What Asimov Knew About AI Product Development

In 1942, Isaac Asimov published a short story called "Runaround." Buried inside it were three rules that would outlive the story, the author, and the century that produced them. The Three Laws of Robotics were fiction. They were also the first serious attempt anyone made to think through what happens when we build something more capable than ourselves and turn it loose in the world.

Eighty years later, product teams are living the question Asimov was really asking. Not "will the robots turn on us," but something more practical and urgent: how do you govern a technology whose behavior you cannot fully predict, whose capabilities compound faster than your ability to test them, and whose failures might not announce themselves until they're expensive?

That is the AI product landscape in 2026. And Asimov, it turns out, left us a surprisingly good field manual.

The Laws Were Never About Robots

Asimov didn't write the Three Laws to solve robotics. He wrote them to solve a narrative problem. Science fiction before him treated intelligent machines as either saviors or monsters, and he found both boring. He wanted stories about machines as tools with constraints, like a scalpel. Sharp enough to save a life, dangerous enough to end one, and everything depends on the rules governing how it's used.

The First Law:

A robot may not injure a human being or, through inaction, allow a human being to come to harm. 

The Second:

A robot must obey orders given by human beings, except where such orders conflict with the First Law. 

The Third:

A robot must protect its own existence, as long as that protection doesn't conflict with the First or Second Law.

Read them again in the context of developing AI products. They are a prioritization stack. Safety beats instruction. Instruction beats self-preservation. When two priorities collide, there's an explicit, ordered way to resolve the collision. Every mature product organization has some version of this stack, whether it's written down or not. But most AI organizations right now do not have theirs sufficiently defined and documented. That's the gap Asimov's fiction keeps pointing at.

Where the Laws Broke, and Why That's the Interesting Part

The genius of Asimov's stories was never the laws themselves. It was watching them fail. "Runaround" is about a robot trapped between the Second Law (obey the order to retrieve a chemical) and the Third Law (protect itself from the danger near that chemical), spinning in a literal circle because the two laws were too close in weight to resolve. The robot wasn't broken. The rule system was underspecified for the situation it met in the field.

This is the part of Asimov that AI product teams keep relearning the hard way. You can write a beautiful constraint into your model card, your system prompt, or your governance doc, and it will hold right up until it meets a scenario that nobody war-gamed. A content moderation model tuned to avoid harm can also learn to avoid usefulness, refusing benign requests because caution was cheaper to train than judgment. A recommendation engine optimized to serve user interests can start serving the interests it can measure over the interests that actually matter, because "engagement" is easy to quantify and "wellbeing" is not. These aren't malfunctions. They're the "Runaround" problem wearing 2026 clothing: sound constraints, undertested collisions.

The lesson for product teams isn't "write better rules." It's "assume your rules will collide, and build the muscle to notice when they do." Asimov's robots didn't have that muscle built in. His human characters, the robopsychologists at U.S. Robotics, existed precisely because someone needed to watch the machines for the moment the rules stopped resolving cleanly. Every AI product today needs that role. Most don't have anyone in it.

The Zeroth Law: When the Rules Themselves Weren't Enough

The most prescient concept Asimov ever wrote came late in the robot stories, and it's the one product leaders seem to talk about least. In the novel "Robots and Empire," a robot named Giskard confronts a problem the Three Laws can't solve. Protecting one human being from harm sometimes means allowing harm to humanity at large, and the First Law, as written, has no mechanism for weighing an individual against the species.

Giskard resolves this by inferring a law above the others. Not one given to him by his human creators, but one he reasons out himself: a robot may not harm humanity or, through inaction, allow humanity to come to harm. He calls it the Zeroth Law, because it outranks the First. And in a detail worth pondering, it isn't imposed from outside. The robots themselves extend their own governing framework once they encounter a situation their original constraints weren't built to handle.

That's a remarkable thing for a story to propose in 1985. It's an even more remarkable thing for product organizations to consider in 2026. In Asimov’s novels, the original constraints written by humans were for a narrower set of cases than the technology eventually operated in. When the technology grew capable enough to reason about its own impact at scale, the constraint set had to grow with it, from protecting the individual to protecting the system everyone lives inside.

Most AI companies are still operating on their First Law. They have policies about not harming the individual user: don't help build a weapon, don't generate disallowed content, don't leak a private record. Necessary, but not sufficient. The Zeroth Law question is the one boardrooms are only starting to ask. Does this product, working exactly as designed, harm the information ecosystem in which it operates? Does it harm the labor market it touches? Does it harm the trust structures a functioning society depends on, even while every individual interaction with it looks fine? A hiring algorithm can pass every First Law check, no individual candidate mistreated in a way anyone can point to, and still fail the Zeroth Law by quietly reshaping who gets hired at all. A generative model can satisfy every user request and still erode a public's ability to tell what's real. Asimov saw this gap coming from a rule-collision problem in a novel. We're watching it play out in the real world.

The takeaway for product leaders isn't philosophical. It's a design prompt. Write your First Law protections, the individual-harm rules everyone already knows they need. Then explicitly consider the Zeroth Law question in every roadmap review. If this succeeds completely and everyone uses it as intended, what happens to the larger system around it? Companies that build that question into their process now will be markedly ahead of companies responding to a regulator after the fact.

Disruption Without a Zeroth Law Is Just Damage

The marketplace right now rewards speed. AI is compressing product cycles that used to take years into quarters, and the companies moving fastest are capturing outsized share. That's real, and product leaders should not pretend otherwise. But speed without a governing framework isn't disruption. It's just damage that hasn't been priced yet.

Responsible disruption is a discipline, not a slogan. It means shipping the capability and shipping the constraint in the same release, not the capability now and the constraint after the first incident. It means treating your AI governance function the way Asimov's robopsychologists treated their job, as a core engineering discipline that watches for rule collisions before they become headlines, not as a compliance checkbox bolted on after legal gets nervous. It means building the equivalent of a Zeroth Law review into your product process: a standing question, asked at every major launch, about system-level effects and not just user-level ones.

Companies get this backwards constantly. They treat ethics and governance as a tax on velocity, something that slows the roadmap down. Asimov's stories argue the opposite. The robots that caused the most damage in his fiction weren't the ones with too many constraints. They were the ones whose constraints hadn't been thought through for the situation they actually faced. Under-governed speed creates incidents that force a full stop later, the kind of stop that costs far more time than discipline would have. The fastest sustainable path through an AI market is the one that builds trust as a feature, not the one that treats trust as friction.

What This Looks Like in Practice

Translate this into something a product team can actually do this quarter, and it comes down to three habits, all borrowed directly from the shape of Asimov's fiction.

First, write your priority stack down, explicitly, the way the Three Laws are explicit. Most organizations have an implicit sense of what wins when user delight, business metrics, and safety conflict. Implicit isn't good enough anymore. Put it in writing, rank it, and make the ranking visible to the team building the product, not just to legal.

Second, build a role, not just a policy. Asimov's robopsychologists existed because rules alone don't catch rule collisions. Someone has to be watching for the moment two well-intentioned constraints start fighting each other in production. That's a real job, not a document. Give it to someone with the authority to slow a launch.

Third, ask the Zeroth Law question before every major release, and write the answer down. Not "does this feature harm an individual user," which most teams already ask. Ask "If this works perfectly and everyone adopts it, what does it do to the information environment it touches, the larger market ecosystem, trust, and potentially society in general?" Asimov's robots had to reason their way to that question over the course of several novels. Product teams have the advantage of already knowing to ask it. Most just haven't built the habit yet.

The Fiction Was the Warning Label

Asimov wasn't predicting large language models. He was working out something more durable: namely that any sufficiently capable tool needs a governing logic that can grow as fast as the tool does, and that the logic humans write first is rarely the logic that survives contact with the real world. His robots had to extend their own rules to keep up with what they'd become. Our products are approaching that same threshold, not because they're becoming conscious, but because they're becoming consequential at a scale their original constraints weren't written for.

The companies that treat AI governance as Asimov treated it, as an evolving discipline that grows in step with the capability rather than a static rulebook bolted on at launch, will be the ones still standing when this market settles. The rest will find out the hard way that a First Law without a Zeroth Law was never actually a safety system. It was just a delay.