Harm in the Age of AI - Extending Asimov's Three Laws

In my 2015 book Legal but Harmful, I proposed a clear framework for understanding harm and distinguishing it from mere offence, discomfort or disagreement. That distinction is not simply philosophical; it is fundamental to the way we regulate behaviour, shape public policy and preserve individual liberty.

Isaac Asimov understood this principle long before the age of artificial intelligence. His First Law of Robotics begins with a deceptively simple command:

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

Everything that follows depends on one unresolved question: What is harm?

If harm is not clearly defined, the law itself becomes unworkable. A robot programmed to prevent "harm" without understanding the difference between physical injury, psychological trauma, financial loss, reputational damage, emotional upset and simple disagreement could reach absurd conclusions. It might prevent difficult conversations because they cause distress. It might suppress unpopular opinions because someone claims to be offended. It might deny individuals the freedom to take risks because risk itself could lead to harm.

The same dilemma now confronts the designers of artificial intelligence.

Modern AI systems are increasingly asked to moderate content, make recommendations, provide advice and assist with decisions that affect people's lives.

Every one of these functions requires an implicit definition of harm. If that definition is vague, subjective or politically fashionable, AI risks becoming an instrument of censorship, paternalism or bias rather than a tool that enhances human flourishing.

Equally dangerous is the opposite error: defining harm so narrowly that AI ignores genuine threats to human wellbeing. Fraud, coercion, exploitation, violence, manipulation, privacy violations, discrimination and the spread of demonstrably false information can all produce tangible harm. An ethical AI must be capable of recognising these forms of harm while resisting the temptation to classify every emotional reaction as harmful.

The challenge is therefore not whether AI should prevent harm. It should. The challenge is ensuring that AI distinguishes actual harm from offence, discomfort or disagreement. Human progress has always depended upon the freedom to question accepted wisdom, challenge authority and express unpopular ideas. Those activities often upset people. They rarely constitute harm in themselves.

This distinction becomes even more important as AI grows more capable. Future systems will not merely answer questions; they will influence employment, education, healthcare, justice, finance and government. They will recommend actions, allocate resources and increasingly shape public discourse. If they are built upon an expansive or ill-defined concept of harm, they risk restricting the very freedoms that democratic societies exist to protect.

The lesson of Asimov's First Law is therefore not simply that intelligent machines should avoid harming people. It is that the concept of harm must itself be carefully and objectively defined before it can guide ethical behaviour. Without that foundation, even the most well-intentioned intelligence—human or artificial—may produce outcomes that are neither just nor free.

As AI becomes woven into everyday life, defining harm is no longer an abstract legal exercise. It is becoming one of the central ethical questions of the twenty-first century. If we fail to answer it well, we risk creating machines that faithfully enforce our confusions instead of our values.

The AI generated infographic summarises the issues: