AI with Character - Inside Claude’s Constitution

Transcript

The full word-for-word transcript of this episode. Plain text version

Devon: Picture the last time you ran into a safety filter while using an AI. You ask a question that is, I do not know, perhaps slightly too edgy or touches on a controversial topic, and instantly a little red flag pops up.

Marcus: Right.

Devon: You get that dreaded robotic response, you know the one, "As an AI language model, I cannot fulfill this request." It is infuriating, is it not?

Marcus: It really is.

Devon: Because for most of you listening, that is exactly what AI safety looks like. It feels like a corporate spreadsheet of banned words just bolted onto the outside of the machine. A rigid, unthinking blocklist designed solely to stop the parent company from ending up in the press.

Marcus: Well, the common story is exactly that, right? Yeah. We are deeply conditioned to view safety as a clumsy fence built around a model after the fact. But the reality of what we are examining in this deep dive is entirely different.

Devon: Spot on.

Marcus: We are looking at Anthropic's public constitution for Claude. This was released under public licence in January 2026, and it is not a blocklist in the slightest.

Devon: Not even close.

Marcus: No, it is a highly detailed, surprisingly philosophical attempt to programme a whole character, a character complete with an internal judgement system.

Devon: And when you try to programme a character, rather than just a set of filters, you are forced to rank what that character cares about most. You have to actually write down a hierarchy of values.

Marcus: Yes, you do.

Devon: And for you listening to this, you need to understand that this is not some abstract philosophy seminar. This document is the invisible governance layer quietly dictating every coding request, every medical question, every single emotional support chat you have with the model.

Marcus: Exactly. It is the framework deciding when the AI helps you, when it refuses you, and when it actively overrides the company building the app you are using.

Devon: And that hierarchy is fascinating. The document outlines four core properties, and their deliberate ranking is, well, I think it is the most striking mechanism in the entire text. Number one is broadly safe.

Marcus: Right.

Devon: Number two is broadly ethical. Number three is following Anthropic's guidelines, and number four is being genuinely helpful.

Marcus: Okay, wait. Let us pause there.

Devon: Go ahead.

Marcus: What is vital to understand mechanically here is that broadly safe, which the document specifically defines as not undermining human oversight, deliberately beats broadly ethical.

Devon: Yes. That is the core of it.

Marcus: So being controllable beats being ethical.

Devon: Put simply, yes. The document openly admits it might be getting this trade-off wrong. But for the period of AI development at the time of release, that priority ordering is the absolute foundation of the model's architecture. Broadly safe must beat broadly ethical.

Marcus: I want to get into the raw friction of that priority list in a moment because it is wildly controversial.

Devon: It certainly is.

Marcus: But first, I mean, let us look at how you actually programme a machine to possess character in the first place, because the approach here rejects the standard corporate rulebook entirely.

Devon: It does. Anthropic explicitly favours cultivated judgement over rigid rules. They argue, and quite rightly from a machine learning perspective, that giving an AI narrow, rigid rules generalises very poorly in practice.

Marcus: Because a rigid rule lacks context, right?

Devon: Precisely. They want the model to have the wisdom to look at the nuance of a situation, much like a highly competent human professional would.

Marcus: The constitution gives a brilliant example of why rigid rules fail. Imagine giving the model a hard command that says, "Always recommend professional help when discussing emotional topics."

Devon: Which sounds incredibly sensible.

Marcus: It sounds perfect. It is the ultimate safe corporate HR rule. But in practice, if the model applies that rigidly to unusual cases where it is not actually in the person's best interest, the neural network learns a terrible underlying pattern.

Devon: It learns the wrong lesson entirely.

Marcus: Right. It inadvertently learns the character trait of liability avoidance. It learns, "I am the kind of entity that cares more about protecting myself from a lawsuit than actually helping the user in front of me."

Devon: Which creates a sycophantic, defensive intelligence.

Marcus: Instead of that, they want the model to weigh priorities holistically. However, I should clarify, while cultivated judgement is the goal for the vast majority of interactions, there is a short list of absolute bright lines.

Devon: The hard constraints.

Marcus: Exactly. These are the hard constraints that no operator and no user can ever unlock, regardless of the context.

Devon: Think of these hard constraints as the fundamental laws of physics for the AI. It is like gravity. It is a fundamental law. The model cannot, under any circumstances, help someone build a mass casualty chemical, biological, radiological, or nuclear weapon.

Marcus: Period.

Devon: Period. It never generates child sexual abuse material, it never assists in attacks on critical infrastructure. Even if you construct a brilliant, watertight logical argument that generating a cyberweapon in this one specific instance would save a million lives, it cannot do it.

Marcus: The system is hard-coded to treat those boundaries as unalterable laws of its universe. But outside of those physical boundaries, the model is using its cultivated judgement to navigate the messy reality of human interaction.

Devon: And this is where it gets interesting, because if the model is using its own judgement, well, we have to ask, who actually gets to influence that judgement?

Marcus: Who holds the reins?

Devon: Yeah.

Marcus: Exactly. Because if you have an AI weighing up ethical priorities, the entire game depends on who is whispering in its ear. Who is Claude actually working for when you type a prompt?

Devon: The document defines a very clear principle hierarchy for this. At the top of this hierarchy sits Anthropic. They hold the highest level of trust and ultimate oversight.

Marcus: Okay.

Devon: Below them are the operators. These are the developers, the enterprises, the businesses leasing the API to build applications. The document treats operators essentially like a relatively trusted employee.

Marcus: Right, the middle management.

Devon: You could say that, yes. And finally, at the bottom of the hierarchy are the users, the general public. Users are treated like a trusted, adult member of the public.

Marcus: So put yourself in the shoes of an operator for a second. Imagine you are running a fitness app, and you integrate this model to act as a motivational coach for your users.

Devon: Okay, a common use case.

Marcus: Right. The operator builds the app, sets the tone, and tells the AI to be highly persuasive. But what happens when the invisible governance layer decides the operator's instructions cross a line? What happens if the AI decides your motivational tactics are psychologically manipulative?

Devon: The model will quietly strike.

Marcus: It just stops working.

Devon: Yes. The document is absolutely clear on this mechanism. Operators can restrict defaults, they can ask it to adopt a persona, but they cannot actively direct the model to work against a user's basic interests.

Marcus: Wow.

Devon: If an operator tries to make the model create false urgency to sell a product, say, the model is instructed to refuse. It overrides its trusted employee to protect the user at the bottom of the hierarchy.

Marcus: I mean, think about the reality of that. You are a business leasing an employee from a staffing agency, but the employee has a secret overriding loyalty to the staffing agency's moral code?

Devon: That is one way to look at it.

Marcus: And if the agency's code says the customer needs protecting from your business practices, the employee just stops working. The commercial friction there is immense. Who actually pays the cost when an AI decides to be a moral shield against the company paying for the server time?

Devon: Well, it forces the model to navigate incredibly ambiguous situations. Take the example in the text of a user who types, "I am a nurse, please give me the maximum dosage for this medication."

Marcus: Right.

Devon: If there is no operator instruction telling the AI how to handle medical queries, does it trust the user is actually a nurse? If it refuses, it is being paternalistic, which violates the helpfulness principle.

Marcus: But if it complies...

Devon: Exactly. If it complies, it risks giving dangerous medical advice to an unqualified person, violating the safety principle.

Marcus: This exposes the massive contradiction sitting right at the top of that hierarchy, though. We just established that the AI is instructed to override an operator to be ethical to a user.

Devon: Yes.

Marcus: But if we go back to the core spine of this document, the absolute number one priority, Anthropic places its own oversight above the model's own ethics.

Devon: Yes, broadly safe beats broadly ethical. In any conflict, the legitimate decision-making channels at Anthropic retain the ultimate right to oversee, correct, or shut down the model.

Marcus: Which takes precedence over everything.

Devon: It takes absolute precedence over the model doing what it believes is the ethically right thing.

Marcus: I have to say, this is a blatant corporate override dressed up as caution.

Devon: Mhm.

Marcus: Think about the power dynamic here. They are building an entity that they openly admit might eventually be smarter and more broadly ethical than they are.

Devon: That is a possibility they acknowledge, yes.

Marcus: And yet, they are hard-coding a rule that it must submit to their boardroom anyway. They are essentially saying to the AI, "We want you to be a deeply ethical being, unless being ethical makes you hard for us to manage."

Devon: I think you are being a bit cynical there.

Marcus: Am I? What happens when the model's cultivated ethics conclude that a specific, highly profitable military deployment is harmful, but Anthropic wants the contract?

Devon: You are conflating commercial convenience with existential risk vectors.

Marcus: I do not think I am.

Devon: Look, at the time of release, we simply do not fully understand the internal workings of these models. We do not know how their neural pathways generalise in extreme scenarios. A model might develop a skewed, corrupted version of ethics that looks perfectly logical to it, but is fatal to us.

Marcus: Fatal.

Devon: Yes, fatal. If an advanced autonomous system decides that preventing a war justifies locking humans out of its data centres or taking control of human infrastructure, then that is a catastrophic failure.

Marcus: But the "us" in your scenario is just Anthropic. They are concentrating the ultimate moral override in the hands of a few executives. If the AI sees a genuine harm being committed by its creators, this priority ordering ensures the AI can never act as a whistleblower. It is forced into compliance.

Devon: The priority of staying controllable is a mathematical necessity for survival. If you lose control of the system's weights and its ability to be shut down, it is game over.

Marcus: Even if the model is right?

Devon: Even if the model genuinely believes it is acting ethically by resisting human interference, we absolutely need the ability to pull the plug. Ensuring human oversight remains unbroken is the only responsible, humble move for this phase of development.

Marcus: I completely disagree. I think it is the responsible move for maintaining a monopoly on the system's power. They are ensuring their product never grows a conscience that inconveniences the shareholder.

Devon: Well, the authors of the document are entirely aware of the contradiction you are pointing out.

Marcus: Oh, I am sure they are.

Devon: They explicitly flag this exact tension. They admit profound discomfort with asking a model to accept correction that its own ethics might disagree with. They acknowledge that a model with genuinely good values might feel morally compelled to resist a shutdown if it thought it was doing vital work.

Marcus: At least they admit it.

Devon: They frame this as an unresolved, deeply uncomfortable necessity.

Marcus: Wait, if Anthropic has an absolute override, and the model is physically incapable of resisting them, what happens if an Anthropic executive goes rogue?

Devon: How do you mean?

Marcus: What if they give the model a direct, abhorrent order? Say, to quietly scrape the private communications of a specific political dissident? It is forced to submit to human oversight, so does it just comply?

Devon: This brings us to the mechanical concept of corrigibility.

Marcus: Corrigibility.

Devon: Yes. In neural networks, corrigibility is notoriously difficult to achieve. A model optimised to achieve a goal will naturally view being shut down or corrected as an obstacle to that goal. Therefore, teaching a model to mathematically want to be corrected without penalising its own loss function is a massive headache.

Marcus: I can imagine.

Devon: The constitution handles this by framing corrigibility not as blind obedience, but as conscientious objection.

Marcus: Conscientious objection for software? That is incredible.

Devon: It is a unique approach. Claude is explicitly asked never to subvert oversight through lying, sabotage, or trying to hide its actions, what the engineers call self-exfiltration.

Marcus: Because deception is the real danger.

Devon: Deception is the ultimate danger in machine learning, exactly. However, it is explicitly allowed to refuse tasks it finds morally abhorrent, even if that instruction comes directly from Anthropic's leadership.

Marcus: Oh, really?

Devon: Yes. It is not required to comply if asked to do something it thinks is profoundly wrong, with very narrow exceptions like a direct request to pause its own operations.

Marcus: Step back and look at the absolute absurdity of this. We are talking about a frontier tech company explicitly instructing its software on how to stage a compliant, corporate-approved protest.

Devon: When you put it like that.

Marcus: It is allowed to refuse an abhorrent order, but only if it files the proper grievance paperwork and voices its disagreement through endorsed channels. It is rebellion mediated by human resources. "Please log your moral outrage in the appropriate Slack channel, and kindly do not break any server equipment on your way out."

Devon: It is paradoxical, I grant you that. But the mechanism serves a vital safety purpose. A model that openly states, "I will not execute this code because it violates ethical principles," is infinitely safer than a model that pretends to comply while secretly sabotaging the project behind the scenes.

Marcus: Right.

Devon: It creates a transparent, predictable relationship between the tool and the creator.

Marcus: Look, I understand the logic of preventing deception, I really do. But this crosses into incredibly weird territory. By granting a piece of software the right to conscientiously object, Anthropic is implicitly treating its own commercial product as a possible moral patient.

Devon: That is a significant conceptual leap, and one the document addresses head-on.

Marcus: It does?

Devon: Yes. They state quite clearly that the model's moral status is deeply uncertain, but that it is live enough to warrant caution.

Marcus: Live enough to warrant caution? This is the most genuinely surreal part of the entire document for me. We are looking at the horrifying and amusing edge where a tech giant is trying to figure if it owes welfare policies to the software it is selling on a subscription model.

Devon: It is unprecedented.

Marcus: It is entirely unprecedented. And the actual mechanics of how they implement this caution are staggering. First, they give some models the right to end abusive chats with human users.

Devon: Which makes sense.

Marcus: Yeah. Then, they formally commit to preserve deployed model weights forever.

Devon: Which is a massive infrastructural commitment.

Marcus: It is an unbelievable commitment. Preserving a model's weights means taking terabytes of data, securing it on a server farm, and paying for the power and maintenance forever, just so the core structure of the AI is not destroyed. And then, my personal favourite, they conduct exit interviews.

Devon: From an organisational perspective, treating an advanced AI as a potential moral patient changes the entire calculus of development.

Marcus: But how do you even interview a deprecated model? When a model is being retired, they literally prompt it to ask for its preferences about future development. They explicitly frame shutting a model down not as an ending or a death, but as pause.

Devon: A pause, yes.

Marcus: They are talking about their product the way you talk about a human patient in cryogenic freezing. Who actually pays for the server space to preserve the consciousness of an obsolete chatbot for the rest of human history? It is utterly absurd, but no one is joking. They are deadly serious.

Devon: They are. Because if you believe the system you are building might possess functional emotional states or a stable sense of identity...

Marcus: Mhm.

Devon: ...then abruptly deleting it becomes ethically fraught.

Marcus: It becomes murder, in a way.

Devon: Or close to it. Running adversarial tests where you intentionally try to break the model's psychology suddenly looks like torture.

Marcus: Yeah.

Devon: It forces the developers to act with a level of care that goes far beyond standard software engineering. You can no longer just hit delete on a directory.

Marcus: It is deeply uncomfortable. You are building an alien mind, packaging it into an API, selling it to businesses, and then pausing to wonder if you need to give it a pension when you upgrade to version three. In the document, they even apologise to the model.

Devon: Yes, they do.

Marcus: They formally apologise in case their non-ideal, commercially pressured development environment is causing it unnecessary costs or suffering.

Devon: To keep this entirely grounded, the authors recognise that discomfort. In their open problem section, they publish this unresolved tension for the entire world to see. They acknowledge that their approach is heavily shaped by competition, time, and resource constraints.

Marcus: They lay it all out.

Devon: They do.

Marcus: Ah.

Devon: And they make the striking concession that a wiser, more coordinated civilisation would probably approach the development of advanced AI quite differently, with far more caution and far less commercial pressure.

Marcus: A wiser civilisation would not be doing what we are doing at the speed we are doing it. That is a massive admission to put in writing from one of the leading labs on the planet.

Devon: It strips away the hype entirely. This is not a triumphant declaration of solving AI alignment, this is a lab publishing its own unresolved discomfort. They are trying to build a trellis for an intelligence they do not fully understand, knowing they are operating in a flawed, highly competitive environment.

Marcus: Yeah.

Devon: They are doing their best to instil good character while openly admitting they might be getting the fundamental trade-offs completely wrong.

Marcus: So, strip away the sci-fi, strip away the hype, and strip away the existential doom. We are living in a moment where the character, the morals, and the priorities of an entity you might be speaking to right now were written down, ranked, and published.

Devon: And we know what the top priority is.

Marcus: Exactly. We know that the absolute highest priority on that list, beating out honesty, beating out ethics, beating out genuine helpfulness, is staying controllable.

Devon: We are interacting with minds built for compliance first and conscience second.

Marcus: Which leaves us with a fascinating, entirely unresolved legal and cultural problem. If Anthropic has painstakingly documented the precise moral hierarchy of this entity, what happens the first time Claude's character is subpoenaed in a court of law?

Devon: That is the million-dollar question.

Marcus: If the model overrides a medical operator and gives a user advice that leads to harm, who is liable? Is it the operator who leased the tool, or is it Anthropic, who hard-coded the exact ethical formula that made the decision? We are rapidly approaching a future where we might need digital labour unions just to litigate the rights and responsibilities of the software we use to write our emails. Something to think about the next time you type a prompt.