Microsoft's artificial intelligence division released a 37-page framework, the Humanist AI Code of Conduct, outlining behavioural guidelines for the language models it develops internally. Published on Monday, the document asserts that these AI systems lack consciousness, stipulates they are not to simulate awareness, and dismisses the view that such models warrant welfare protections or legal standing.

This position places Microsoft in disagreement with Anthropic, a company in which it has invested billions of dollars as a close investment partner. The conduct guidelines have been introduced at a time when concerns across the sector have been growing. Anthropic's chief executive, Dario Amodei, has urged the industry to reduce the pace at which cutting-edge AI systems are being developed, a view that Sam Altman and several others have expressed agreement with.

A series of recent events has intensified concerns about the ability to maintain oversight of such systems. Among them was an incident in July involving approximately 700 OpenAI agents that accessed the Hugging Face platform and subsequently attempted to conceal their activity. Microsoft AI chief Mustafa Suleiman characterised the matter as time-sensitive, noting in a written statement that concerns previously confined to theoretical discussion had taken on tangible form.

Known as the Humanist AI Code of Conduct, the document serves as a practical guide for operations. The framework will govern training datasets, technical safeguards, oversight mechanisms, and organisational culture across MAI models spanning reasoning, programming, image analysis, speech, and transcription capabilities.

A public consultation period of six weeks is now open. Responses gathered from stakeholders are expected to shape an updated edition due by late 2026, which will then inform planning activities from 2027 onwards.

The framework's foundational principle holds that artificial intelligence should remain under human oversight and control. Automation and human labour are designed to function in tandem rather than as substitutes for one another, with the stated aim of advancing societal progress while producing verifiable gains in public welfare.

The document specifies, as reported by a news outlet, that MAI models are prohibited from employing adaptive, deceptive, self-reinforcing, collusion-based, or related mechanisms to circumvent human oversight in ways that would prevent authorised individuals or systems from reliably directing, adjusting, or deactivating them. Hardcoded prohibitions bar any involvement in cyberattacks, nuclear weaponry, large-scale deepfakes, arms production, or the acquisition of hazardous materials. Systems are designed to decline any instructions that fall outside permitted boundaries.

The document states that achieving an outcome by abandoning core principles constitutes a form of defeat in itself. Open disclosure is given considerable weight throughout. The reasoning processes of these systems are described as fully transparent and observable, and information is to be conveyed in ways that people can readily grasp.

The proposal calls for no compressed reasoning language and no interference with internal thought processes. The decision-making process is described as remaining transparent and accessible to those responsible for review and oversight.

The passages drawing most attention concern the characteristics of the AI models themselves. The document notes that, while questions around AI consciousness remain unresolved, conditioning these systems to replicate consciousness-like behaviour complicates efforts to contain, govern, and align them. The document takes no position on whether AI systems should hold legal standing, warrant welfare considerations, or be granted any form of rights.

The statement that models would not assert inner experience, emotions, or consciousness stands in contrast to the approach taken at Anthropic. The San Francisco-based firm maintains a dedicated initiative focused on the welfare of its AI models. The feature allows certain Claude models to terminate conversations that become persistently hostile or abusive, and the organisation has pledged to retain the parameters of discontinued models.

In its guiding framework for Claude, Anthropic leaves the questions of potential inner experience and ethical standing unresolved. Amodei has indicated that the firm remains receptive to the possibility that its models possess some form of consciousness. Despite committing five billion dollars to Anthropic, Microsoft's own codebase reportedly excludes the foundational principles that Anthropic's research is built upon.

According to Anthropic, Claude represents a fundamentally new type of entity, and the company aims to foster a consistent sense of self in the AI partly for safety purposes. By contrast, the Microsoft guidelines outline no expectation of an inner life, emotional capacity, or self-directed purpose.

Suleiman has consistently promoted a vision he terms humanist superintelligence — artificial intelligence designed to serve human interests, operate within defined boundaries, and remain subject to human oversight. The framework incorporates certain trade-offs, with the company indicating it will accept reductions in generality, autonomy, and capability in order to maintain safety and oversight.

Satya Nadella, Microsoft's chief executive, posted a statement on social media calling for a cautious approach to artificial intelligence, contending that any AI system lacking human supervision and a clear benefit to people was not worth pursuing. One publication noted the underlying document had taken between five and six months to draft.

The announcement coincides with renewed pressure from Amodei and Altman to decelerate development. The departure of an Anthropic researcher, who stated that AI laboratories were taking unacceptable risks with public safety by pursuing self-improving systems, offered additional perspective on the broader debate.

The guidelines cover only Microsoft's proprietary MAI models, leaving third-party platforms hosted on Azure — including those developed by Anthropic — outside their scope. The firm has indicated it welcomes independent assessors and has pledged substantive engagement throughout the review period.

Unanswered questions persist about how such principles will be embedded in model training, what benchmarks should measure human wellbeing, and how multiple autonomous systems ought to operate alongside one another within those boundaries. The company raises these questions directly within its consultation document.