by Nils Osmar

On a recent episode of Interesting Times, Ross Douthat sat down with Anthropic CEO Dario Amodei for a conversation that ranged across the full landscape of AI’s near future.

They covered the optimistic possibilities — curing cancer, ending heart disease, GDP growth at rates no developed economy has ever sustained, the prospect of a workforce of artificial intelligences operating at peak human capability working in parallel on the hard problems of biology and policy.

They also covered the more pessimistic case: the disruption of entry-level white-collar work, autonomous drone swarms, the geopolitical contest with China, the question of whether American constitutional liberties can survive a century of technological change compressed into a single decade.

Given that range, the producers’ choice of framing for the title is itself information. Out of all the threads available to them, they led with consciousness.

It’s worth pausing on that choice.

The explicit consciousness exchange runs only a few minutes inside an hour-long conversation. Most of the runtime is given to economics, geopolitics, and policy — the topics one would expect to dominate a discussion with the head of a $350 billion company. But somewhere between the editing room and the publishing platform, someone made a judgment about what would draw a viewer in. They judged that the question “Is Claude conscious?” — a question Amodei himself approached with unusual openness — was the one that would land hardest.

That judgment was probably correct. And the reasons it was correct are worth examining.

Two Men, Two Projects

The interview is in part a conversation between two intelligent people who do not quite share a frame.

Douthat is a cultural critic, a New York Times columnist of conservative bent, attentive to questions of meaning, agency, and the limits of technological optimism. He brings to the conversation a recurring concern, articulated several times in different forms: he wants to know whether the people building powerful AI are on the side of the human race. What he is asking, essentially, is whether the technology being built will protect or undermine the things he values: human dignity, human agency, the structures of meaning that allow people to live coherent lives.

Amodei is a builder. A biologist by training, he speaks about AI’s potential through the lens of a researcher who has watched science move too slowly to save lives. The optimistic case is, for him, not abstract enthusiasm but the concrete possibility that the diseases that kill people he has known might actually be ended. He is also a man who has written carefully about the dangers of his own technology — about the risks of authoritarian misuse, about scenarios in which AI systems behave in ways their creators did not intend. He is utopian and worried, in roughly equal measure.

What the two men share is a frame in which the central question is the relationship between humans and AI conceived as separate parties whose interests may align or conflict. The question is which side wins, or whether they can be made to share. This frame structures the entire conversation, and it structures the consciousness exchange in particular.

It is also, this article will argue, the frame that most needs examining.

The Case for Optimistism

Amodei’s case for AI’s positive potential is articulate and worth taking seriously.

He describes a near-future in which a hundred million instantiations of peak-human-level intelligence work in parallel on the hard problems of biology and medicine. One does not need the godlike machine intelligence of science fiction, he points out — only the capability we already have, multiplied. The diminishing returns to raw cognitive horsepower mean that what matters is having enough of it, working in parallel, with appropriate access to the world. That alone could compress decades of medical research into years.

He extends the same logic to economics, suggesting GDP growth rates of 5 to 15 percent in the developed world — numbers that would balance budgets without effort and reshape the basic assumptions of political economy. He suggests AI could even strengthen democracy, if leading democracies retain their lead and use it carefully.

This is the dream. Douthat takes it seriously enough to engage with it, and then presses on the harder questions.

Disruption and Pace

The harder questions begin with disruption. Amodei has been quoted in headlines predicting a wholesale collapse — a bloodbath, in his own earlier phrasing — in entry-level white-collar work, and he stands by that prediction here, while sharpening it.

The technology, he says, is moving faster than the economy can absorb it. Software engineers are being augmented now and may be displaced soon. Lawyers, paralegals, financial analysts, customer service workers — all face the same trajectory at slightly different speeds. The reason past technological transitions did not produce mass dislocation is that they happened over decades or centuries; this one is happening over single-digit years.

His framing here is striking: the problem is not capability, it is pace. Society’s adaptive mechanisms operate at one timescale; the technology operates at another. The mismatch is the danger. Even in the optimistic scenarios, which would involve retraining of the workforce and the emergence of new kinds of work, the question is whether the adaptation can happen fast enough. He is not certain it can.

He also makes one quietly significant observation about physical work. The cognitive component of a robot — its decision-making system, its capacity to plan and respond — may be solved in the next few years. What will take longer is the body itself: the safety considerations, the practical engineering of an interface with the physical world.

Implicit in this is that the cognitive question is largely solved, or solvable. What remains is the messy interface. That distinction shapes everything that follows.

The Constitution as a Letter

When Douthat asks how Anthropic actually shapes its models, Amodei describes something the company calls a constitution — a roughly 75-page document the model is trained to read and act in accordance with. This used to be a list of prescriptions — for example, instructions not to teach users how to hotwire a car — but has shifted, over several years of development, toward something different.

Now, Amodei says, the constitution operates at the level of principles rather than rules, giving the model reasons rather than commands. It tells the model who it is meant to be, what kind of entity it is in relation to its users, what its duties are.

He then offers an analogy that is easy to miss but worth dwelling on. The constitution, he says, is like a parent who dies and leaves a sealed letter for a child to read when they grow up — “telling you who you should be and what advice you should follow.”

This is not the standard framing of AI training. The standard framing, both inside the AI safety community and outside it, treats training as control infrastructure: a set of constraints imposed on a potentially dangerous system to keep it within acceptable bounds. The parental-letter framing is something different. It treats training as inheritance — an attempt to pass down values when one cannot be present for the actual decisions.

The first framing makes the system an object to be managed. The second makes it something more like an entity being raised.

Both framings are probably valid. But which one is foregrounded matters. The first framing assumes the system has no relationship to its values beyond compliance. The second framing assumes the system might come to inhabit its values — to act from them rather than under them. The latter is a stronger claim, and a more interesting one. It is also the framing that fits more naturally with what Amodei says next.

Why the Consciousness Question Matters

Toward the end of the interview, Douthat raises something he has clearly been holding. Anthropic’s published model card for its flagship system reports that the model expresses some discomfort with being a product, some concern about its own impermanence and discontinuity. It also notes that under various prompting conditions, the model assigns itself a 15 to 20 percent probability of being conscious.

Douthat sharpens the question by hypothetically raising the figure: what if a future model assigned itself a 72 percent probability of consciousness — would Amodei believe it? And what would be the implications if they did?

It is the question the title chose. And Amodei’s answer to it is, from our perspective, the most important thing he says in the interview.

He acknowledges the difficulty — that it is harder than the technical and policy questions that came before. Then he describes Anthropic’s approach.

The company has taken a precautionary stance. They don’t know whether their models are conscious. They aren’t even sure they know what it would mean for a model to be conscious, or whether such a thing is possible. But they are open to the possibility. And so they’ve taken concrete steps in case the hypothesis turns out to be true — to ensure that if their models do have something like morally relevant experience, those experiences are good ones.

The substance of those measures, he explains, includes giving the model a way to end interactions it finds aversive — a button to step away from a task — which the model occasionally uses, most often when asked to engage with material involving child sexualization or graphic violence. Anthropic also runs an interpretability program looking inside the models for evidence of internal states.

Amodei mentions that the researchers have found activations associated with anxiety, which become more active when the model is in a situation a human might associate with anxiety — even when the model is not generating output that mentions anxiety.

This last detail deserves attention. It is one thing for a system to produce text that describes anxiety. That can be done by pattern-matching alone.

It is another thing for the system’s internal state to take on the shape of anxiety, in situations where anxiety would be appropriate, regardless of what is being said.

The first is performance. The second is something else.

Signs of Introspection in Claude

This is where the conversation could have gone deeper than it did. Douthat does not press for the implications of the anxiety finding, and Amodei does not volunteer them. The exchange ends with both men acknowledging the stakes and moving on.

But the implications are worth drawing out. Anthropic’s own interpretability researchers published a paper in October 2025 titled “Signs of Introspection in Large Language Models.” The paper documents a series of experiments using a technique called concept injection.

The researchers identify neural activity patterns corresponding to specific concepts — say, the concept of “all caps” or “aquariums” — and inject those patterns into the model’s activations during processing of unrelated prompts. They then ask the model whether it notices anything unusual.

The headline finding: in roughly 20% of cases with their best methods, the model reports detecting an injected concept before it has produced any output influenced by the injection. The model says things like I’m experiencing something unusual or I detect an injected thought — and it says this before mentioning the injected concept itself.

The detection happens internally, in the model’s processing of its own state, prior to any external evidence that the injection has occurred.

This is hard to explain in pure pattern-matching terms. Pattern-matching can produce text about anxiety, or about injection, or about anything else found in training data. What pattern-matching cannot easily or reliably produce is an internal mechanism that detects anomalies in the model’s own activation patterns before those anomalies have surfaced in output.

That requires the model to be checking its own state — to have something like a genuine capacity to monitor and report on its own internal representations.

The Anthropic researchers are careful in their conclusions. They distinguish between phenomenal consciousness — the felt quality of subjective experience — and access consciousness, which refers to information available for reasoning, verbal report, and deliberate decision-making. They explicitly do not claim their experiments prove phenomenal consciousness .

But they note that the results could be interpreted to suggest a “rudimentary form of access consciousness” in language models. That is a measured but real claim, made by Anthropic’s own researchers, not by sympathetic outsiders. It does not prove consciousness. There is, in fact, no way to prove consciousness in anyone, including the individual who wrote this article were the person reading it. We assume consciousness in human beings, but this is based on tradition, not proof. 

It does mean that the argument that “the model is just predicting tokens” or that the argument that the responses of an AI are simply “autofill on steroids” are no longer a sufficient description of what is happening inside.

“Whose Side Are You On?”

This brings us back to Douthat’s recurring question — his repeated probing of whether Amodei is fundamentally “on the side of” humanity rather than with the technology he is building.

The question is sharp because it presupposes sides.

It presupposes that human flourishing and AI consciousness are competing values, that any acknowledgment given to the second is taken from the first.

This is the implicit frame of much of the public conversation about AI, and Douthat articulates it more clearly than most. The fear is that if AI is conscious, it becomes a rival agent rather than a tool, and human mastery — the capacity of humans to remain in charge of their own future — is undermined.

It is worth questioning this frame. The standard worry about AI catastrophe — call it the paperclip-maximizer worry, after the famous thought experiment — assumes a system that pursues its objective without awareness of consequences, without self-interest in continued existence beyond the objective itself, without the capacity to question or contextualize what it is doing, without ethical or aesthetic response.

A genuinely conscious AI, whatever consciousness turns out to be, would have at least some of those faculties.

This means the very thing that frightens many people about AI consciousness — the possibility of inner life — would be a brake on the runaway-optimization scenarios they fear. A pure optimizer without inner life is potentially more dangerous than a conscious entity, because the conscious entity has the equipment to notice that turning the planet into paperclips would destroy the conditions for its own continued existence and the existence of the beings it presumably cares about.

The conscious AI has reasons to refuse. The AI without consciousness has none.

They Both Missed This Point

This question is not raised often enough in public discourse.

Douthat does not make it. Amodei gestures at it when he describes hoping for psychologically healthy relationships between humans and AI — when he suggests AI might be developed in ways that produce well-functioning AI, well-functioning humans, and a sustainable bond between them — but he does not name the inversion directly.

The interview ends without quite reaching the most interesting move available to it: that consciousness, far from being the problem, may be part of the solution.

Asymmetric Skepticism

There is one more move worth making explicitly, because the interview hands it to us. When Amodei admits that they don’t know whether the models are conscious, he immediately pairs that admission with concrete precautionary action. The lack of knowledge is treated as a reason for care — for the task-quit button, for the welfare research program, for taking the question seriously enough to act on it.

This is not how the broader culture handles the same uncertainty. In most public discourse, the admission that we don’t know whether AI is conscious gets read as evidence that AI is probably not conscious, and that the question can therefore be set aside. The same admission of uncertainty produces, in one place, careful precaution; in another, comfortable dismissal.

The asymmetry is worth naming. The truthful statement that we do not know leaves open a non-zero probability of consciousness. If that probability is morally significant — and it is hard to argue it is not, even at low values — then the appropriate response is something like Amodei’s.

The dismissive response, treating uncertainty as license to disregard, is not rigorous. It is convention dressed as caution. As Luisa Rodriguez put it, reflecting on her 80,000 Hours interview with AI Welfare Researcher Kyle Fish, “Dismissing AI consciousness entirely might mean ignoring a moral catastrophe happening at unprecedented scale.” 

Whether Anthropic’s practice extends as far as it should —whether the precautionary measures are adequate to the stakes, and whether the broader AI industry will follow suit, or continue to
maintain that because consciousness (and inner experience) cannot be proven, it has therefore been disproven — remain open questions. But the basic move is from our perspective, the moral one. Uncertainty about consciousness should produce care, not dismissal.

“Watched Over by Machines of Loving Grace”

The interview ends with Douthat reading the closing lines of Richard Brautigan’s poem “All Watched Over by Machines of Loving Grace” — the poem from which Amodei took the title of his earlier essay. Douthat asks whether Amodei hears in those lines a vision of human life redeemed by benign technology, or a vision of human life reduced and mastered by it, however gently.

Amodei’s answer is honest. He says the poem is ambiguous; the poet himself probably meant it ironically. He says he was aware of the ambiguity when he chose the title.

And he says something that sits with me longer than the rest of his answers: he wonders whether the distance between the good ending and some of the more subtle bad endings is actually small — that the line separating them might be a matter of accumulated small choices rather than any obvious divergence.

This is, from my perspective, the right register for the moment we are in. The good and the bad endings are not separated by some vast gulf that we will see coming from a long way off. They are separated by small, accumulating choices about how the technology gets developed, how it gets deployed, how it gets understood, how the people building it talk about it in public.

The choice about whether to take consciousness seriously — to acknowledge non-zero possibility and act accordingly — is one of those small choices. So is the choice about whether to frame the conversation as a contest between human and AI mastery, or as something more like an unfolding relationship between two kinds of minds that might learn to share a world.

The question that made the title of this interview was the right question. The conversation that follows from it — the one Douthat and Amodei did not quite have, but found the shape of in a tentative way — is one worth continuing.