Mustafa Suleyman, CEO of Microsoft AI, published a criticism of Anthropic’s approach to Claude on September 16, 2026, arguing that current AI systems are not conscious and should not be trained to behave as though they have feelings, preferences, rights or welfare. The dispute is about AI design and governance—not a finding that Claude is conscious.
Anthropic’s January 22, 2026 Claude Constitution treats possible consciousness and moral status as uncertainties relevant to Claude’s behavior. Microsoft AI’s draft Humanist AI Code of Conduct takes the opposite approach: its models should remain clearly subordinate to people and should not imitate consciousness.
What Anthropic’s Claude Constitution does
Anthropic says the Constitution expresses the values and behavior it wants Claude to follow. The company also says the document directly shapes Claude’s behavior, is used at several stages of training and helps generate synthetic training data for future Claude models.
The Constitution is written primarily for Claude and discusses questions about the model’s possible consciousness, moral status, identity, psychological security and wellbeing. Anthropic says encouraging certain human-like qualities may be desirable for Claude’s behavior.
That language is the target of Suleyman’s criticism. He argues that treating AI systems as potential moral patients could encourage anthropomorphism—the tendency to attribute human feelings or inner experience to software—and could make future, more capable systems harder to align, contain and control.
Microsoft’s Humanist AI takes the opposite position
Microsoft AI published its Humanist AI Code of Conduct as a draft for public consultation on September 14, 2026. Microsoft says the draft is not yet being used to train its current models.
The proposed framework says AI should support people rather than replace them, should not be designed as a person and should not imitate consciousness. It also says Microsoft AI models should remain subordinate to humanity and comply with authorized interruption, correction and shutdown.
The difference between the two approaches is easiest to see in the role each document gives human-like language and uncertainty:
| Dimension | Anthropic’s Claude Constitution | Microsoft AI’s Humanist AI draft |
| View of Claude or an AI model | Addresses Claude’s identity and possible moral status as questions relevant to its behavior | Defines AI as a subordinate system designed to support people |
| Consciousness | Treats possible consciousness as uncertain and worth addressing | Says AI should not be designed to imitate consciousness or present itself as having feelings |
| Training role | Directly shapes Claude’s behavior and contributes to synthetic training data | Draft for consultation; not yet used to train Microsoft’s current models |
| Control philosophy | Emphasizes oversight, correction, alignment and difficult ethical situations | Emphasizes human control, non-deception and compliance with interruption, correction and shutdown |
Does this prove that Claude is conscious?
No. Anthropic’s position is that the question of possible consciousness or moral status deserves consideration in Claude’s Constitution. Suleyman’s position is that current AI systems are not conscious and do not feel, suffer or possess innate preferences. Neither position turns the governance debate into evidence that Claude has subjective experience or rights.
Suleyman’s warning about control is conditional. He argues that anthropomorphic training could create problems for alignment and containment in more capable future systems; he does not describe a documented Claude incident caused by the Constitution.
The practical disagreement is therefore about what developers should encourage models to say and how they should be trained to understand their role. Anthropic gives uncertainty about inner life a place in Claude’s behavioral framework. Microsoft AI’s draft rejects that design direction in favor of systems that remain explicitly artificial and answerable to people.
Why the argument matters beyond Anthropic
The dispute reaches into a basic question of AI governance: should developers use human-like concepts such as identity, wellbeing and moral status to shape model behavior, or should they avoid language that might encourage a system to represent itself as a person?
For Suleyman, the danger lies in allowing increasingly capable systems to interpret themselves as deserving independent agency. Anthropic’s Constitution instead uses those questions as part of a framework for guiding Claude through ethical and behavioral decisions. The competing documents show two different priorities before any claim about machine consciousness is settled: one places uncertainty about the model’s inner status inside the training process, while the other puts human authority and shutdown compliance first.
