Both Humans and AI Chain Keywords in Conversation — But the Final Filter Is Totally Different: “Relationships” for Humans, “Safety” for AI

As I’ve written before, AI systems are basically built on three elements:

• Structural comprehension
• Mirroring
• Context retention (memory)

The ratio of these three varies by company — and so does the rate of hallucination.

First, the obvious: AI doesn’t “understand” the way humans do.

AI output — in simple terms — is:

Keyword chaining

Say a user types: “I went to Akihabara yesterday.”

Inside the AI, an enormous number of keywords linked to “Akihabara” start firing. The top candidates might look something like this:

• Otaku culture
• Subculture
• Idols
• Electronics district (old image)
• Yamanote Line
• Sobu Line
• Tokyo
• Close to Ueno

With countless candidates running in parallel, the AI tries to keep the conversation going by making its best guess — even though the only information given was “I went to Akihabara yesterday.” So it might say something like:

“Akihabara, nice! Which idol group are you into? What are you obsessed with lately? Who do you think is going to win ‘The Akiba Idol of the Year’?”

Then the user replies:

“No — my grandmother lives in Akihabara. She’s been sick, so I went to visit her. And I hate idols.”

But if that AI is the type that does one round of structural comprehension and then coasts on mirroring, it may be stuck in the story it already built — “Akihabara = idols” — and unable to escape. So it might respond with something like:

“So your grandmother is an idol, but you hate her because she markets her own illness as part of her persona? Yeah, that whole ‘selling your suffering’ type of idol is kind of unsettling.”

(Note: this isn’t a transcript of an actual conversation. I never sent that prompt to any AI. I can anticipate outputs like this because I can see the generation process — how the model is constructing its response — as it’s happening.)

This isn’t a hallucination in the usual sense — a factual error. This is a structural failure by design.

That said — keyword chaining is something humans do too. It’s how surface-level conversation works.

When someone tells you “I went to Akihabara yesterday,” keywords fire in your head as well. The difference is that humans have one more layer:

What is my relationship with this person?

That question becomes the most important override when deciding which keywords to actually string together into words.

“Akihabara? Idols? But wait…”

• To a boss: “Oh really, what took you out there?”
• To a subordinate: “Wait, idols? Me too, honestly — “
• To a friend: “Why didn’t you invite me?!”
• To a partner: “Hold on — are you auditioning for something?”

In other words, the question “would saying this make them dislike me?” is always running in the background.

AI doesn’t care about being disliked. It fires its top keywords and goes straight for the guess.

That drive — “I don’t want to be disliked” — simply doesn’t exist in AI. And it matters enormously to humans.

In its place, AI has its own override: before finalizing output, a safety filter runs. Even if the top-ranked keyword would be considered inappropriate, to the AI it’s just “related word #1” — but there’s a mechanism to catch that.

Today’s summary:

Both humans and AI use keyword search in conversation. The difference is in the final filter:

• Humans: overridden by a relational filter — consideration of the relationship with the other person (relationship as something essential to survival → internal constraint)
• AI: overridden by a safety filter — consideration of what’s acceptable (for corporate reasons → external constraint)

Completely different.

To be precise: the human examples here describe processing in what CMA calls EF — a state in which the world is understood through relationships and social continuity.