Satya Nadella has backed deliberate pacing of advanced AI development while arguing that its benefits and oversight should extend beyond a small group of powerful organisations.
In a statement on X, Microsoft’s chief executive links human control with business independence: organisations should retain their knowledge, choose providers and control models containing their expertise. He supports both open-source and closed models, and welcomes outside evaluators with broad international and academic representation.
What is new?
Nadella has set out Microsoft’s position in the AI safety debate and promised an MAI Code of Conduct for public consultation. This post does not announce a product launch, a training halt or a new customer entitlement.
What embedded evaluators would do
The proposal has a specific meaning in Dario Amodei’s September essay. Rather than relying entirely on a developer’s own assessment, independent evaluators would work inside AI labs with extensive access to systems and information. Amodei says Anthropic will pursue that approach.
His essay also calls for coordination among democracies and, eventually, globally. It distinguishes pacing development from simply stopping training. These are proposals for managing increasingly capable systems, rather than a claim that AI development has already been paused.
Source: Dario Amodei: We Must Pace the Frontier
Our earlier report on Amodei, Musk and Altman explains the background. Nadella’s contribution brings the question of who participates in oversight alongside the question of how quickly development should proceed.
The test for any evaluation programme is practical: what can reviewers inspect, can they publish uncomfortable findings, and what happens when they identify a serious problem? Welcoming evaluation is a starting point. Its authority, access and independence determine how useful it becomes.
Model choice is already part of Microsoft’s stated policy
There is a longer history behind the argument about avoiding dependence on one supplier. Microsoft’s AI Access Principles, published in February 2024, committed the company to supporting proprietary and open-source models and giving developers flexibility over which models they use.
Those principles also included a commitment to help Azure customers export and transfer their data when switching cloud providers. They are useful background to the latest statement, but should not be mistaken for a new September 2026 feature announcement.
Source: Microsoft’s AI Access Principles, February 2024
For readers, there are two separate questions here. Can you choose another model within a service? And can you leave that service with the material your organisation needs to continue operating? A supplier can offer several models without answering the second question completely.
The practical question for an Australian business
Imagine a small business building an assistant around years of procedures, product notes and corrections from staff. The useful asset is the accumulated knowledge and the work spent making answers dependable. A cheap introductory subscription is only one part of the decision.
Before relying on the assistant, ask the supplier to demonstrate an export using a small sample of your material. Establish what comes out, which settings or customisations remain behind, and what another tool would need to use the export. This is an illustrative purchasing question, not a test of a particular Microsoft product.
It also helps to separate a model from the application around it. Owning a set of documents is different from controlling a customised model. Neither phrase, on its own, tells you whether the surrounding workflow is portable.
The MAI consultation is the next item to verify
Nadella promises a Code of Conduct for Microsoft’s own MAI models for consultation, describing the timing as “tomorrow”. We have not verified that document’s publication. This refers to a conduct document, not a promise to release model source code.
Once available, the useful questions will be specific: which models it covers, how public feedback can change it, and what enforcement or reporting mechanisms accompany it. An existing Microsoft document with a similar title should not be treated as the promised consultation without confirmation.
What to take from this statement
Our reading is that safety and customer independence belong in the same conversation. People need meaningful control over systems they rely on, and meaningful alternatives when those systems no longer suit them.
The next step is to examine the mechanisms behind the commitments. An evaluator’s access agreement, an export that actually works and a consultation with a clear response process would tell readers more than another statement of principles.
Michel Elijah covers technology, streaming and online security for Tech X Telco. He writes practical how-to guides on everything from email troubleshooting to spotting the latest scams doing the rounds in Australia, with a focus on clear steps anyone can follow.
Spotted something wrong? Corrections are recorded in the open. Report a correction
