Microsoft announced in September 2026 a new AI code of conduct that makes human control the top priority, bars the development of models deemed unsafe, and explicitly rejects any notion that its systems have consciousness, inner life or rights.
Key provisions of the code
The Decoder reports that the document sits at the top of Microsoft’s internal rulebook and guides training, technical controls and evaluation of its "MAI" models. The core rule, according to the source, is that “human control comes first.” Microsoft says it is prepared to sacrifice “generality, autonomy, or performance” to preserve that control.
AI chief Mustafa Suleyman is quoted as telling The Information, “If it isn’t safe we shouldn’t build it.” The code therefore includes a safety‑first building rule: any model judged unsafe must not be built. It also bans the use of “Neuralese” or other incomprehensible communication forms, whether in internal reasoning or when models talk to each other.
Finally, the code states that Microsoft’s AI should not claim to have feelings, motivations or rights. The company draws a clear line between its models and the anthropomorphic language sometimes used by competitors.
Sector‑wide implications
Microsoft’s stance arrives amid a wave of calls for slower, more responsible AI development. By codifying human‑control primacy, the firm signals that large‑scale model developers may need to embed similar safeguards to stay competitive, especially as customers and regulators scrutinise opaque AI behavior.
The prohibition on “Neuralese” could pressure other firms to make model outputs more interpretable, a trend already evident in academic research. If major cloud providers adopt comparable rules, developers may face higher engineering costs to meet transparency standards.
Because the code applies only to Microsoft’s own models, third‑party models running on Azure are not automatically covered. However, the public consultation that concluded in August 2026 – a six‑week process documented by The Decoder – suggests Microsoft is gathering broader stakeholder input. A revised version is slated for release later in 2026, with the updated rulebook expected to guide model development from 2027 onward.
Microsoft’s financial backdrop
Understanding the company’s capacity to enforce the new policy benefits from a look at its recent financial statements, filed with the SEC. The most recent 10‑K, dated 29 July 2026, shows a net income of $133.749 billion for fiscal year 2026, total assets of $758.376 billion, shareholders’ equity of $442.387 billion and 7.427 billion shares outstanding. By contrast, the latest quarterly filing from 2011 records revenue of $36.148 billion for the six‑month period ending 31 December 2010.
| Metric | Value | Period | Unit |
|---|---|---|---|
| Revenue (10‑Q) | 36,148,000,000 | 2010‑07‑01 to 2010‑12‑31 | USD |
| Net income (10‑K) | 133,749,000,000 | 2025‑07‑01 to 2026‑06‑30 | USD |
| Total assets (10‑K) | 758,376,000,000 | 2026‑06‑30 | USD |
| Shareholders’ equity (10‑K) | 442,387,000,000 | 2026‑06‑30 | USD |
| Shares outstanding (10‑K) | 7,427,000,000 | 2026‑06‑30 | shares |
| Source: Microsoft SEC filings, 10‑Q (2011) and 10‑K (2026) | |||
These figures underscore Microsoft’s scale: with more than 221,000 employees and a market‑cap that comfortably exceeds its equity, the company can allocate substantial resources to compliance and monitoring of its AI governance framework.
Outlook and unanswered questions
The code’s immediate effect is internal – Microsoft will not train models that fail its safety test, and developers will need to avoid “Neuralese” communication. The broader industry impact depends on whether competitors adopt comparable rules or whether regulators codify similar requirements.
- Will cloud customers demand that Azure‑hosted third‑party models adhere to the same human‑control standards?
- How will enforcement be measured? The packet does not detail audit mechanisms or penalties.
- What impact will the rule against “Neuralese” have on research that relies on emergent communication among AI agents?
These gaps are acknowledged in the research packet: “The story meets the publication’s criteria… No contradictory information was found.” The lack of concrete enforcement details means the sector will watch closely for any follow‑up guidance from Microsoft later in 2026.
In the meantime, the code adds a new layer to the ongoing debate over AI safety. By making human oversight a non‑negotiable principle, Microsoft may set a de‑facto benchmark that shapes both private‑sector best practices and future regulatory proposals.