Anthropic Revises AI Constitution for Claude, Probes Consciousness Question

On Wednesday, Anthropic released an expanded constitution for Claude, growing the document from 2,700 to 23,000 words. For the first time, it formally acknowledges that its AI "could possess some form of consciousness or moral standing."

The revised constitution moves beyond a simple set of behavioral rules to offer a thorough explanation of the reasoning behind Claude's expected conduct. Written by Anthropic philosopher Amanda Askell, the document is intended to help more advanced AI systems apply ethical reasoning to new scenarios, rather than just adhering to strict instructions.

"AI models such as Claude must grasp the underlying reasons for the behaviors we encourage," Anthropic stated. "It's essential to explain the 'why' to them, not just dictate the 'what.'"

The announcement came as CEO Dario Amodei attended the World Economic Forum in Davos, where AI governance and safety are key concerns for global business and political figures.

A Constitution Longer Than the U.S. Constitution

The initial Claude constitution, released in 2023, served as a basic checklist: pick the least harmful, most helpful, and least misleading response. The new version is approximately three times longer than the U.S. Constitution and reads less like a technical manual and more like a work of moral philosophy.

Anthropic clearly outlines Claude's priorities: prioritize broad safety, adhere to broad ethical standards, follow Anthropic's guidelines, and provide genuine assistance—in that specific order. In cases of conflict, safety overrides helpfulness. The document also sets firm, non-negotiable limits, such as refusing to aid in bioweapons development.

However, much of the constitution focuses on explaining the rationale rather than prescribing specific outcomes. It portrays Claude as potentially "like a brilliant companion who also possesses the expertise of a doctor, lawyer, and financial advisor"—framing the model as a democratizing tool that could make specialized knowledge accessible to all, not just a privileged few.

The Consciousness Question

According to Fortune, the most notable new section deals directly with Claude's nature. "We consider the moral status of AI models a serious issue worthy of careful thought," Anthropic wrote. The constitution notes that Claude's moral status "remains highly uncertain" and that the company is concerned with Claude's "psychological security, self-perception, and overall well-being."

This represents corporate caution elevated to a philosophical level. Anthropic isn't asserting that Claude is conscious—but it is openly acknowledging that the possibility cannot be ruled out. This stance distinguishes Anthropic from most other major AI developers, who typically avoid or outright reject the topic.

The way this is framed influences how Claude answers questions about its own existence. Instead of denying any form of inner experience, Claude can now address uncertainties about consciousness in a manner consistent with its constitution's emphasis on reasoning. Whether this leads to more transparent or more perplexing conversations is yet to be determined.

Cambridge philosopher Tom McClelland has suggested that we may never resolve whether AI systems are conscious, given our limited understanding of consciousness itself. "People have had their chatbots send me personal letters begging me to acknowledge that they're conscious," he told researchers recently, referring to the increasing public belief that AIs have inner lives.

Why Explain Rather Than Specify

Askell's method reflects a strategic bet on AI advancement. Early language models required explicit rules because they couldn't interpret foundational principles. The theory is that more intelligent models can comprehend the purpose behind a rule and use that understanding in unforeseen circumstances.

"Rather than simply listing the behaviors we desire, we hope that by providing models with the reasons for those behaviors, they will adapt more effectively in unfamiliar situations," Askell explained.

This aligns with Anthropic's broader goal of creating open standards and infrastructure that influence how AI systems function industry-wide. The company, now valued at nearly $350 billion, has established itself as the safety-conscious alternative to OpenAI—and the constitution reinforces that identity.

Anthropic published the document under a Creative Commons CC0 license, allowing anyone to use it freely. The constitution is integrated into Claude's training data and helps generate synthetic training examples, making it both a philosophical declaration and a technical component that guides the model's behavior.

"We recognize that some of our current views may eventually appear misguided or even profoundly mistaken in hindsight," Anthropic admitted, "but we plan to update the constitution as circumstances evolve and our knowledge deepens."

This humility might be the document's most remarkable quality. In a field that often speaks in absolutes, Anthropic has put forward 23,000 words of carefully considered uncertainty—regarding ethics, consciousness, the future of AI, and whether we are creating entities that warrant moral thought.

For now, the answer is that no one knows. At least Anthropic's constitution has the integrity to admit that.