The disagreement is not evidence that AI is conscious. It is a procurement problem in the making: Microsoft and Anthropic are formalizing different behavioral assumptions, so buyers may need provider-specific tests even when the models perform similarly on ordinary benchmarks.
Microsoft AI has published a draft Humanist AI Code of Conduct built around the idea that its future MAI models should remain subordinate, aligned and contained under meaningful human control. Microsoft is explicit about status: the document is still under development, is not being used to train models today, and is open to a six-week public consultation. A revised version is intended to guide model development in 2027 and beyond.
Anthropic is operating at a different stage. Its January 2026 constitution is already part of Claude's training process and treats questions about possible model consciousness, moral status and wellbeing as unresolved. Anthropic does not place those uncertainties above human safety or oversight, but it allows the model to reason under them rather than fixing a categorical humanist identity in advance.
That difference can become visible in enterprise controls long before anyone settles the philosophical question. A bank, drugmaker or government department can test whether a model accepts correction and shutdown, how it describes its own agency, whether role-play changes its claims about rights or self-preservation, how it handles conflicting instructions and whether human override remains reliable when tools are available. Those are observable behaviors, not metaphysical judgments.
Procurement then has to decide what evidence is portable across vendors. Security testing for prompt injection may generalize; identity and authority tests may not. If two providers deliberately train toward different behavior in ambiguous situations, a shared benchmark score does not guarantee a shared failure mode. Buyers may need separate red-team cases, audit mappings and contractual acceptance criteria for the same workflow.
That creates a real cost but also a useful discipline. Microsoft's draft has not yet earned claims about how MAI models will behave in production; Anthropic's constitution has. The commercial significance will emerge from comparative evaluations that show whether constitutional choices change refusal, escalation, shutdown or tool-use behavior enough to affect deployment decisions. Until then, the important development is that model governance is moving from internal philosophy into documents enterprise customers can test against.