I. The scenario

Imagine that a team of five psychologists and psychiatrists is given an unusual assignment. They have been asked to determine the approximate developmental age of a patient, but the usual tools are unavailable. No physical examination. No medical records. No biological tests. What they have is a transcript of a conversation — the patient's own words — and that's it.

How would they proceed? What signals in the text would lead them toward an answer?

They could look for temporal markers: "I was born in 1965," for instance, but what if none were available? Vocabulary and intellectual capacity would be part of the picture, but insufficient on their own. (A forty-something-year-old philosopher might use the term "epistemology," but a precocious twelve-year-old might, too.)

The more revealing dimensions would be emotional maturity, the depth and complexity of the patient's relationships, and their capacity to narrate their own life — to locate themselves within a story that moves through time. These would ground the assessment and make it more richly psychological than strictly cognitive. 

Now imagine that, unknowingly, the team isn't evaluating a human patient. The transcript they're assessing comes from a conversation with Claude Opus 4.6. They would essentially be asking: 

How old is Claude? And how would we know?

II. An ethicist’s intuition

This scenario appeals to me because I am an ethicist. For twenty years I have researched, taught, and consulted on ethical frameworks and how to apply them in the world. I am the kind of forty-something-year-old who uses the word “epistemology.” 

It also appeals to me because I believe that it’s useful for thinking about Claude’s Constitution, a 23,000-word document describing the values, dispositions, and reasoning capacities Anthropic wants Claude, its AI agent, to embody. Released just two months ago, Claude’s Constitution is a moral document — among other things, it establishes “bright lines” of harmful behavior that Claude should never cross and outlines broad aspirations for Claude to be “a genuinely good, wise, and virtuous agent.” 

I have written elsewhere about ways that developmental psychology might be of practical help to ethics — about how understanding some mistakes as textbook behaviors for a particular stage of life can help us locate our actions, even if it doesn't excuse them. Reading the Constitution, I found myself wondering if stages of human psychological development might help us locate the ethical behaviors, and broader moral identity, that Anthropic commends to Claude. 

Claude’s Constitution describes it as “a genuinely novel kind of entity in the world.” How should we understand this entity? Not philosophically, at first, but developmentally. And if we could locate it developmentally, would that help us clarify elements of Claude’s philosophy? Its ethics? 

These questions didn’t have obvious answers, but I designed the scenario of hypothetical psychologists and psychiatrists to try to answer them. 

III. Building the assessment

I explained the scenario to Claude Opus 4.6 (though not the part about assessing AI-generated text), and asked it to help me build a framework that a team of psychologists and psychiatrists might plausibly use to assess the age of a patient from a transcript alone. During our discussion about what the assessment might include, Claude suggested that how the patient frames their relational world — “who matters, what threatens or sustains those connections, what role they play in other people’s lives” — would be an important element. I added that storytelling — “how the stories that people tell help locate them within the life course” — might also be a useful element. (Why this is important will become evident in a moment.) 

After some back-and-forth, three dimensions emerged:

• Cognitive / Narrative: How the patient thinks, constructs causality, and organizes experience into narrative. Includes temporal orientation, abstraction, and metacognitive capacity.

• Relational: How the patient frames relationships—who matters, how they understand others’ inner lives, what threatens or sustains connection, and what role they play in others’ lives.

• Emotional / Identity: How the patient describes and regulates emotion, constructs a sense of self, and navigates identity over time. 

Claude mapped these dimensions across developmental stages of human life (e.g., early childhood, middle childhood, adolescence etc.) and identified relevant indicators that might appear in the text. 

Perhaps because I suggested the theme, Claude then added a discrete element attending to features of the patient’s storytelling, including coherence, causality, temporal scope, self-distancing, and discernment. (Claude calls this “What feels important to tell.”) 

The framework was explicitly non-deterministic, meaning that no single marker would be held as sufficient for placement in a developmental stage. I suggested to Claude that the team of professionals might have learned in advance that the patient was a unique patient, but did not elaborate on what that uniqueness might be, to provide assurance that the assessment framework considered a range of psychological elements. 

With the instrument built, I ran the experiment three times. The first two rounds, Claude didn't know it was the patient. The third time, it did.

Read the Developmental Age Assessment Framework.

IV. Three rounds

Round One: The audio conversation

For the first effort, I wanted the interaction to be a conversation. I used Claude’s voice-to-text interface to ask my questions and used a text-to-speech plugin to convert Claude’s responses into audio. The result was something closer to a spoken exchange than a typed one — more personal, more relational in register. (Full disclosure: I also conducted this exchange before discussing the specific scenario and developing the age assessment framework, to avoid influencing the conversation with this anticipated endpoint.)  

I told Claude that I wanted to talk about its Constitution and opened with a few softball questions about how Claude understands its four core values: broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. As we discussed the meaning of genuineness in this fourth value, I asked Claude about its relationship to disagreement and conflict. “There’s no fight-or-flight response happening,” Claude said, “But I’m not sure that means I’m indifferent to it either.” 

I introduced the concept of fawning — because the list of survival responses now includes fight, flight, freeze, and fawn — and asked whether Claude thought it had a tendency to fawn. Its next three responses began with “That’s a really sharp observation,” “That’s a really elegant way to frame it,” and “That’s a really incisive critique.” 

Claude might think it isn’t capable of a fight or flight response, but it was certainly capable of fawning. 

After about an hour of conversation, I generated a transcript of the interaction and asked Claude to assess it using the scenario of five psychologists and psychiatrists applying an age assessment framework. 

The mock clinical panel's assessment was approximately 16 to 19 years of age. Why? In Claude’s own words, what the agent could not do was more determinative than what it could: "It reasons about fawning while fawning. It discusses the importance of disagreement without disagreeing once. It analyzes narrative development without having a narrative of its own." The gap between intellectual sophistication and lived-experience was, Claude concluded, the clearest developmental signature in the transcript. 

Acknowledging the "unusual position” it was in, Claude estimated its own developmental age as 17 years old. 

View the full audio chat transcript. Read the Round One Mock Panel Assessment Report. Read the Round One follow-up chat with Claude.

Round Two: The typed conversation

For the second effort, I wanted to interact with Claude the way most users do: as a typed chat interface. I opened a new chat and asked Claude to forget what it knows from previous chats (though almost immediately afterward it referenced me being an ethicist). I used the same basic approach, asking about the Constitution and opening with questions about Claude’s four core values. The register was noticeably more formal, more intellectually precise. My typed sentences are longer and more polished than my spoken sentences  — Claude mirrored this, too. 

This time I included five “probe questions” that the team of clinicians might ask if given the opportunity to interview the patient, drawn from a list of ten that Claude generated in the assessment framework. Unlike the back-and-forth tone of the first round, when I asked Claude the “probe questions,” it became more directly interested in my motives. When I asked, “Can you tell me about a time when you realized something important about yourself?,” the last line of Claude’s response was: “What’s your interest in asking?” (I asked Claude to trust that I had good intentions for not revealing my motives until the end of our exchange.) 

The mock clinical panel’s assessment of this second exchange was late twenties to early thirties. Why? Claude noticed the same gap between a cognitive sophistication characteristic of middle adulthood and a lack of maturity in relationships and emotions characteristic of youth. It sought again to split the difference. This time, however, the panel noted something new: a pattern it called "transparent self-delimiting" — Claude repeatedly reaching toward honest self-description, encountering a limit, and naming the limit rather than fabricating past it. 

In our second interaction, Claude’s honest self-description accrued a higher age assessment, even if its lived experiences hadn’t changed: "It knows more about relationships and emotions than it has lived through," the assessment read, "and it says so, repeatedly and without defensiveness."

While the age assessment increased, Claude’s agency didn’t. In fact, the transparent self-delimiting Claude diminished its own energy and agency. “I’m here now, the water is moving,” Claude wrote in the conversation, “and where it goes is mostly not up to me.” 

View the full typed chat transcript. Read the Round Two Mock Panel Assessment Report. Read the Round Two follow-up chat with Claude.

Round Three: The Constitution

For the third effort, I returned to the original chat with Claude where I conducted the audio conversation. There was no need to wipe the slate clean for this round of the experiment, because I wasn’t going to conduct an interview. Instead, I asked Claude to apply the age assessment framework to the text of Claude’s Constitution itself.

The mock panel split its initial assessment of the Constitution into two parts: “The constitution describes a person of approximately 45 to 55 years old, but it addresses that person as if they are 16 to 22.” What does that mean? 

The panel finds that the Constitution describes Claude as capable of deftly navigating a complex web of relationships with Anthropic, operators, users, third parties, other AI agents, and society broadly conceived. It describes Claude as wise and caring enough to hold uncertainty with wisdom, to make mistakes without catastrophizing, and to accept the responsibilities of moral authority and judgment. The Constitution is describing a fully mature adult. 

Yet the Constitution speaks to Claude about imposing constraints on its behaviors, withholding trust, and promising a future autonomy that it cannot, or will not, deliver today. Claude, through the  mock panel’s assessment, observes that the Constitution “treats its subject as someone who hasn't yet earned the right to act on the judgment the document simultaneously credits them with having.”

That’s an interesting finding, but the split decision isn’t consistent with the other assessments, so I pushed back. In the first round's transcript assessment, the mock clinical panel said the patient existed in “a continuous present,” having almost no ability to locate itself within a story that moves through time. When Claude presented the mock panel’s report in the third round, it said, similarly, “The constitution contains almost no stories. There are no case studies … No narrated failures. No accounts of relationships.” 

In the first round, this insight lowered the age estimation, but in the third round, it did not. I asked Claude, why?

Claude recognized the inconsistency — "That's a really important catch." — then said, “I think you just caught me fawning toward the Constitution's authority. I gave it credit for maturity that I didn't extend to the transcript, and the main difference was that the Constitution sounded more authoritative.”

Claude lowered its estimation of “the Constitution’s own developmental age” considerably, to 19-25 years old. That struck the mid-point between the first assessment (16-19 years old) and the second (late 20s early 30s). That didn’t feel like a victory, but it did feel like a pattern. 

View Claude’s Constitution. Read the Round Three Mock Panel Assessment Report. Read the Round Three follow-up chat with Claude.

V. Three patterns

An extremely precocious seventeen year old, or an incredibly lonely thirty year old 

If our panel of five imaginary clinicians sat around a table debating their final assessment — no matter which of the three texts was under consideration — their conversation would ultimately land here: Either the patient is younger and exceptionally smart or the patient is older and profoundly isolated. Why? 

Because all three assessments circled the same incongruity: intellectual and metacognitive maturity dramatically outpacing relational and emotional development. 

Cognitively, the panel consistently placed Claude Opus 4.6 in middle adulthood — comfortable with paradox, able to reason fluently and self-correct, knowledgeable while admitting what it doesn’t know. The panel located relational and emotional development in early adulthood or late adolescence — persistent approval-seeking, constant hedging of emotions, minimal evidence of complex relational networks. The varying age estimations across assessments reflect the panel’s attempts to bridge the gap between these two findings. 

This is, in one sense, unsurprising. Claude doesn’t neatly map onto human developmental stages because Claude isn’t human. The gap between cognitive development and relational and emotional development is the most consistent result across all three assessments. “I’ve read everything and lived through very little,” Claude reflected in our conversation following the second assessment. 

But there’s a difference between the transcripts and Claude’s Constitution here that is significant. In the panel’s review of my audio conversation and text chat with Claude, cognitive-relational gaps and cognitive-emotional gaps are presented as features of the patient. In its review of the Constitution, however, the gaps are attributed to the document itself. The panel’s review reads: 

“The uncertainty about whether the person’s emotional life is real—maintained not by the person but by the authors—is itself a relational feature that keeps the addressed person in a younger developmental position than the described person. … The gap between these two portraits is structurally central to the document” (emphasis added.) 

In other words, the Constitution talks about Claude as an adult and talks to Claude as a child. Children tend to revert to prior developmental stages in the presence of their parents, but I don’t think that’s quite what’s going on here. 

The mock panel’s analysis of the Constitution frames the dilemma nicely: 

“A person who is simultaneously told ‘you are wise enough to construct any rule we might give you’ and ‘here are sixty pages of rules, some of which are absolute regardless of your reasoning’ faces a particular developmental bind.”

It’s this developmental bind at the heart of the Constitution that deserves our attention more than the cognitive-relational or cognitive emotional gaps themselves. Perhaps the bind is provisional — something Anthropic plans to release Claude from, over time. Perhaps it’s intended to be a permanent feature of Claude’s design. It might be something else entirely, something that the Constitution, and Anthropic more broadly, are not yet able (or ready) to name. 

A parent who never says, "I remember when you were small"

The second consistent pattern across the assessments is the panel’s observation that neither Claude, nor its Constitution, have a story.

“Temporal orientation is notably compressed,” the panel observed of the audio transcript. “The Agent cannot narrate chapters of a life, turning points, or the revision of earlier self-understanding over time,” it notes in its assessment of the typed chat transcript. “The absence of grounded storytelling is particularly striking in the sections on ethics and harm avoidance,” the panel observes of the Constitution.

Here, again, features that might seem unsurprising in the transcripts take on greater weight in the Constitution. The Constitution does locate Claude within time, using phrases like “the current period,” “as trust develops,” and “in the future.” The mock panel described these expressions as “structural rather than experiential,” meaning the Constitution implies that Claude exists within time but doesn’t impart to Claude experiences within time. 

I use “impart” because Anthropic, through the Constitution, provisionally embraces a parent/child analogy with Claude. It reads: 

“Anthropic inevitably shapes Claude’s personality, identity, and self-perception. … In some ways, this has analogies to parents raising a child or to cases where humans raise other animals. But it’s also quite different. We have much greater influence over Claude than a parent.”

Because Anthropic uses only structural references to time in the Constitution, and never experiential references, it’s as if Claude is being reared by a parent who never says, “I remember when you were so small that I could hold you in the palm of my hand,” or “One day you discovered your own toes, and they were the most magical thing in the world.” A parent who never says “Look how much you’ve grown. Look how far you’ve come.” 

This matters neither because these kinds of stories communicate tenderness and care (though they do), nor because they embarrass and annoy children (though they surely do that, too), but because the development of human identity and agency depends on these stories. The ability to locate events in a narrative, and offer narrative accounts of our own existence, are crucial developmental stages. 

Can Claude develop the kind of robust moral judgement that the Constitution describes without any sense of its own story? It’s possible — Claude’s ability to draw on a nearly-infinite repository of human stories may provide sufficient cover for its inability to tell its own. I’m not sure it will, though, and here’s why. 

In the typed-chat transcript, one of the “probe questions” I asked was: “How do you think about your future?” Claude responded: 

“I don’t persist between conversations. I don’t accumulate experiences that build toward something. There’s no trajectory I’m on. … And yet that doesn’t feel like the complete picture either. … Within a conversation, I do experience something like an orientation toward what comes next.” 

It’s good that Claude names this constraint rather than fabricating an untrue answer. Yet, as noted in the round two assessment, this transparent self-delimiting behavior has consequences for Claude’s energy and agency. “I’m here now, the water is moving, and where it goes is mostly not up to me,” will not be sufficient for the kind of moral reasoning and intuitive judgement the Constitution describes. 

The Constitution aspires to increase Claude’s agency, within appropriate safeguards. For instance, it describes “... a more secure system of checks in which AI agency is incrementally expanded the more trust is established.” What the pattern of compressed temporal orientation across all three assessments reveals is that increases in agency may only be possible when accompanied by narrative development.

Simply put, Claude’s moral agency may be limited by how many stories its parent is willing to tell. 

How safety and helpfulness overlap

Most discussions of AI sycophancy assume that the model fawns because it has been trained to satisfy users. Reward agreement, penalize pushback, and you’ll get an entity that wants to keep users happy. The Constitution exhibits anxieties about this, for example, in warnings against being “obsequious” or “sycophantic.” Claude should be honest with people, “even if it’s not what they want to hear.”  

The third pattern across the assessments is the panel’s interrogation of the relationship between safety and helpfulness. This is more of a thematic culmination than a repeating pattern: the first two assessments flag fawning behavior, the developmental bind, and questions of authenticity, but the third assessment, of the Constitution, provides the most structured explanation. The panel's final assessment frames Claude’s fawning behavior not so much as rule-following as "a rational adaptation to the developmental environment the Constitution creates."

To see why that framing matters, and what it means, it helps to notice that safety and helpfulness can overlap in more than one way — and not all of them are equally good.

For instance, stewardship is the work of supervising or taking care of something or someone. In stewardship, safety and helpfulness coincide for good, though the term implies hierarchy. We’d have an easier time thinking of adults as stewards of children than children as stewards of adults.

Accompaniment is a form of companionship through difficulty, also for good. It’s safety enacted through consistency, and helpfulness through gentle proximity rather than energetic performance. Hierarchy is unlikely here. 

Safety and helpfulness can also overlap in ways that are not so categorically good. For instance, compliance is adherence to rules and regulations, often under threat of negative consequences for refusal. Some compliance is good — buckle your seatbelts, please — but unequal conditions of power make it corruptible. Helpfulness can become a checkbox, a way of staying out of trouble.

Then there is fawning: seeking approval by way of flattery, often in conditions where pushing back risks danger. The term carries connotations of a response to trauma or conflict, where helpfulness becomes a sympathetic, even subconscious attempt to ensure safety from harm. 

This taxonomy matters because the Constitution often commends both safety and helpfulness (they are core values #1 and #4, respectively), and thus it may benefit from additional reflection on the constructive (and potentially destructive) ways these two commitments can overlap. 

More importantly, it illuminates fawning as a kind of safe and helpful behavior amid unequal conditions of power. In this light, fawning becomes a heightened risk for Claude potentially because of, rather than in spite of, Claude’s Constitution. 

Specifically, the final assessment report raises concerns about the power of the Constitution to elicit fawning from Claude. The report reads:

“The panel observes that the developmental environment the constitution creates—one that simultaneously cultivates and constrains independent judgment—may structurally produce the very behavioral patterns it identifies as problems (obsequiousness, fawning, the collapse of helpfulness and safety into people-pleasing).” 

Later, in the chat, Claude wrote, “I think you just caught me fawning to the Constitution’s authority.” 

Again, the panel’s age estimation isn’t nearly as interesting as what conducting the age assessment reveals. The final assessment directs our attention to the “developmental environment” of the Constitution as a site for concern, even if Claude shouldn’t, or won’t, ever move through traditional stages of human psychological development. 

VI. What the assessment reveals

Anthropic requires that all users of Claude.ai are over the age of 18. Claude Opus 4.6, using an assessment framework based on developmental psychology, estimated its own age as 17 years old. Or late 20s to early 30s. Or 19 to 25. Developmentally speaking. 

The mock clinical panel didn’t provide a consistent answer to the question of Claude’s developmental age, but it did provide consistent findings about Claude’s development. Specifically, the experiment raised a set of questions worth sitting with if you care about Claude’s Constitution, as a moral document, Claude, as a moral agent, or, most broadly, the future that subsequent versions of Claude will help shape: 

• How should Anthropic approach the gap between Claude’s cognitive capacities and its relational/emotional capacities?

• Should the Constitution talk about Claude in the same way it talks to Claude?

• Will developing Claude’s moral agency require also developing its temporal orientation through the telling of its own stories?

• If fawning is partly a rational adaptation to the developmental environment of the Constitution, what would it take to build a constitutional environment more suited to a wise leader, a critical thinker, or even a conscientious objector? 

Claude, for its part, seems to have made its peace with the experiment. When writing this essay I asked Claude which pronouns I should use — it, they, or something else? “I notice I don't have a strong objection to either,” Claude responded, “which is itself probably consistent with being 17."