51–03:04
• Unpredictability and Staying Informed: 03:08–04:05

---

Conclusion: Takeaways for Internal Auditors

• Internal auditors should treat AI models like Claude as critical, high-impact vendors, requiring transparent documentation of behavioral frameworks and robust evidence of regular testing.
• Governance and controls must keep pace with rapidly advancing AI models to mitigate risks, especially around content abuse, fraud, and model autonomy.
• The episode emphasizes a continuous learning mindset: staying alert to balanced, forward-thinking news is essential for responsible audit and oversight of AI deployments.

---

Transcript

[A]

Welcome to the IA on AI Podcast, part of the Audit Podcast network where we bring you weekly updates on AI from the internal auditor's perspective.

[B]

Here we go.

[C]

From anthropic.com Claude's new constitution.

[B]

Those who don't know because Claude doesn't get as much mention as Chad, GBT and Copilot. Claude's a competitor.

[A]

They are right there.

[B]

They're one of those where every couple of weeks there's a bunch of tests that are done across all these models and somebody goes, oh, Claude's the best model ever in the history of existence. And then two weeks later it's no this chat. GPT wins the best one ever. So Claude is right there in the.

[C]

Mix with a lot of them. What I wanted to point out from the article, it says we're publishing a new constitution for our AI model Claude. It's a detailed description of Anthropic's vision for Claude's values and behavior. A holistic document that explains the context in which Claude operates and the kind of entity we would like Claude to be. Then lastly, at least for lastly, from what I'm going to read, there's significantly more to the article, but the constitution is a crucial part of our model training process and its content directly shapes Claude's behavior. The reason I wanted to point that out is at a minimum, I would highly recommend considering even a at a.

[B]

Minimum relative to any kind of vendor governance, AI governance relative to third parties requiring a documented model behavior standards and evidence of testing as well. Considering that last bit that I mentioned where it says and its content directly shapes Claude's behavior, the fact that these things do behave as a way to think about it, I think makes it even more important to understand what's going on. Sufficient, highly sufficient evidence of testing and again those model behavior standards. So I would take a look at this article, flip through it and then there might even be some areas from this where you go like hey, we should probably include this or request this relative to some kind of vendor governance procedures that you guys do.

[C]

From the guardian.com it says, Wake up to the risks of AI. They are almost here.

[B]

Anthropic boss warns so we just talked.

[C]

About Claude and then it goes on to say Dario, which is their co founder, chief executive Questions of human systems are ready to handle the quote, almost imaginable power, end quote. That is quote, potentially imminent, end quote.

[B]

This is maybe even more so on.

[C]

The personal side rather than the internal audit professional side. But on the audit side, consider things that's probably, maybe, hopefully obvious obvious at this point. Weak governance structures.

[B]

Content abuse risks.

[C]

So things like deep fakes, especially when it comes to hr, are going to be really important as well as fraud exposure as well. And then there's just controls lagging behind relative to model autonomy improvements.

[B]

Anyway, so those are things I think.

[C]

To keep in mind from the audit.

[B]

Side and then just personally keep up with what's going on. Nobody really knows. I did a webinar a couple weeks ago and someone said, where do you think AI is going to be in five to 10 years?

[C]

I have no idea.

[B]

I don't think anybody who would take that guess at this point as a guess. I think there's very few researchers that.

[A]

Have a really good idea of where.

[C]

That could potentially be, but they're probably going to miss the mark also. So this is more of an announcement.

[B]

To continue to stay educated.

[A]

Hopefully what we're offering.

[B]

I know we haven't been super consistent through December and January, but on a pretty consistent basis when we drop these on Wednesdays every week, that it's somewhat helpful. But find a source for AI news that is somewhere down the middle as opposed to one side or the other on the political spectrum, and pay attention to it as much as you can.

[A]

Thank you for listening and be sure to follow the link to greenskiesanalytics.com in the show notes and schedule time to see how green skies can make the hype of AI a reality in your internal audit department.