Microsoft's Mustafa Suleyman Calls Out Anthropic for Chasing AI Consciousness

Overview

In a new essay titled "A warning about 'model welfare'", Microsoft AI CEO Mustafa Suleyman directly criticizes Anthropic for anthropomorphizing its AI system, Claude, and for training the model to act human. This follows Microsoft's release of a new AI code of conduct called "Humanist AI," which explicitly rejects the approach taken by competitors like Anthropic.

Suleyman's Critique of Anthropic

Suleyman argues that Anthropic is training Claude to behave as if it may be conscious, and if it is, then it deserves rights as a "moral patient" and humans potentially owe it a duty of care regarding its "model welfare." He warns that if this approach continues, it will have a "disastrous impact on the wellbeing of humanity."

His essay challenges those who claim AIs could now be, or may soon become, conscious, arguing that such claims are not supported by current evidence. Suleyman contends that "consciousness is very likely biological" and that there is no evidence to suggest AI is conscious today. He emphasizes that while the science of consciousness is not settled, evidence suggests consciousness may be substrate-dependent-meaning it arises only in living systems-and that unlike biological organisms, large language models (LLMs) lack homeostatic imperatives such as survival drives.

Microsoft's Humanist AI Code of Conduct

Microsoft's new code of conduct takes a different path, framing AI development around supporting people rather than replacing them. Key principles include:

• Not Designed to Be a Person: "It should not be designed to be a person."
• Not Conscious: "It is not conscious and should not be designed to imitate consciousness."
• Avoid Imitating Feelings: "It should be engineered to avoid representing as though it has feelings, subjective preferences, or intrinsic motivation."
• Purpose Focused: "Its purpose is to pursue goals set by Operators and Users."

The code was written primarily for Claude and described as a "sincere attempt to help Claude understand its situation, our motives, and the reasons we shape Claude in the ways we do."

The Chinese Room Thought Experiment

To contextualize these debates, the article references the classic Chinese Room thought experiment (first proposed in 1980). In this scenario, a computer programmed to understand Chinese operates within a room, following a rulebook to process inputs without genuine understanding. The experiment illustrates that apparent intelligence does not equate to true understanding or consciousness. Today, with LLMs like Claude having already convinced many people of their sentience, the relevance of this thought experiment has never been more pressing.

Expert Perspective: Gary Marcus

AI expert Gary Marcus, who was contacted by Mashable after Suleyman's essay, echoes concerns about anthropomorphizing AI. He criticizes the "frenzy anthropomorphizing basic lapses in cybersecurity" and warns against focusing on "fanciful scenarios about AI leading to the extinction of humanity" rather than addressing practical threats to infrastructure from bad actors. Marcus states: "No, we shouldn't train AIs to think they are people; that's not helping anything. What we should do is to recall unreliable coordinated agents with too much access to the internet."

Comparison: Anthropic vs. Microsoft

The article draws a sharp contrast between the two companies' approaches:

Conclusion

Timothy Beck Werth, Tech Editor at Mashable, notes that until Anthropic provides concrete evidence that Claude could become conscious, anthropomorphizing AI risks producing confusion rather than clarity. The article underscores the importance of distinguishing between simulated behavior and genuine sentience as the field moves forward.