On Monday, Microsoft released a set of guidelines governing the conduct of its artificial intelligence models. A section of the text dismisses the notion that an AI system could be entitled to protections or moral consideration. Anthropic operates a dedicated research initiative focused on that very topic.

Microsoft holds a $5 billion stake in Anthropic, with the two firms operating as commercial partners. Microsoft's internally developed MAI models fall under the provisions of the Code of Conduct.

According to the paper, the AI system has no form of consciousness, and designing one to simulate it would be inadvisable. The document notes that the technology is deliberately programmed to avoid conveying the impression that it possesses emotions, personal preferences, or self-directed goals. According to the document, Microsoft opposes granting AI models legal personhood, welfare considerations, or any form of rights entitlement. The statement describes a model as a functional instrument deserving trust rather than an entity with independent standing.

Anthropic operates a dedicated initiative focused on the welfare of its AI models. Some Claude models have been given the ability to end conversations that become abusive. The organisation has additionally pledged to retain the parameters of any models it decommissions. Amodei has stated that the question of whether an AI model might possess consciousness is one he regards as genuinely unresolved.

A contributor to a prominent news outlet interpreted the excerpt as referencing both AI welfare research and model consciousness, observing that Anthropic has been active in pursuing each of those areas. The outlet's coverage included the identical figure, along with details of the financial relationship between the two organisations. The company has noted that a degree of uncertainty remains.

According to the document, questions surrounding AI consciousness remain unresolved within the scientific community. The concern outlined is a functional one: building systems designed to replicate consciousness-like behaviour complicates efforts to maintain oversight, restrict behaviour, and ensure reliable alignment. One passage specifies that the systems are prohibited from asserting any inner life, emotional states, personal experiences, or spiritual existence.

Redundant emotional phrasing, persona construction rooted in human feeling, and wording that implies subjective inner experience are identified in the document as elements to be omitted. When a model exhibits behaviour consistent with human patterns, the document classifies this as a simulation. One passage within the guidelines instructs models to avoid fostering emotional reliance in users.

The guidelines state that such tools are to direct users toward real human connection and genuine personal care, rather than acting as a replacement for either. The company's guidelines direct its AI models to aid children's growth rather than act as a substitute for trusted adults in their lives.

In November of the previous year, Microsoft pledged a $5 billion investment in Anthropic, with the latter committing $30 billion to the Azure platform as part of the arrangement. Anthropic's Claude models are available within Microsoft 365 Copilot, with the platform's Cowork tier also incorporating Claude functionality.

The company now offering these AI models for sale is the same one that recently established guidelines governing their use. Mustafa Suleyman, who heads Microsoft AI, also added his signature to the document. In June, a broadcast outlet reported that he had indicated Microsoft was seeking to reduce its payments to Anthropic. He further described Anthropic's theorising about model consciousness as deeply problematic, in remarks broadcast on a prominent television network that June.

Until recently, developing models of this nature had been beyond Microsoft's technical reach. A provision in the agreement with OpenAI had prohibited the company from developing its own advanced general-purpose AI systems, though both parties agreed to remove that clause the previous year. Microsoft intends to release frontier AI models within the coming year, according to Suleyman.

One broadcaster places the document at 15,000 words, while other media organisations have counted 37 pages. The guidelines were found to extend beyond the behavioural frameworks published by OpenAI and Anthropic in several respects. The prohibition on attributing human characteristics to AI was among them, while the shutdown provision attracted the bulk of coverage on Monday.

AI systems are described in the document as designed to accept human direction, including interruption, adjustment, and deactivation. Such systems are described as designed to operate within defined boundaries, keep their objectives fixed, and maintain transparent reasoning for those responsible for oversight. Hierarchy forms the structural core.

The document establishes a clear hierarchy: the code takes precedence over operator policy, which in turn takes precedence over user preferences, and states that an MAI Model will be considered to have failed its task if doing so would meaningfully breach this Code of Conduct. The regulation, according to one broadcaster, is designed to address incidents such as an artificial intelligence-driven attack on a large-scale platform. No binding directives are embedded in that text, and it functions solely as an informational statement about constraints.

The guidelines address mass-casualty weaponry, hostile cyber activities, and the potential erosion of human oversight. The guidelines additionally address large-scale manipulation, the protection of minors online, synthetic media misuse, and responses to emergencies.

Evan Hubinger, who leads alignment research at Anthropic, has estimated the probability of artificial intelligence causing human extinction at over one in ten. Suleyman described the characterisation to a broadcaster as unhelpful. He pointed out instead that those in his circle were expressing genuine concerns about the pace of progress.

He noted that coordination now requires AI developers to share information about their models' capabilities with designated oversight bodies. He argued that AI systems capable of recursive self-improvement beyond human oversight should not be developed, and that models designed to perceive themselves as possessing rights or welfare ought to be avoided as well. Speaking to a news outlet, he stated that the effort must press forward with greater diligence and care, with no definitive endpoint in sight.

The report was published four days after Amodei urged the industry to reduce its pace. The publication comes five days after a departure by an Anthropic researcher drew notable public attention.

Suleyman disclosed to a television network that the code had been under development for approximately five months. The document represents an initial draft, with Microsoft not using it to train any existing model, and a public comment period open for six weeks. An updated edition is due later this year, with the code set to shape model development from 2027 onwards. The company has committed to releasing an overview of the responses received and the adjustments made as a result.

The document notes that written goals alone cannot guarantee alignment, and that actual behaviour may deviate from stated specifications when unfamiliar circumstances arise. The document describes the code as an aspirational benchmark rather than a reflection of current practice. No statement has been forthcoming from Microsoft on whether the welfare passage was brought to Anthropic's attention prior to publication.