What the Romans Got Right About AI

Two thousand years before ChatGPT, the Romans solved the same problem we’re getting wrong. They understood something about systems, humans, and the relationship between the two that the AI industry is only now beginning to grasp.

The Roman Empire ran on infrastructure.

Not legions. Not emperors. Infrastructure. Roads that connected every corner of the empire. Aqueducts that delivered water to cities of a million people. Legal systems that resolved disputes across dozens of cultures. Communication networks that moved information faster than any civilisation before them.

These were not engineering achievements. They were systems achievements. The Romans didn’t just build things. They built things that worked with human behaviour rather than against it. And they maintained them for centuries.

The AI industry is building the most powerful systems in human history. It is also building them with a fundamental misunderstanding that the Romans would have recognised immediately.

The Roman Principle: Infrastructure Serves People, Not the Other Way Around

The Roman road network covered over 400,000 kilometres at its peak. Every road was built to the same standard: raised, drained, surfaced, and marked with distance markers. A Roman citizen could travel from Britain to Mesopotamia on a continuous, maintained network of roads that worked the same way everywhere.

But here’s the detail that matters: the roads were not built for the Roman state. They were built for the people who used them. Trade moved on Roman roads. Information moved on Roman roads. Soldiers moved on Roman roads, but so did merchants, diplomats, travellers, and settlers. The infrastructure was designed to be useful to anyone who needed it, not just to the system that created it.

The AI industry has inverted this principle. The most powerful AI systems are being built to serve the companies that create them, not the people who use them. The product metric is engagement, not utility. The design goal is cognitive offloading, not cognitive enhancement. The business model is dependency, not capability.

A Roman road made you more capable. You could travel further, trade more, communicate faster. The road amplified your agency. It didn’t replace it.

An AI assistant that writes your emails, summarises your research, makes your decisions, and navigates the digital world on your behalf is not amplifying your agency. It is replacing it. The road is walking for you.

The Aqueduct Principle: Deliver the Resource, Don’t Replace the User

The Roman aqueduct system is one of the most impressive engineering achievements in history. Eleven aqueducts delivered over a million cubic metres of water to Rome every day. The water flowed by gravity, over distances of up to 90 kilometres, through precisely calculated gradients that maintained consistent pressure throughout the system.

But the aqueduct didn’t drink the water.

This sounds absurd, but it’s the precise distinction the AI industry is collapsing. The Roman system delivered a resource to the people who used it. The users decided what to do with the water. They drank it, bathed in it, used it for manufacturing, for agriculture, for sanitation. The aqueduct was infrastructure. The application was human.

AI is both infrastructure and application. It doesn’t just deliver information; it processes it, synthesises it, evaluates it, and presents a conclusion. The user receives not a resource but a finished product. The cognitive work that the user would have done with the resource, the analysis, the synthesis, the judgement, has already been done.

This is the difference between a water system and a robot that drinks for you. The first makes you more capable. The second makes you dependent on a capability you no longer possess.

The Legal Principle: Systems Need Rules That Humans Can Understand

Roman law was revolutionary not because it was sophisticated but because it was accessible. The Twelve Tables, Rome’s first legal code, were displayed publicly so that every citizen could read them. The principle was simple: a system that governs human behaviour must be understandable by the humans it governs.

This principle held for centuries. Roman law evolved, expanded, and became more complex, but it maintained the core requirement that the rules be legible. You could challenge a legal decision. You could appeal. You could argue your case. The system had transparency, not because transparency was a nice-to-have, but because a legal system that operates as a black box doesn’t function as law. It functions as arbitrary power.

The AI industry is building systems that operate as black boxes. The user asks a question. The system produces an answer. The process by which the answer was generated is opaque. The sources are aggregated beyond recognition. The reasoning is embedded in billions of parameters that no human, including the engineers who built the system, can fully explain.

When a Roman citizen was subject to a legal decision, they could ask why. They could examine the reasoning. They could challenge the conclusion. The system’s legitimacy depended on its transparency.

When an AI produces a summary, a recommendation, or a decision, the user cannot ask why. They can accept or reject the output, but they cannot scrutinise the process. The system’s authority depends on its fluency, not its reasoning.

The Romans understood that a system you can’t interrogate is a system you can’t trust. The AI industry is betting that fluency is a substitute for trust. History suggests otherwise.

The Military Principle: Standardisation Enables Scale, but Diversity Ensures Survival

The Roman military was the most effective fighting force in the ancient world, and its effectiveness was built on standardisation. Every legionary carried the same equipment. Every cohort followed the same tactical playbook. Every fort was built to the same plan. A legion raised in Syria could deploy to Britain and operate with the same efficiency because the system was standardised.

But the Romans also understood the limits of standardisation. When they encountered terrain, enemies, or cultures that their standard playbook couldn’t handle, they adapted. They incorporated local knowledge. They used auxiliary units composed of local soldiers who understood the terrain. They modified their tactics for forests, deserts, mountains, and urban environments.

The standardisation was the backbone. The adaptation was the survival mechanism.

The AI industry is building standardised systems and deploying them universally. The same model processes legal contracts and writes poetry. The same system summarises medical research and generates marketing copy. The same architecture handles a question from a CEO and a question from a student.

This works because the underlying capability is genuinely impressive. But it fails in the same way that a Roman legion would fail if it tried to fight in every terrain with the same tactics. The standardisation that enables scale also creates brittleness. The system that works well on average works poorly at the edges, and the edges are where the important decisions happen.

The Romans solved this with local adaptation. The AI industry is attempting to solve it with bigger models, more data, and more compute. The Roman approach was cheaper, more effective, and more sustainable.

The Communication Principle: Speed Without Verification Is Dangerous

The Roman communication system, the cursus publicus, moved information across the empire at unprecedented speed. Mounted couriers could cover 80 kilometres per day. Relay stations every 15-20 kilometres maintained fresh horses and supplies. A message could travel from Rome to the frontier in days rather than weeks.

But the system had a built-in verification mechanism. Messages required official seals. The authenticity of a message could be checked at any relay station. The system was designed to move information quickly and reliably, not just quickly.

The AI industry has optimised for speed without building equivalent verification mechanisms. An AI can synthesise information from dozens of sources in seconds. It can produce a coherent, well-structured analysis faster than any human. But the verification, the checking of sources, the evaluation of evidence, the stress-testing of conclusions, is either absent or delegated back to the user, who is precisely the person the AI was supposed to help.

The Roman cursus publicus was valuable because you could trust the message. An AI-generated analysis is valuable only if you can trust the analysis, and the system is not designed to help you do that. It is designed to make the output so fluent that you don’t feel the need to check.

This is the most dangerous pattern in the current AI landscape. Speed without verification is not efficiency. It is the systematic distribution of unchecked conclusions at a scale that makes individual verification impossible.

The Succession Principle: Systems That Replace Humans Don’t Last

The Roman Empire lasted, in various forms, for over a thousand years. Its longevity was not due to the brilliance of any individual emperor or general. It was due to the quality of its systems. When an emperor died, the system continued. When a general was replaced, the legions functioned. When a governor was recalled, the provinces administered themselves.

The systems were designed to survive the loss of any individual component, including the most important one. This is what made the empire resilient. It didn’t depend on any single person’s capability. It depended on the system’s capability to function regardless of who was in charge.

The AI industry is building systems that replace human capability rather than systematising it. When the AI is working, the human doesn’t need to think. When the AI fails, the human doesn’t know how to think. The system creates dependency, not resilience.

A Roman general who was incapacitated could be replaced because the system preserved the knowledge, the procedures, and the decision-making frameworks that any competent successor could use. The system didn’t replace the general. It ensured the general was replaceable.

An AI that replaces a human’s analytical capability doesn’t make the human replaceable. It makes the human dependent. When the AI fails, there is no fallback. The knowledge, the procedures, and the decision-making frameworks that the human would have developed through practice have been outsourced to a system that the human doesn’t control and can’t replicate.

The Romans built systems that preserved human capability. The AI industry is building systems that erode it.

What the Romans Would Do Differently

If a Roman systems architect were designing AI today, they would make five changes:

1. Build infrastructure, not dependency. The AI should make the user more capable, not less. It should deliver resources (information, analysis, options) that the user processes, not finished conclusions that the user accepts.

2. Maintain transparency. The user should be able to interrogate the system’s reasoning. Not the full technical architecture, but the logical chain: what sources were used, what assumptions were made, what alternatives were considered.

3. Design for local adaptation. A universal system that works well on average is less valuable than a system that works excellently in specific contexts. The Roman approach was standardised backbone plus local adaptation. AI should follow the same principle.

4. Build verification into speed. The cursus publicus moved fast because it was trustworthy. AI should move fast and help the user verify, not move fast and hope the user doesn’t notice the gaps.

5. Preserve human capability. The system should ensure that the human can function without it. This means using AI to augment analytical processes, not replace them. It means maintaining the human’s ability to think, judge, and decide, even when the AI can do those things faster.

The Question the AI Industry Isn’t Asking

The Romans built systems that lasted a thousand years because they understood a principle that the AI industry is ignoring: the purpose of a system is to make the human more capable, not to make the human unnecessary.

Every Roman infrastructure project, roads, aqueducts, law, military, communication, was designed to amplify human agency. The citizen could travel further, trade more, resolve disputes, defend themselves, communicate faster. The system made them better at being human.

The AI industry is building systems that make humans better at producing outputs. The distinction matters. Being better at producing outputs is not the same as being more capable. A person who uses AI to write their emails, summarise their research, and make their decisions is producing outputs faster. They are not becoming more capable. They are becoming more dependent on a capability they don’t possess.

The Romans would have recognised this pattern immediately. It is the pattern of a system that serves itself rather than the people it was built for. And systems that serve themselves rather than their users don’t last.

The Roman Empire lasted a thousand years. The question for the AI industry is whether the systems it’s building will last a decade.

David Chadderton spent his twenties and thirties teaching people how to make life-or-death decisions at forty thousand feet. He now applies the same principles to consumer psychology, which, depending on the brief, can feel equally high-stakes. He’s the creator of the STAR Framework and the author of The STAR Framework: Rewriting the Rules of Consumer Engagement (NYC Big Book Award 2025), The STAR Operating System, and Dear Algorithm, It’s Not Me, It’s You. By day, a Chief Marketing Officer. By night, a behavioural science obsessive who writes The Unoptimised Human because he can’t stop thinking about why people do what they do.