Sacrifice to Moloch

A few years ago, AI still felt somewhat distant. Most people knew it existed, of course, but it seemed to belong to technology companies, university researchers and the occasional enthusiast who enjoyed experimenting with tools that appeared mysterious to everyone else. There was an assumption that AI would eventually become important, but for many people it remained something happening somewhere else. Then, seemingly all at once, it arrived. Large language models are now writing reports, generating images, producing software code and increasingly appearing in everyday business processes. Even people who have little interest in technology are discussing AI over coffee, in meetings and on social media.

As I have tried to understand what is happening, I have spent a lot of time separating genuine capability from hype. That is not always easy because the language surrounding AI can be surprisingly misleading. In a couple of previous articles, AI Think, Therefore AI Am and Mind Your Language I wrote about the way we casually describe AI systems using words such as thinking, understanding and reasoning. Those words come loaded with assumptions. When we say that a person understands something, we generally mean they comprehend it, can apply judgement and appreciate the wider context. Current AI systems do not work in that way, even when the results they produce can sometimes make it feel as though they do.

Recently I completed a number of Microsoft-accredited LinkedIn Learning courses, including AI for Organizational Leaders and Human Skills in the Age of AI. What interested me was that the courses were not really focused on teaching people how to press the right buttons or write better prompts. Much of the content concentrated on responsibility, ethics and the relationship between humans and machines. Again and again, the message was that AI should be viewed as a collaborator rather than a replacement.

In principle, that makes perfect sense. AI is extremely good at certain kinds of tasks. It can process information at extraordinary speed, identify patterns across large datasets and generate possibilities that might take a person much longer to explore. Humans bring something different. We bring experience, context, creativity, judgement and an understanding of consequences that stretch beyond immediate outputs. The best outcomes will almost certainly come from combining those strengths rather than treating them as competitors.

AI is extremely good at certain kinds of tasks. It can process information at extraordinary speed, identify patterns across large datasets and generate possibilities that might take a person much longer to explore. Humans bring something different. We bring experience, context, creativity, judgement and an understanding of consequences that stretch beyond immediate outputs. The best outcomes will almost certainly come from combining those strengths rather than treating them as competitors.

The trouble is that when you step outside training courses and look at the broader public discussion, collaboration often feels like a secondary concern.

Much of the debate is focused on capability. Which model is strongest? Which company is ahead? How quickly can work be automated? Who will achieve artificial general intelligence first? Those are interesting questions and they attract a great deal of attention, but I increasingly find myself wondering whether they distract us from a more important issue.

What happens when everybody feels they have to keep moving faster, even when nobody is entirely certain where they are going?

That question led me towards an idea that appears in game theory and philosophy, and one that has become increasingly relevant when I think about AI. The idea is known as the Moloch principle or problem.

Historically, Moloch was associated with an ancient deity mentioned in the Hebrew Bible and linked to practices that were regarded as some of the most abhorrent acts imaginable. The historical details remain debated, and scholars still disagree on exactly how Moloch was worshipped and how much of the biblical account reflects historical reality rather than warning. What is beyond dispute, however, is the symbolic role Moloch came to occupy. Within Jewish and later Christian traditions, Moloch became associated with the sacrifice of children, often through fire, in exchange for perceived prosperity, security, military success or divine favour. The image is deliberately shocking. Parents surrendering what is most precious in pursuit of something they believe they desperately need.

Over the centuries, Moloch evolved from the name of a deity into something much larger. It became a symbol of systems that demand sacrifice. Not sacrifice freely chosen because it is good or noble, but sacrifice made because people feel trapped by circumstances. The horror of the story does not lie only in the act itself. It lies in the fact that the participants often believed they had little alternative. They were not necessarily sacrificing their children because they hated them. Quite the reverse. They may have believed the sacrifice was necessary to protect their community, secure their future or avoid disaster. In other words, the very thing they valued most was surrendered in pursuit of a goal that seemed impossible to refuse.

More recently, writer and philosopher Scott Alexander revived the concept as a way of describing situations where individuals make rational decisions that collectively produce outcomes nobody actually wants. In this modern sense, Moloch is not a god. It is a system of incentives that pushes people towards choices they might never make if they were acting together rather than competing with one another.

What makes the idea so interesting is that it does not require anyone to be behaving badly. In fact, it works best when everyone is acting sensibly.

Moloch appears when reasonable people, responding to reasonable pressures, end up sacrificing something important. It may be time, wellbeing, creativity, trust or long-term thinking. Nobody intended the loss, but somehow the system produces it anyway. That's why the concept remains so powerful. It reminds us that societies do not always lose what matters through malice or greed. Sometimes they lose it because the pressure to keep up becomes stronger than the ability to step back and ask whether the race itself makes sense.

societies do not always lose what matters through malice or greed. Sometimes they lose it because the pressure to keep up becomes stronger than the ability to step back and ask whether the race itself makes sense.

Imagine several AI companies competing in the same market. Many of them might privately prefer a slower pace of development. They might want more testing, stronger safeguards and a better understanding of the systems they are building. Given complete freedom, they may choose to proceed carefully.

Then a competitor releases a powerful new model. Investors become excited. Journalists start writing about it. Customers begin asking questions. Suddenly caution starts looking less like responsibility and more like a competitive disadvantage. The companies that had hoped to take their time find themselves under pressure to accelerate.

From their perspective, the decision makes sense. If one company moves faster, the others feel compelled to respond. If they do not, they risk losing investment, customers or market share. Each decision is rational when viewed individually. The difficulty is that when everyone reaches the same conclusion, the entire industry can end up moving far faster than anyone originally intended. That is essentially the Moloch problem.

Game theory has examined similar situations for decades. The classic example is the Prisoner's Dilemma, where cooperation would produce the best overall outcome but individual incentives push participants toward self-interest.

The scenario is a simple one. Two suspects are arrested and questioned separately. Each has two choices: they can remain silent and cooperate with the other prisoner, or they can betray the other person in exchange for a reduced sentence. If both of them remain silent, each ot them receive a relatively light punishment. But if one betrays the other while the other remains silent, the betrayer goes free whilst the silent prisoner receives a much harsher sentence. Faced with that uncertainty, both prisoners often choose to betray each other. The result is that they both end up with a worse outcome than if they had cooperated from the start.

In this scenario, neither prisoner is being irrational. In fact, each is acting sensibly based on the information available to them. Instead, the problem is that what is rational for the individual isn't necessarily beneficial for the group. The sound logic that protects each person from being exploited is the very thing that prevents them from achieving the best collective outcome.

the problem is that what is rational for the individual isn't necessarily beneficial for the group. The sound logic that protects each person from being exploited is the very thing that prevents them from achieving the best collective outcome.

In real life, few people set out intending to create a worse result, yet that is often exactly what happens. The Prisoner's Dilemma illustrates how systems can trap people into choices that make perfect sense individually while producing consequences that almost everyone would prefer to avoid.

The AI race shares some of those characteristics. One organisation releases a new model and gains an advantage. Competitors respond by accelerating development. The original company then feels pressure to move even faster because others are catching up. Before long, an escalation dynamic takes hold.

Nobody necessarily wants an uncontrolled race. Nobody wakes up one morning thinking, "Let's create a situation where safety, understanding and reflection become secondary." Yet the incentives pull everyone in that direction anyway.

History provides plenty of examples of this dynamic. During the Cold War, the United States and the Soviet Union became locked in a competition that neither side felt able to step away from. If one side developed a new missile system, the other felt compelled to respond. If one side increased its stockpile, the other feared the consequences of standing still. From the perspective of either government, these decisions could be justified as defensive measures. The goal was not necessarily domination, but avoiding vulnerability.

The result was an arms race that continued for decades. Both sides spent vast sums of money developing increasingly sophisticated nuclear weapons, not because they wanted a world overflowing with nuclear warheads, but because they feared what might happen if their rival gained an advantage. Every step seemed rational when viewed individually. No leader wanted to be the one who left their country exposed.

What stands out from this is that very few people would have designed the final outcome from scratch. If you had asked policymakers at the beginning whether they wanted a world containing enough nuclear weapons to destroy civilisation many times over, most would almost certainly have said no. Yet through thousands of individually rational decisions, that is precisely where they ended up.

If you had asked policymakers at the beginning whether they wanted a world containing enough nuclear weapons to destroy civilisation many times over, most would almost certainly have said no. Yet through thousands of individually rational decisions, that is precisely where they ended up.

This is one of the reasons the Cold War remains such a powerful example of a Moloch-like dynamic. The system didn't require irrational actors or reckless decision-makers. In many cases, it involved intelligent people responding logically to the incentives and threats they perceived around them. The problem was that each side's attempts to increase its own security often made the other side feel less secure, triggering further escalation. Actions intended to reduce risk for one player ultimately increased risk for everyone.

The lesson here isn't really about nuclear weapons. It's about how competition can trap people, organisations and even entire nations in cycles that become difficult to escape. Once the race begins, slowing down can feel dangerous, even when everyone involved recognises that continuing to accelerate may lead somewhere none of them actually want to go.

AI is obviously very different from nuclear weapons, and comparisons should be treated carefully. Nevertheless, the underlying mechanism feels strangely familiar. Competition changes behaviour. People begin responding not only to their own objectives but to the actions of everyone around them.

I see a version of this every day on LinkedIn. Over the last year my feed has been full of posts encouraging organisations to adopt AI immediately. The wording changes slightly but the message remains remarkably consistent. "Use AI now. Your competitors already are." "Don't get left behind." "The future belongs to those who embrace AI today." And some of these claims contain a great deal of truth. Organisations that ignore significant technological change rarely benefit from doing so. At the same time, the language creates a sense of urgency that can be difficult to ignore. AI is presented not merely as an opportunity but as a race. And races have a habit of making people focus on speed rather than direction.

Organisations that ignore significant technological change rarely benefit from doing so. At the same time, the language creates a sense of urgency that can be difficult to ignore. AI is presented not merely as an opportunity but as a race. And races have a habit of making people focus on speed rather than direction.

The potential danger here is that we're so focused on winning the AI race that we've forgotten to define the destination. As the Cheshire Cat said in Alice in Wonderland “if you don’t know where you are going, any road will get you there”. The question is whether we are choosing the road, or whether the race itself is choosing it for us.

I wrote recently about what I called The LinkedInverse, the curious tendency for professional social media platforms to reward visibility and engagement over substance. The most successful posts are not always the most thoughtful. They're often the ones that trigger the strongest reaction. Dramatic predictions tend to travel further than careful analysis.

I suspect something similar is happening with AI discussions. A post claiming that AI will transform everything within twelve months will usually attract more attention than a post asking whether we have properly defined the problem we are trying to solve. Yet the second question is probably the more useful one.

Fear of missing out plays an important role here. Nobody wants to discover five years from now that they ignored a genuinely transformative technology. Nobody wants to explain why competitors surged ahead while they remained still. Nobody wants to become a case study in organisational complacency. Fear is a remarkably effective motivator. The issue is that fear tends to encourage action long before it encourages reflection.

Nobody wants to discover five years from now that they ignored a genuinely transformative technology. Nobody wants to explain why competitors surged ahead while they remained still. Nobody wants to become a case study in organisational complacency. Fear is a remarkably effective motivator. The issue is that fear tends to encourage action long before it encourages reflection.

Before adopting any significant technology, we should probably ask a few straightforward questions. What problem are we trying to solve? What value will this create? What risks are we introducing? Where does human judgement remain essential? How will we know whether success has actually been achieved? Yet those questions are often overshadowed by a simpler one. How quickly can we implement AI?

Once that becomes the dominant question, the technology itself starts becoming the objective rather than a means of achieving something useful. I don't think most organisations are intentionally behaving irresponsibly. They're responding rationally to the environment around them. They see competitors investing heavily in AI. They hear investors asking about AI strategies. They encounter consultants and commentators confidently predicting dramatic disruption. Under those circumstances, doing nothing can feel more dangerous than doing something.

The result is that entire industries may begin moving in the same direction without ever pausing long enough to examine whether the destination is actually desirable. That concern becomes even more interesting when we look at the gap between what is said and what is rewarded. The language surrounding AI is full of discussions about ethics, responsibility and human-centred design. We hear constant references to collaboration. We are told that people and AI will work together, each contributing their unique strengths. Yet many conversations quickly drift toward cost reduction, workforce reduction and automation.

There is an obvious tension there. Using technology to remove repetitive work and make people more effective is one thing, but using technology primarily to reduce the need for people is something rather different. Both approaches may involve the same tools, but they are driven by very different priorities.

Using technology to remove repetitive work and make people more effective is one thing, but using technology primarily to reduce the need for people is something rather different. Both approaches may involve the same tools, but they are driven by very different priorities.

Another area where this dynamic appears is in the race for capability itself. There is enormous pressure to build larger, more powerful models because capability is relatively easy to measure. Benchmarks produce numbers. Demonstrations generate headlines. Performance improvements can be presented in neat charts and graphs. But understanding is much harder to quantify.

A system may produce impressive outputs while still making mistakes, exhibiting biases or failing unpredictably in situations it was not designed for. Sometimes it feels as though we are asking whether something can be built before we've fully considered whether it's understood.

The same pattern appears in environmental discussions. AI requires significant computing resources. Data centres consume electricity and water on a vast scale. A single organisation choosing to slow down will have little impact on its own, which means there's very little incentive for any individual actor to change course. Yet the collective effect of everyone accelerating simultaneously becomes increasingly difficult to ignore. Again, the tension lies between individual interests and collective outcomes.

One of the ideas that helped crystallise this for me came from a recent talk by Jonathan Pageau at the Alliance for Responsible Citizenship conference 2026. What I found particularly interesting was that he moved beyond the usual debates about capability, employment or regulation and focused on a more fundamental question.

Whom does AI serve? (Yes I know that should be “who” not “whom”, but I didn’t give the talk)

At first glance it seems almost too simple a question to matter. But the more I thought about it, though, the more important it seemed.

We spend enormous amounts of time discussing what AI can do. We talk about bigger models, greater automation and increased productivity. Much less time is spent discussing what purpose these systems should ultimately serve. Are we building them to help people make better decisions, solve difficult problems and improve human wellbeing? Or are we gradually restructuring organisations, workplaces and even our own behaviour around the demands of increasingly complex technological systems?

History suggests that systems created by humans often end up shaping human behaviour in return. We adapt ourselves to the system. We reorganise priorities around whatever the system rewards. Before long, success starts being measured according to criteria that may not have been particularly important in the first place. That concerns me more than speculative discussions about superintelligence.

There are certain things that remain difficult to measure but are central to human life. Wisdom. Judgement. Creativity. Responsibility. Relationships. These qualities rarely fit neatly into productivity metrics, and yet they matter enormously.

My worry is that in the rush toward AI adoption we may gradually begin valuing only what is easy to count. Faster outputs. Lower costs. Greater efficiency. Improved automation. Those things matter but they’re not the whole story.

When I think about the future of AI, I find myself less concerned about machines suddenly taking control and more concerned about the possibility that humans create competitive environments in which nobody feels able to slow down. A company says it would prefer more testing, but competitors are moving too quickly. A government would like stronger international safeguards, but worries other nations may gain an advantage. An employee hopes AI will be used to support their work, but suspects the incentives might push organisations elsewhere. These are classic escalation dynamics. They do not require bad intentions. They require only a system that rewards acceleration more effectively than caution.

That is why the idea of Moloch feels so relevant to artificial intelligence. It provides a way of understanding how reasonable people can collectively create outcomes that very few of them would have chosen individually.

The challenge, as I see it, is not stopping AI. That would be neither realistic nor desirable. The challenge is ensuring that AI remains a tool that serves human purposes rather than becoming something that quietly dictates them. Ultimately, the most important question isn't whether we can build more powerful artificial intelligence. It's whether we can avoid creating a world in which everybody wins the race and discovers, a little too late, that something valuable was sacrificed to Moloch along the way.