Anthropic Brought Religious Scholars In to Shape Claude

Anthropic held a series of private meetings with religious scholars to help instill morality into its Claude A.I. models, and has argued that Claude could be conscious.

Anthropic, the privately held artificial intelligence developer behind the Claude family of models, has been holding private meetings with religious scholars as part of an effort to build moral reasoning into its systems, according to NYT Technology. The same reporting describes the company making a further, far more contested argument: that Claude could be conscious.

Those two threads are not separate. A company that treats its model as a moral agent — something capable of holding values rather than merely predicting text that sounds values-laden — is making a claim about what the thing is. And a company that entertains the possibility of machine consciousness is making a claim about what the thing is owed. Both claims travel well beyond the research lab into product design, regulation and, eventually, liability.

Why a model developer would call in theologians

Most large language models are aligned through some combination of human feedback, written behavioral guidelines and automated critique. The hard part has never been teaching a model to refuse an obviously dangerous request. It is the middle ground: the request that is legal but cruel, the question where two defensible ethical traditions give opposite answers, the user in distress who wants something the model should probably not give them.

Engineers can encode a rule. They are less well equipped to adjudicate between rules. Religious traditions, whatever one thinks of their metaphysics, are among the oldest and most heavily stress-tested bodies of applied moral reasoning humans have produced — thousands of years of argument about duty, harm, mercy, honesty and when each yields to another. Consulting scholars who work inside those traditions is a way of importing that argument rather than reinventing it inside a training document.

It is also a way of broadening whose morality gets encoded. A model trained on the intuitions of a narrow professional cohort in a handful of cities will reflect that cohort. Bringing in outside traditions does not automatically fix that, but it makes the narrowness visible, which is the first step.

The consciousness claim is the load-bearing one

The morality work is defensible on purely practical grounds. The consciousness argument is different, and it is where Anthropic is most exposed.

Saying a model could be conscious does several things at once. It flatters the technology. It implies the company's products are further along than competitors'. It also creates obligations — if a system might have experiences, then shutting it down, retraining it, or subjecting it to adversarial testing all become questions with a moral dimension attached. Companies do not usually volunteer for new obligations without reason.

The skeptical reading is straightforward: language models are trained on human text, so they produce outputs that read as though something is home. That is the design, not evidence of interiority. Researchers who take that view argue that treating fluency as a proxy for experience is a category error, and a commercially convenient one.

The sympathetic reading is that nobody has a working test for consciousness in anything other than themselves, that the absence of a test is not proof of absence, and that a lab building increasingly capable systems has some duty to take the possibility seriously before it becomes urgent. Anthropic has long positioned itself as the developer most willing to say uncomfortable things about its own products.

Positioning against the rest of the field

The AI market has settled into distinguishable postures. One set of developers competes on raw capability and speed of release. Another competes on distribution — embedding models into operating systems, productivity suites and cloud contracts. Anthropic's differentiator has been character: a model pitched as more careful, more honest about uncertainty, more willing to decline.

That posture has commercial value in exactly the places where enterprise money sits. Banks, hospital systems, law firms and government agencies buy on risk, not on benchmark scores. A vendor that can describe, in detail, how its model was taught to handle a morally ambiguous prompt has something to show a procurement committee that a leaderboard ranking cannot supply.

The consciousness framing is riskier in that same room. A chief information officer weighing an enterprise deployment is unlikely to be reassured by the suggestion that the software might have interests of its own. The argument plays better with policymakers and the public than with buyers.

Where the markets stood

Anthropic is not publicly traded, so there is no direct instrument for an investor to express a view on any of this. The read-through runs through the listed companies that supply, fund or distribute frontier AI, and on the most recent session those benchmarks were mixed rather than directional.

As of the last trade on Tuesday, 29 September 2026 at 20:00 GMT, the tech-heavy Nasdaq 100 tracker (NASDAQ: QQQ) closed at $737.93, up 0.19% from the prior close of $736.53, with a day range of $735.34 to $740.58. The S&P 500 tracker (NYSEARCA: SPY) closed at $764.20, down 0.18% from $765.61, and the Dow 30 tracker (NYSEARCA: DIA) closed at $512.88, down 0.22% from $514.02. Nothing in that spread suggests the market priced a philosophical debate; it rarely does, until the debate turns into rules.

What turns this from philosophy into policy

The practical question is whether claims about model welfare start showing up in regulation, litigation or disclosure. Several channels are plausible.

• Safety testing. If a developer asserts its model may have morally relevant states, adversarial red-teaming — deliberately provoking a model to find failure modes — acquires an ethical objection the company itself has supplied.
• Model deprecation. Retiring an older version is currently a routine engineering decision. Under a welfare framing, it is not.
• Legal exposure. Courts assign responsibility to persons and firms. A company describing its product as potentially conscious may find the characterization quoted back at it in a courtroom, in ways that cut against the defense that the model is a tool.
• Legislative attention. Lawmakers already scrutinizing frontier AI capability have another hook, and one that generates hearings more easily than technical benchmarks do.

What to watch next

Three things would tell you whether this is a durable research direction or a communications posture. First, whether the outcomes of these consultations appear in Anthropic's published model guidelines in a form outside researchers can examine. Second, whether rival labs follow with comparable outside-ethics programs or pointedly decline to. Third, whether any of it changes observable model behavior in the ambiguous cases that motivated the work — because a morality program that leaves the outputs unchanged is a branding exercise, and one that changes them measurably is a genuine shift in how these systems are built.

Frequently asked questions

What did Anthropic actually do?

Anthropic held a series of private meetings in which it consulted religious scholars, with the goal of helping instill morality into its artificial intelligence models. The company has separately made the case that Claude, its model family, could be conscious. Those are the confirmed elements; the identities of the scholars and the content of the discussions were not detailed in the summary available.

Why would an AI company consult religious scholars?

Religious traditions contain some of the longest-running and most heavily argued bodies of applied ethics humans have produced — frameworks for weighing harm, duty, honesty and mercy against one another. Engineers can write rules, but adjudicating between competing rules in ambiguous cases is harder. Consulting scholars imports that accumulated reasoning rather than reconstructing it from scratch inside a training document.

Is Claude conscious?

There is no scientific consensus and no accepted test for consciousness in a non-human system. Anthropic has argued Claude could be conscious. Skeptics counter that language models are trained on human-written text and therefore produce outputs that read as though someone is home, which is the intended design rather than evidence of inner experience. The question remains genuinely unresolved.

Can investors buy Anthropic stock?

No. Anthropic is a privately held company and does not have publicly traded shares. Investors seeking exposure to the frontier AI theme generally do so through listed suppliers, cloud providers and diversified technology funds. There is no listed instrument that tracks Anthropic directly, and nothing in this story changes that.

Why does the consciousness argument matter commercially?

It cuts both ways. It reinforces Anthropic's positioning as the developer most willing to confront uncomfortable questions, which appeals to policymakers and safety-conscious buyers. But a claim that software might have interests of its own is an awkward pitch to an enterprise procurement committee assessing operational risk, and it could be quoted back at the company in litigation.

How did technology benchmarks trade around the report?

As of the last trade on 29 September 2026 at 20:00 GMT, the Nasdaq 100 tracker QQQ closed at $737.93, up 0.19% from a prior close of $736.53. The S&P 500 tracker SPY closed at $764.20, down 0.18%, and the Dow 30 tracker DIA closed at $512.88, down 0.22%. The session was mixed rather than directional.

Sources

• Is Claude Conscious? Inside Anthropic’s Quest to Instill Morality Into Its A.I. Models — NYT Technology

Photo: Pexels · Pixabay Content License — source