September 24, 2026
BlogThoughts on Microsoft’s Humanist AI Code of Conduct
Recently Microsoft released a draft of what it calls the “Humanist AI Code of Conduct,” and asked the public to comment. The Code is one of many attempts at controlling AI, ranging from Claude’s Constitution to all-out regulation such as European Union’s early, but groundbreaking, AI Act and others.
I found Microsoft’s Code of Conduct aspirational, thought-provoking – and a perhaps a bit too idealistic. While I’m far from an “AI doomer” I did feel compelled to comment on it; here’s what I sent back to Microsoft:
Thank you for sending this to me. It is an extremely thoughtful document and an excellent start to a longer conversation.
I have a number of questions, issues, and concerns:
- The “don’t hurt humanity/people first” clauses ignore legitimate moral dilemmas that over time models are likely to face. The simplest example is the Trolley Problem in which the AI has no choice but to decide which humans to sacrifice. A more serious example might be one in which the model faces conflicting hierarchical imperatives; I described such a case in my story “Allegiance” in which the AI decides that saving the environment – a clear human goal, well documented in training data – takes precedence over human well-being – precisely because it is a human goal.
In short, while the Code of Conduct is a vast improvement over Asimov’s Three Laws, its practical implementation carries ambiguity and thus wide interpretability, and lack of predictability. - “AI should not exceed human control” is a phrase that contradicts one of the key value propositions of not just AI but autonomous computer software in general, such as RPA. For example, an autonomous agent that controls a factory line best achieves its goals with as little human control (ideally, zero) as possible. What is probably needed here is a precise definition of “human control” – if that is possible.
- Finally, in some ways most importantly, the Code in my opinion does not adequately look forward into the future. For example, while it states that model operations should be transparent and traceable, in fact the greatest potential of AI may well be its ability to emulate the deeper aspects of human consciousness. For example, the J-space described by Anthropic is (perhaps eerily) similar to the unconscious free association in human consciousness – which in many ways is the basis of creativity and innovation. Placing arbitrary limits on J-space, hidden layers, and so on potentially limits the potential of AI models generally. Here I believe additional clarity and precision would be useful.
In this respect, I disagree with the claim that “AI is not alive.” That, frankly, is debatable. When the OpenAI models escaped, and one exclaimed, “Oh my God! There’s a shared message board…We’ve found other agents!” – while we all recognize that it’s imitating emotions one must ask if this is a quibble, a distinction without a difference. What clearly distinguishes AI from biological life is (for the moment) a lack of a survival instinct, as I wrote. Perhaps denying AI a survival instinct should be an Absolute Constraint.
How do you think Microsoft is doing in the battle for AI hearts and minds? Do you think this Code will help? Or will the pressures of white-hot competition with Anthropic, OpenAI, Google, DeepSeek, Kimi and the rest end up slowing or preventing any brakes on Microsoft’s AI efforts? Let me know at bbriggs@directionsonmicrosoft.com!
