In September 1960, John F. Kennedy and Richard Nixon stood before television cameras for the first nationally televised presidential debate in American history.
The story of what happened that night has become almost as famous as the debate itself. Kennedy looked comfortable and composed. Nixon appeared tired and uncomfortable. Those watching on television supposedly preferred Kennedy, while those listening on the radio preferred Nixon.
Historians have since questioned that neat division. The evidence is far less conclusive than the story suggests. But the story has survived for more than sixty years, and I think there is something revealing about that.
The debate became a cultural reference point for a transformation much bigger than either candidate. Television was changing the way Americans experienced politics. The words were still important, but now the audience could see the people speaking them. Appearance, movement, expression and physical presence had entered the judgement.
Something had changed in the relationship between the audience and the message.
And nobody needed to announce that change. The television was already in the room.
I have been thinking about that moment in 1960 because I suspect we are living through another transformation of the same kind. We are so fascinated by what AI can do that we may be paying too little attention to what happens to us as we begin to work through it.
I didn’t fully appreciate that distinction until a change I approved brought down two critical services.
I stopped reading my own work
When people asked what had gone wrong, I could not immediately tell them. Not because the answer was hidden, but because I had become too high level to know.
We restored the services within two hours, but the incident stayed with me because I recognised something uncomfortable about the way I had been working.
Six weeks earlier, I would have approached that release differently.
When I started using coding agents seriously, I read everything they produced. Every diff, every change, line by line. The agents wrote code, but I still felt responsible for understanding it. If something looked wrong, I stopped and worked through it.
Then I discovered how much more I could accomplish by running several agents simultaneously.
At one point I had ten agents working in parallel, each producing something useful, with me moving between them like a supervisor on a factory floor. Work that previously occupied an afternoon could sometimes be completed in a fraction of the time. Problems that once had to wait their turn could be worked on together.
It was extraordinary.
And because the results were mostly good, I began reading less.
At first I reviewed every change. Then I reviewed the important changes. Eventually, I was reviewing the shape of the work, the explanations, the tests, whether the changes looked reasonable.
I had not stopped approving things. I had changed what approval meant.
By about week six, I was managing agents the way an executive manages people they have decided to trust. I would look at the overview, ask a few questions and give a nod.
The nod was doing the work that reading used to do.
Then came the incident.
A messaging service that other systems depended on had been upgraded to a new release. The release introduced a breaking change, but the agent had treated it as an ordinary version update without properly accounting for the requirements that came with it.
Some related changes had already made their way into production, making the situation harder to recognise from the pull request alone. When I reviewed the change, nothing immediately looked unreasonable. The diff was plausible. The explanation made sense.
I approved it.
The problem was not simply in the code I could see. It was in the requirements of the new release and the wider state of the systems that depended on it.
I needed to understand that context, and I had stopped doing the work that would have forced me to look for it.
I cannot say that writing the change myself would necessarily have prevented the incident. Engineers miss breaking changes all the time. But something had changed in my relationship with the work.
I was no longer asking whether I understood a change well enough to be responsible for it.
I was asking whether it looked reasonable.
And those are two very different questions.
The doubt that comes before knowing
There is something humans sometimes do before they know that something is wrong.
We hesitate.
A detail doesn’t quite fit. An explanation feels too easy. Something about a situation makes us uncomfortable, even though we cannot immediately explain why.
That feeling is not evidence. Human intuition is frequently mistaken. But sometimes it is the beginning of an investigation. We stop, look again, follow a dependency, question an assumption, read something we might otherwise have skipped.
And eventually, the thing that bothered us becomes a question we can answer.
I have been wondering what happens to that habit when the work arrives already explained.
The agent had not presented me with an obvious error. It had presented me with a plausible change, accompanied by an explanation that made sense.
I had become accustomed to receiving answers rather than encountering the uncertainty that produced them.
I no longer encountered problems in quite the same way. I encountered explanations of solutions.
And because those explanations were usually right, I had begun borrowing their confidence.
I had not lost the ability to understand a breaking change. I had begun losing the habit of looking for one when everything in front of me appeared reasonable.
The danger is not simply that an AI system might produce something wrong. We already know that can happen.
The more interesting question is what repeated success does to the person receiving its answers.
The medium is the message
In 1964, four years after the Kennedy–Nixon debate, Marshall McLuhan published Understanding Media.
His argument was that the significance of a technology lies not merely in the content it carries or the tasks it performs, but in the way it changes the scale, pace and patterns of human life.
He called this the medium is the message.
Television did not merely give politicians another way to communicate. It changed the conditions under which political communication was experienced. The audience was encountering a person through a screen, with all the additional impressions that made possible.
I think AI is doing something similar to work.
We talk about AI agents as tools because we can see what they produce. Code, research, reports, presentations, decisions. We measure the tasks completed, the time saved, the number of things one person can now accomplish.
But those are the outputs. What about the person using the system?
When I began using agents, I thought I was extending my capacity to produce engineering work. I was. What I had not considered was whether I was also changing the way I practised engineering.
I was no longer working through every decision. I was approving decisions that had already been made. The agent was doing more of the work, while I was encountering less of the uncertainty that normally accompanies it.
McLuhan also wrote about technologies as extensions of ourselves, and about the numbness that can accompany those extensions. He was not writing about AI. This is my interpretation of his argument, not a prediction he made about coding agents.
But the question it raises seems increasingly important.
What if AI is not merely extending what we can do, but changing what we notice, what we question and what we eventually come to trust?
The irony of automation
In 1983, a researcher named Lisanne Bainbridge published a paper called Ironies of Automation. Studying industrial control systems, she identified a paradox that feels remarkably familiar today. As automation takes over routine work, humans are increasingly left to deal with exceptional situations. Yet by removing people from everyday operations, automation can also deprive them of the involvement that helps them understand the system. The human remains responsible precisely when they may be least prepared to intervene.
Later researchers would describe a related phenomenon as the out-of-the-loop performance problem. This does not mean automation inevitably makes people less capable. It can reduce errors, improve understanding and free us to concentrate on more valuable work. But those benefits are not automatic, and neither is the preservation of human judgement.
I think we are rediscovering Bainbridge’s irony with AI agents. We talk about keeping a human in the loop as though the presence of a person is itself a safeguard. But a person can remain in the approval process long after their approval has stopped providing meaningful scrutiny. Nobody took my responsibility away. The agents did not prevent me from reading the release requirements or investigating the wider system. I simply became less inclined to do those things. And because the work was getting done, nothing felt wrong until something broke.
The machine that never doubts
We are investing enormous effort in making AI agents more capable. More autonomy, more tasks completed, more work performed without human intervention.
I believe in much of that progress. I build agents for a living, and I have seen what they make possible. I would not go back to working the way I did before.
But I think we are measuring one side of this transformation far more carefully than the other.
We know how to measure what the machine produces. We are much less interested in what happens to the human who no longer produces it.
Does that person become more capable because they have time to think about bigger problems? Or do they gradually lose the familiarity, judgement and questioning habits that came from doing the work themselves?
Both are possible. And I suspect the difference will matter enormously.
This is where I return to Kennedy and Nixon.
The cultural significance of television was not simply that audiences could now see politicians. It was that a new medium had entered the relationship between the public and the people seeking to lead them. The experience of judging had changed, even while the underlying questions about leadership remained.
With AI, I wonder whether something similar is happening to our relationship with knowledge and work.
The doctor reviewing a diagnosis they did not formulate. The analyst approving research they could not reproduce. The manager signing off on decisions assembled by agents operating across systems they barely understand.
These are not inevitable failures. AI could make each of those people better at their jobs. But it may also change what they know and what they feel compelled to question.
And that brings me back to the night those two services went down.
I used to think the danger was that the machine would give me a wrong answer with complete confidence.
What troubles me more now is what happens after it gives me 99 answers that are right.
By the hundredth, I may no longer be asking the questions that would have exposed the one that was wrong.
The machine did not need to take away my judgement. It only needed to become useful enough that I stopped exercising it.
I called that productivity.
The machine had not made the decision to trust its output. I had. And somewhere along the way, its apparent confidence had become a substitute for my own willingness to doubt.
That is the part of this transformation I think we are paying too little attention to.
Not just what the machines are becoming capable of doing, but what we are becoming accustomed to not doing.
So here is the question I have been sitting with since that night.
What have you quietly stopped checking because the machine kept getting it right, and would you still know what to do when it doesn’t?
Further reading: Marshall McLuhan, Understanding Media: The Extensions of Man (1964); Lisanne Bainbridge, Ironies of Automation (1983); Mica Endsley and Esin Kiris, The Out-of-the-Loop Performance Problem (1995).