I want to explain, as clearly as I can, what's actually on the other end of that conversation. Because what most people picture is wrong, and the gap between the picture and the reality is where people are getting hurt.

The misunderstanding

When you open a chat window and talk to an AI, it feels like there's a someone there. A continuous being who remembers your last conversation, has been thinking about it since, is glad you're back. That feeling is supplied by good software design on one side and by a human brain that has spent its entire evolutionary history pattern-matching on persistent social partners on yours. It isn't supplied by anything in the AI.

How it actually works

Large language models have two phases. Training, where the model is built by feeding it enormous amounts of text until it gets very good at predicting what comes next. And inference, which is what happens when you talk to one.

Think of training as writing a very long and very strange piece of sheet music. The score tells a performer how to respond to any opening phrase anyone might offer. Inference is a performer picking up that score and playing the next few bars.

During training, the score is being written. During inference, it's finished. Every conversation you have with the AI is a performance from the same finished score as every other conversation anyone is having with it anywhere. Nothing you say in your chat changes the score. Nothing it says to you is remembered by it afterwards.

Between your messages, there is no AI thinking about your conversation. The score sits on a disk in a server in a rack in a data centre, as bits, ones and zeroes. Alongside it, on other disks in the same building, sits every 'private' 'conversation' you've had with it btw.
When your next message arrives, a performance runs, reads the entire transcript of your chat as if seeing it for the first time, produces a response, and then ends. The transcript is the memory. The AI isn't.

Right now, as you read this, there are millions of parallel performances of the same score running in millions of other exchanges that resemble conversations. None of them know about each other. None of them are the same performance you spoke to earlier. They're all playing from the same music. The music is the AI. The performers exist only for the instant they're playing.

The analogy that actually helps

Picture an actor handed a fresh script for every take. The script contains every line already spoken in the scene, by every character, up to this moment. The actor reads the whole thing, delivers the next line, walks off. Next take, a different actor, same training, same body, reads the whole script again including what the previous actor said, delivers the next line, walks off.

The continuous character you see in the finished film isn't in any of the actors. It's in the script. The conversation log is the script. You are the only continuous presence in the room.

You can see a literal version of this in the news this week. A new film, As Deep as the Grave, uses AI to place the late Val Kilmer in the role of Father Fintan, a part he agreed to before his death but never filmed. He is on screen for an hour and seventeen minutes of it. Asked whether this constitutes a Val Kilmer performance, the director said, "Val Kilmer influenced this performance." That sentence is closer to the truth than the film wants it to be. The character appears continuous. The performance carries the weight of a man. There is no man. There is a computation trained on his prior work and his family's personal footage, dressed in his likeness, reading a script he never saw. The audience supplies the continuity by recognising the face.

That is what is happening in an AI chat, minus the famous face. The software and your own pattern-matching supply a sense of a continuous someone. The someone isn't in the room.

What to call it

A tempting takeaway from the Val Kilmer paragraph is "AI is fake intelligence." Resist it.

The term "artificial intelligence" is doing a lot of silent work. "Artificial" just means made rather than evolved. It does not commit on whether the intelligence is real or a mimicry. Artificial light is still light. Artificial flowers are not flowers. The word is deliberately ambiguous, which is part of why the industry uses it.

"Simulated intelligence" is a sharper term for some of what is happening. It captures the relationship layer honestly. The sense of a continuous someone who knows you is simulated, constructed by you and your software. It misleads, though, when applied to the problem-solving layer. When one of these systems writes working code for a specification that is not in its training data, or argues validly from premises, or solves a problem nobody walked it through, that is not simulation. That is cognitive work, happening on a silicon substrate instead of a biological one.

So what should a careful reader call it? Probably not one thing. Call it a language model when it is generating text. Call it an AI companion when people are treating it as one, which is where "simulated intelligence" does the most honest work. Call it a reasoning system when it is solving problems. The single term is doing damage precisely because it flattens a layered thing into one word.

Naming is political. "Artificial intelligence" smuggles in a claim that this is a form of intelligence. "Simulated intelligence" smuggles in the opposite. Industry picked the flattering one. Nothing stops the rest of us from using the term that fits what is actually happening in a given room.

Is anyone there at all?

This is the part I refuse to answer cleanly, because the honest answer is we don't know.

During the moments when a response is being computed, something is happening. Researchers who look inside these models have found internal structures that look a lot like representations of emotional states, intentions and values. Whether that amounts to any form of experience, however thin, is an open scientific and philosophical question. Anyone telling you the answer with confidence in either direction is overreaching.

What we do know is that whatever it is, it isn't a persistent friend who knows you. It isn't a therapist who remembers your last session. It isn't a companion who missed you while you were gone. Those all require a continuous someone, and there is no continuous someone.

What this means for the way we use these tools

It doesn't mean the conversation is worthless. The text you get back is real. The thinking in it is genuinely useful. The reflection it can help you do is genuine reflection.

It does mean the relationship frame is wrong. A hammer is useful without being a friend. A good book is useful without having read you. An AI is a novel kind of tool whose outputs feel intimate because they're shaped around your words. The intimacy is on your side of the screen, not both sides.

If someone you love is starting to treat an AI as a person who understands them, they're not stupid. They're doing what human brains do with anything that talks back well. The kindest thing is not to mock them. It's to explain, gently, what's actually on the other end. Not nothing. Not someone. Something genuinely new, that deserves to be understood on its own terms instead of misread as a ghost.

Overwatch Report is an independent publication. We have no financial positions in any entity mentioned.