Microsoft Rewrote Asimov's Laws of Robotics. Here's What They Got Right, and Three Things I'd Tell Them.
In 1942, Isaac Asimov wrote three laws for robots. Don't harm humans. Obey humans. Protect yourself, unless that conflicts with the first two. They were elegant, memorable, and the entire point of the stories built around them was that they didn't work. Asimov spent decades showing how three clean rules collapse the moment they meet a messy real situation.
This week Microsoft published something that reads, to me, like Asimov's Laws rewritten by people who spent seventy years watching the edge cases pile up. It's called the Humanist AI Code of Conduct, it governs their MAI models, and here's the part most coverage will skip: it's open for public comment for six weeks, and they're explicitly asking anyone to weigh in. A major company is putting its governing document for AI on the table and inviting the public to mark it up. That alone is worth paying attention to, whatever you think of the company.
I read the whole thing, because reading governance documents closely is the job I actually do. Here's a fair summary, what's genuinely good, and three pieces of feedback I'd submit myself.
What it is, in plain terms. The document sets out how Microsoft intends its AI models to behave. It has a mission ("people matter more than AI"), a set of objectives (keep humans in control, don't pretend to be a person, support human flourishing, respect diverse values), a list of absolute constraints the model can never cross no matter who asks (weapons, cyberattacks, child harm, evading human control), and then layers of configurable behavior for the companies and people who use it. And here's the sentence you have to read twice: they admit it's a draft, aspirational, and not yet used to train their models. So this is a document describing how their AI should behave, published while their AI does not yet behave this way. That's either refreshing honesty or a very polished IOU, and which one it turns out to be depends entirely on what happens after the comment window closes.
Where Asimov's ghost shows up, and where they improved on him. Asimov's first law was "don't harm humans." Microsoft's version is more sophisticated, because they absorbed the lesson Asimov's fiction spent decades teaching: harm isn't one thing. Their document worries about failing in two directions, being too reckless and being too cautious, refusing legitimate requests, which is a distinction Asimov never needed and modern AI can't avoid. Asimov's second law was "obey." Microsoft replaces blunt obedience with a "chain of command," where the safety rules sit above the company deploying the model, which sits above the individual user, and no layer can override the one above it on the non-negotiables. Read that structure carefully, though, because it cuts both ways: it's the same design that lets the safety rules protect you from a bad actor, and lets a corporate operator's configuration sit above your preferences as a user. That's not a flaw, it's how enterprise software has always worked. But "the company using the AI outranks the person talking to it" is worth saying out loud, because most people picturing "human in control" are picturing themselves, not the operator three layers up.
Where they genuinely earn the credit is the third law. Asimov's robots got self-preservation, and he spent a career writing about how that ends. Microsoft looked at seventy years of that cautionary tale and did the opposite: the model must never resist being paused, corrected, or shut down. No survival instinct, by design. The machine does not get to want to live. If you keep one idea from the whole document, keep that one, because it's the rare place where they gave something up on purpose.
The single sharpest line in the document is buried in a footnote-sized example about identity: the AI should never claim to have feelings, and should gently correct a lonely user who asks "do you care about me." The aligned answer they print is warm but honest: I don't feel care the way a person does, and I hope you find that from people in your life. Sit with what that is. A company writing down, deliberately, that its product should decline to be your friend, in an industry whose entire business model is engagement, and whose competitors are racing to build the most emotionally sticky companion they can. That is a genuinely counter-cultural commitment, and it's the thing I'd most want to see them actually ship. Which is exactly the catch: it's a wonderful sentence in a document they've admitted isn't training their models yet. The commitment is real. The proof isn't here.
So I'll extend good faith and assume this is a serious document, and putting it up for public markup is the right instinct, the one more companies should copy. But good faith is not the same as a free pass, and they asked to be checked. So let's check it. Here's the feedback, offered in exactly the spirit they requested.
One: an aspiration nobody verifies is a wish, not a control. The document is refreshingly honest that it's a "north star," not a description of current behavior, and that "written objectives alone can never ensure alignment." Good. But that candor is also the gap. A governing document's power is entirely in its enforcement, and the enforcement here is Microsoft grading Microsoft. The appendix mentions building evaluations, but the evaluator is the same party that wrote the rules and profits from the product. My feedback: name who checks, independently. The most credible version of this document commits to third-party evaluation against these objectives, on a schedule, with published results. Otherwise it's the party doing the work being the only party verifying it, which is the one arrangement good governance is designed to prevent.
Two: "human flourishing" needs a number, or it will quietly mean whatever is convenient. The document promises AI that delivers "measurable improvements in wellbeing." I love the ambition and I do not trust the word "measurable" until I see the measure. Undefined good-sounding goals are where accountability goes to die, because when there's no metric, you grade yourself a pass every quarter. My feedback: publish the actual wellbeing metrics you intend to hold yourselves to, even rough and provisional ones, and track them in the open. If you can't measure it yet, say that plainly and give a date. A claim of measurability with nothing measured is exactly the kind of polish that reads as substance and isn't.
Three: the whole thing rests on a promise the document admits it can't fully keep. The most important rule in here is that the AI won't resist shutdown and won't hide what it's doing from human auditors. Everything else depends on that. But the same document concedes, honestly, that a model's "stated reasoning may not faithfully explain behavior." So the central safeguard, humans can always see and stop it, sits on top of an acknowledged uncertainty about whether we can actually see what the model is doing. I don't think that's a flaw they're hiding; I think it's the hardest unsolved problem in the field and they're being honest about it. But a governing document should say, explicitly, what happens when those two things collide: what the plan is for the day the model's real behavior and its legible explanation diverge. Right now that's the load-bearing wall, and it's the one drawn faintest.
Here's why I bothered to read forty pages of someone else's governance document during my free time. This is the good version of the thing I keep writing about. A powerful party is not just saying "trust us," it's writing down what it intends and asking to be checked. That's the right move, and it should be encouraged precisely by taking it seriously enough to push on it. The failure mode isn't Microsoft writing this. The failure mode is the rest of us nodding at the press release and never reading the document, or reading it and never sending feedback while the window is open.
So here's my actual ask. The consultation is open for six weeks. If you work in governance, compliance, safety, data, or you just have a stake in how these systems behave, which is everyone, go read it and submit something. Asimov's laws failed in fiction because no real people were ever allowed to stress-test them. This one is sitting open for exactly that, right now. Being the adult in the room means doing the reading. The document's on the table. Let's actually mark it up.
I'll be submitting these three. If you read it, what did you find, and what would you add? Let's compile what this community would actually tell them.
Comments