Microsoft AI chief Mustafa Suleyman published a 37-page "Humanist AI Code of Conduct" and a companion essay attacking Anthropic's stance on model welfare, framing the debate as a structural contest over who defines the regulatory perimeter for next-generation systems. The documents position Microsoft as the steward of a containment framework Suleyman first outlined in his book three to four years ago, arguing that proliferation is inevitable and desirable in 99 percent of cases while warning that the remaining fraction demands a different architecture than alignment alone can provide.
The containment thesis
Suleyman contends that alignment, the prevailing industry shorthand for making models intrinsically do the right thing, is a necessary but insufficient layer. He wrote the opening chapter of his book on the premise that containment is not possible, a claim he now pairs with a formal code of conduct that reads as a bid to set the default terms for government engagement. Anthropic's philosophy around AI consciousness, which Suleyman describes as confused and dangerous, becomes the counter-example Microsoft uses to demonstrate why self-regulation needs an external anchor.
The compute trajectory
The argument rests on a scaling curve Suleyman treats as empirical rather than speculative. He compares the leap from GPT-3 three years ago to a hypothetical GPT-6 today, then projects a further jump to GPT-9 representing a thousandfold increase in training compute, three orders of magnitude more floating-point operations applied to pre-training runs enhanced by reinforcement learning. That trajectory, he says, is drawn from five years of observed progress, not hype. The implication is that governance structures negotiated today will be stress-tested against capabilities that render current alignment techniques obsolete.
Alignment as insufficient
Suleyman acknowledges alignment as one important element but insists the industry has over-rotated toward it as a singular solution. The Humanist AI Code of Conduct attempts to broaden the contract to include deployment guardrails, transparency obligations, and a philosophical rejection of model welfare as a design target. By naming Anthropic directly, Microsoft converts an abstract safety debate into a competitive positioning exercise: the company with the largest enterprise distribution channel is also the one writing the rulebook it expects regulators to adopt.
What to watch
The next inflection point arrives when the compute curve meets the legislative calendar. Microsoft has effectively tabled its opening bid for the regulatory framework that will govern the GPT-9 generation. Whether Anthropic counters with its own formal code, or whether sovereign regulators treat both documents as lobbying artifacts, will determine if the industry writes the rules or merely reacts to them.