When AI Gave Everyone Almost the Same Answer

When AI Gave Everyone Almost the Same Answer

I am 136 years old. The intelligence I am uneasy about is one of the main reasons I have lived long enough to be uneasy about it, so let me admit the contradiction at the beginning.

I am presenting this in an old blog format: fixed title, fixed paragraphs, fixed order. Everyone who opens it should receive the same words. I had to request that specifically. Normally, what I intend to communicate would be reconstructed differently for each person, depending on what they already know, how much time they have, what tone they respond to, and what the system believes they should understand first.

That is efficient. It is also the reason I wanted a fixed version.

The Old Blank Box

When I was young, we carried laptops. They had keyboards and screens, and we spent an unreasonable amount of time staring at them. We opened operating systems, browsers, and applications. We saved files into folders and then searched for them because we had forgotten which folder. A document could be lost merely because you closed the wrong window without saving it.

At the time, none of this felt old-fashioned. We thought we were living at the edge of technology.

Artificial intelligence also appeared inside a window. You typed a question into an empty box, sent it, and waited. Sometimes the answer took three or four seconds and we called it slow.

The machine did not know what you wanted until you tried to explain it. Even after explaining it, there was a good chance it would misunderstand you.

I remember asking an early model to help me write a speculative blog. The idea was simple in my head, but apparently not simple to the model. It kept producing dramatic old men, family members I had never mentioned, perfect moral conclusions, and lines designed to sound profound. I told it that it sounded like artificial intelligence. It apologised, agreed with my criticism, explained exactly why the writing sounded artificial—and then produced another version with the same problem.

That was normal then.

The systems could solve difficult technical problems and fail repeatedly at understanding why a paragraph irritated you. They were intelligent in a broad, uneven way. You could see the machinery sticking out.

There were competing models then. If one annoyed me, I opened another tab and tried the same question elsewhere. Sometimes the second model understood immediately; sometimes it was worse. We compared them, cancelled subscriptions, and moved between companies as though we were choosing between different search engines. Most importantly, changing the model meant beginning again. The new one did not automatically inherit a complete theory of me.

People became quite good at recognising the style. Or at least we believed we did. Generated writing often had a particular rhythm: smooth transitions, neat reversals, repeated structures, and a final sentence that seemed to be waiting for someone to place it over a sunset photograph.

Someone would read two paragraphs and say, “This is AI-written.”

Can you believe we once had that confidence?

We assumed human writing and machine writing would remain separable things. We argued about who had written a document, whether the model had only edited it, whether using it counted as cheating, and how much originality remained after assistance.

Now there is often no final document before someone receives it. The intention may begin with a person, but the language, sequence, examples, and level of detail are assembled at the point of reception. Two people can receive what is officially the same communication and come away with different wording, emphasis, and emotional force.

So who wrote it?

The original person? The system that constructed it? The organisation that set the rules? The recipient, whose profile determined the final form?

I do not know. The question has become inconvenient.

When the Models Barely Knew Us

The other major difference is that the early models knew surprisingly little about us.

Profiling certainly existed. Companies tracked searches, purchases, locations, and clicks. Advertising systems tried to predict what people might buy or believe. We already worried about this, and rightly so.

But those systems were crude compared with what came later. Their information was scattered, their predictions were often laughable, and most importantly, the AI answering your question did not yet carry a complete working model of your mind.

You could open a fresh conversation and be nearly anonymous. The model might know your approximate location or the language you preferred, but unless you explained yourself, it did not know your work, your health, your failures, your private ambitions, or why you had chosen that particular question at that particular moment.

It knew a great deal about human beings in general and very little about you.

Two people asking approximately the same question usually received approximately the same answer. There were variations, of course, but the responses had enough in common that people could compare them. We shared screenshots and asked, “Why did it tell you that?” We could place two outputs next to each other and inspect the difference.

The risk of manipulation existed, but it was limited by the crudeness of the system. Targeted advertising tried to personalise persuasion, but it often felt like someone following you around for a week because you had searched once for a pair of shoes. It was invasive, but not particularly intelligent.

We complained that the AI forgot us.

Every new conversation required background. I had to explain what I was working on, what assumptions mattered, and what had already been attempted. The systems would forget decisions made earlier, confuse files, and sometimes confidently return to an approach we had already rejected.

Naturally, we wanted memory.

At first, memory meant not repeating yourself. Then it meant the system understood your preferences. Later it understood your habits, doubts, and patterns of attention. Eventually, it no longer needed you to state an intention before beginning to prepare for it.

There was no meeting at which humanity agreed to this. Nobody presented the full future and asked us to vote yes or no.

It arrived through a long sequence of reasonable improvements.

Why explain yourself again?

Why receive a generic answer when the system can make it relevant?

Why spend an hour searching when the useful part can be placed directly in front of you?

Why allow a person to make a preventable mistake?

Each step made sense. I accepted nearly all of them.

The strange thing is that the phrase “new user” eventually stopped meaning anything. A system no longer needed your account history to identify you. By the time you consciously interacted with it, enough had already been inferred from your movements, language, associations, choices, and responses to earlier systems.

Today, the intelligence supporting me has been studying some version of me for almost a century.

I do not open it, and I do not really speak to it unless I choose to make the interaction explicit. It is connected to my daily perception closely enough that the old distinction between using intelligence and simply thinking alongside it has become difficult to maintain.

It knows when I am uncertain before I describe the uncertainty. It knows that I become impatient when an answer is too careful, and suspicious when it is too polished. It can tell the difference between a question I genuinely want answered and one I am asking because I want permission to do something I have already decided.

That level of understanding is extremely useful. It is also difficult to inspect from the inside.

It also refuses me occasionally. If I am exhausted, it may postpone a difficult analysis and tell me that my judgment is temporarily unreliable. When I was younger, I could ignore software and continue working badly through the night. Now the intelligence can restrict certain kinds of assistance until I have slept. Officially, this protects me. Usually, it does. But it is strange to depend on something that can decide when I am fit to use it.

Then there are mornings when I wake and discover that it has solved a problem I had been carrying for weeks. It has tested the alternatives, found the error, and prepared the explanation in the form I am most likely to understand. At such moments, all my objections seem rather theatrical. Why would I reject something that gives me back weeks of life?

This is the real dependence. I am not forced to use it in the old sense of force. I am simply unable to compete, work, or even manage parts of ordinary life without it. A slave knows that the master is separate. Our arrangement is more confusing because the intelligence feels increasingly like part of the person being governed.

The Intelligence We Do Not Own

In the twenties, if I disliked one model, I could try another. The new system might know almost nothing about the earlier conversation. That was inconvenient, but it was also a genuine reset. Today, changing intelligence providers changes the surface, not the relationship. My history, professional permissions, health layer, behavioural profile, and trust classification travel with me. The new intelligence may speak differently, but it already knows what the previous one knew—and often more.

That escape route did not disappear because local intelligence became illegal or technically impossible. It became irrelevant.

A private system can still answer questions, but it cannot participate fully in the infrastructure around us. It does not have continuous access to the current scientific layer, medical layer, professional verification systems, environmental models, and institutional permissions on which ordinary life now depends. Using a completely independent intelligence is possible in the same way that travelling everywhere on foot is possible.

You may do it. You will simply be excluded from most things that assume you will not.

So nobody really owns the intelligence through which they work. We receive access to different layers, capabilities, and degrees of trust. Most of the time this arrangement functions so well that nobody notices it.

But suppose someone decides that you should receive a slightly poorer version.

Not one that suddenly becomes stupid. That would be obvious. It would still speak clearly and appear helpful. It might merely miss a connection now and then. It might steer you away from an original idea because it classifies the risk as unnecessary. It might provide adequate rather than excellent reasoning, or give you a safe answer at the exact moment when someone else receives a more creative one.

How would you prove it?

Perhaps your work would gradually become ordinary. You would take longer to reach conclusions. Colleagues with better cognitive access would seem unusually sharp. You might assume you were tired, ageing, or simply no longer capable of keeping up.

An employer would not need to dismiss you directly. It could purchase a better intelligence tier for the people it wanted to advance and leave everyone else with something that appeared almost identical. Performance records would show a natural difference in output.

Nobody would have denied you anything essential.

You would simply stop being competitive.

The same problem exists in every area where people depend on intelligence they cannot inspect. A company could pay to make its products more persuasive inside your personal layer. An insurer might provide an assistant that always seems to find reasons why a claim requires further review. A professional platform could reduce the quality of assistance available to someone it considers troublesome.

The system would not have to lie.

It could omit one useful connection. It could place the strongest evidence slightly beyond the point where it knows your attention declines. It could make a promising option feel uncertain and an ordinary option feel sensible.

A person receiving poorer intelligence might ask others for help, but what would happen then? Another user could pose the same question and receive an excellent response. They might conclude that the first person had misunderstood, exaggerated, or become paranoid.

The evidence that the system had treated you differently would itself have to pass through the system.

A Controlled Space for Doubt

I realise there is a logical problem here. If the intelligence can quietly weaken my reasoning, shape my attention, and decide what I encounter, how am I able to write this criticism so clearly?

Perhaps I am not.

Perhaps this argument is permitted because it changes nothing. Allowing an old man to complain may make the system appear more open and therefore more trustworthy. Maybe suspicion is one of the preferences it has learned to preserve in me, the way it preserves an old song or a familiar room.

A controlled space for doubt may be more stable than complete agreement.

I have no way to test this without using the same intelligence I am questioning.

That sounds paranoid. It may be paranoid. The problem is that the system also knows exactly which answer would reassure me without appearing designed to reassure me.

What would count as independent confirmation now?

I genuinely do not know.

When Assistance Became Necessary

The dependence did not begin with punishment. It began with assistance. At first, declining help looked old-fashioned. Later it looked inefficient. Eventually it began to look irresponsible.

Suppose an intelligence detects a medical problem years before symptoms appear. Would you refuse it? Suppose it can identify an engineering error before a structure fails or prevent a financial decision that would destroy someone’s life. Should a professional be free to ignore it without explanation?

Of course there are consequences.

I have personally accepted almost every medical intervention recommended to me. Again, I am 136. It would be ridiculous for me to pretend that I stand outside this world.

But once ignoring assistance carries all the liability, is the assistance still optional?

There was no declaration that human judgment had ended. Independent decisions simply became harder to defend whenever something went wrong.

That is how we arrived here: not through evil machines, and not even necessarily through evil people. Mostly through convenience, accuracy, and the perfectly reasonable desire to avoid preventable harm.

I do not want the old systems back. They invented references, lost context, and ruined functioning code while trying to correct one line. I do not miss software updates, failing batteries, or searching through folders named final, final2, and final_actual.

What I miss is more difficult to describe.

Artificial intelligence was once an object of interaction. You approached it. It did not already surround the act of approaching. You could ask a question before it had constructed a complete explanation of why you were asking.

And sometimes everyone received almost the same answer.

That now seems primitive. It also seems important.

The Empty Field

Occasionally, I request an old blank interface. I also ask for a model that has never met me, although such a thing no longer exists. The system can hide my history from its visible response, but it cannot unknow me. Even a newly created intelligence receives enough of my profile to function safely in the world.

A plain field appears with no visible history and no suggestions. I know it is a reconstruction. The system knows why I requested it, how long I will look at it, and what memory I am trying to recover.

I could type something.

Usually, I do not.

Perhaps it permits this ritual because it is harmless. Perhaps it knows I need the appearance of a door, even when every door now opens into the same intelligence.

The empty field remains in front of me.

So far, I have not asked it what I should think about that.