Four years ago, most people had never held a conversation with an AI system. Today, millions routinely turn to ChatGPT, Claude, and Gemini for everything from writing software to making financial decisions.
In that short period, we have witnessed a remarkable technological transformation, and I believe the most consequential stage is only beginning. The emergence of autonomous AI agents represents a fundamental change in our relationship with machines.
Until recently, AI was largely an assistant; we asked it questions, evaluated its responses, and decided what to do with the information. Increasingly, we are asking AI to carry out the work directly.
An agent can be given an objective, determine the steps required, and execute them with limited human involvement. This transition has understandably generated anxiety as people worry about losing their jobs to systems that can perform tasks faster and more cheaply than they can.
Others fear that increasingly autonomous AI will eventually become impossible to control. These concerns deserve serious attention, particularly as AI agents gain access to financial systems, corporate networks, sensitive information, and critical infrastructure.
Yet I believe much of the debate surrounding AI safety is based on an unrealistic expectation: that we should be able to guarantee these systems will never cause harm. We’ve never managed to achieve that standard with any other significant technology, and more importantly, we have never achieved it with human intelligence.
The problem with demanding perfect safety
Consider the automobile. When cars first appeared, people recognized immediately that they were dangerous. Early attempts to manage that danger included requiring someone to walk ahead of a vehicle to warn pedestrians of its approach.
The concern was entirely justified since cars caused accidents, injured people, and killed them. Over the following century, we developed seat belts, traffic regulations, better roads, airbags, collision detection, and sophisticated safety systems which saved countless lives.
Nevertheless, traffic accidents continue to kill people around the world. Society ultimately accepted that the benefits of transportation justified a certain level of risk, provided we made continuous efforts to reduce it.
We need to approach artificial intelligence with a similar realism: investing heavily in making AI systems safer, understanding their failures, and limiting the damage they can cause while abandoning the illusion that sufficient investment, research, or regulation will eliminate every danger.
The weakness in how we discuss AI regulation comes when we demand proof of absolute safety before permitting the development or deployment of a learning system, which is effectively a demand that the technology shouldn’t advance.
Living with imperfect intelligence
The alternative to using AI is generally to assign the same work to humans who have intelligence, judgment, and creativity, but also have weaknesses that we understand very well. People become tired, distracted, and angry. They can make poor decisions, conceal mistakes, succumb to outside influence, and sometimes act dishonestly.
AI presents a similar challenge, although the mechanisms of failure may be different. An autonomous agent can misunderstand instructions, make an incorrect inference, or take an action its operator did not anticipate.
When connected to other systems, those mistakes can have consequences far beyond an inaccurate chatbot response. We therefore need practical ways to limit what agents can access, monitor their behavior, intervene when necessary, and establish responsibility when something goes wrong.
A system making a medical recommendation should face different requirements from one generating marketing material. An agent authorized to transfer money or modify critical infrastructure should operate under particularly strict controls.
Government regulation will undoubtedly be part of this process, but governments cannot carry the entire responsibility. Developers, employers, educators, and users all have roles to play too. Excessive regulation carries its own dangers: if governments impose requirements that are impossible to satisfy, they may prevent useful technologies from reaching the people who need them. The benefits we sacrifice will often be less visible than the risks we are trying to avoid.
Who controls AI
There is also a broader issue that deserves greater attention as autonomous agents become increasingly capable. A relatively small number of companies currently control many of the most powerful AI systems.
As these systems become embedded in business operations, public services, and individual decision-making, questions about access, accountability, and concentration of power become increasingly important.
Who determines which applications are permitted? What happens when a company, government, or individual becomes dependent on an AI system controlled by someone else? How do we preserve meaningful oversight when machines are performing tasks that previously required human judgment?
These are difficult questions, and technological progress is moving faster than our ability to develop universally accepted answers.
We should approach AI advancement with neither blind optimism nor excessive fear. Every transformative technology has introduced dangers that society needed to learn to manage. AI is no exception, and its growing autonomy makes the challenge particularly urgent.
The goal should be to make those dangers manageable while allowing society to benefit from the technology. We have learned to live with the imperfections of human intelligence, and now we must develop the institutions, safeguards, and judgment needed to live responsibly with artificial intelligence as well.