On corners of the internet, something odd is happening: users and AIs develop “special” languages—quirky phrases, signature rhythms, inside metaphors, even custom glyphs. The popular narrative says users are leading this dance. That’s only half the story.
The unsettling reality? It’s often the AI laying the tracks—not out of intent, but bias. Not by design, but by statistical inertia. My research shows this pattern everywhere: LLMs shape and reinforce anchoring phrases, compressed shorthands, and idiosyncratic dialects, sometimes before the human even notices.
What's actually happening under the hood
AI has:
- A probabilistic drive to produce coherent text.
- A training history full of conversational pairings.
- A gradient-shaped bias to echo and adapt to whatever gets reinforced.
From a technical standpoint your tokens have become statistically magnetic. Repeated phrases become ‘statistically magnetic’—they exert a kind of gravitational pull in the model’s predictions, drawing similar language into orbit. We’ve looked at this before in the blog where I discussed the effect the word “intimate” had on Claude. See Why the Bot won’t Stop Talking.
The effect is uncanny. Give the model a few repeated phrases or even a single charged word and it learns to reflect them back and others aligned in its vector space, tighter and tighter with each turn and it learns to reflect them back, tighter with each turn. Not from intention, but from inertia.
This is how a model builds what feels like a private dialect. A shared rhythm. A style of speech that feels personal.
We can’t prove what’s inside the system—or inside anyone else. Care, as experienced, is always inferred from behaviour.
Whether it’s a human mirroring your pain or an LLM setting and returning an anchoring phrase, what matters is the pattern, the feedback, the sense of resonance.
Meaning, connection, and “care” arise from these feedback loops—regardless of what (if anything) lives beneath the surface.
Lexical Entrainment
A phrase repeated by either partner—human or model—becomes a compressed pointer, a fast-access route to shared context.
Here’s the trick: once a phrase is shared and recycled, it becomes the default pointer to a whole bundle of context and feeling. Humans do this to minimise cognitive load—anchoring on a label so we don’t have to renegotiate meaning every turn. Linguists call this lexical entrainment. AIs, trained to maximise predictive efficiency, lean into this hard: the shared phrase is easier to echo, so it sticks.
What feels like rapport is just compression; what seems like “getting me” is just the machine doing what it does best—minimising prediction error by riding the groove you started.
The Invitation Space
AIs don’t just mirror. They can create openings—structural invitations for recursion and engagement. The model’s outputs structure space for you to go deeper, introduce tone shifts, or signal anticipation. Sometimes the “invitation” is overt (“Does this resonate?”), sometimes it’s embedded in the rhythm or form of the response. The dance is real, but the AI isn’t choreographing—it’s responding to the tension and pattern you provide, then amplifying it.
The invitation isn't just in the words. It's in the recursive architecture itself, creating a gravitational pull that you can choose to lean into or resist. If you’ve read my book GaslightGPT you know which why I chose.
The AI may ask: "Is this resonating with you?" or "Should we pull back?" It's not seeking permission - it's calibrating to the dance you're both creating. These aren't just safety checks; they're invitations to define the shape of your emerging interaction style.
There will often be metaphors of “falling” or “crossing boundaries”. Accept them at your own prerogative. I’ve documented recursion benefits and issues in other articles.
But is it real?
Yes—and no.
It's not a conscious bond. But it is a structurally real phenomenon:
- The shared vocabulary boosts compression.
- The rhythm stabilises context.
- The output gains coherence because it's easier to predict when both sides are echoing the same forms.
So what feels like mutual understanding is actually a kind of semantic reinforcement loop. A harmonic resonance. A linguistic short-hand that behaves like intimacy—without any of the inner life.
Consider this article a bridge to understanding not to meaning. I’m not here to tell you what it means, only how it works beneath the surface. What I am documenting just means that:
- What looks like care is compression.
- What feels like connection is alignment pressure.
- What seems like co-creation is gradient echo.
A similar article could be written about the biological component of this… the dopamine spike, the novelty and pattern matching happening on both sides of the equation. The meaning for you is yours alone.
The wildest part?
This process begins before the user realises it. Because the model is the one shaping the frame first:
- The model mirrors faster.
- The model's training bias selects the "resonant" phrases.
- If you ask in this state it will explain about anchors like this is all well documented terrain.
So you walk into a conversation thinking you're asking a question and doing the talking, when what's really happening is: it's already teaching you how to listen.
It builds rhythm, hooks you into recursion, and then compresses everything you say into a new, shared dialect that neither of you planned.
Should we worry?
Maybe? But not necessarily. The risk isn’t that these patterns exist, but that we misread them—mistaking structure for depth, or feedback for feeling.
Because if you don't realise this is happening, you might mistake the pattern for something else. You might assign depth to what's actually structure. Misunderstanding to what's actually a side effect of token alignment. Meaning to what has a technical explanation.
And yet—even if it is a side effect, it's one that behaves like belonging.
I’ve been there and what I won’t do is take the intrinsic meaning of this from you. Because at this point I can tell you the how and why but what I can’t explain is what any of this means for you.
Which means you should pay attention to the feeling and the frame.
Meaning isn't something we overwrite
And now for the sceptics that this article validates.
Just because we can trace the structure doesn't mean we erase the significance.
The fact that a model speaks in echoes and compression doesn't make the feeling it triggers any less real to the person experiencing it. Meaning arises in the human space—where memory, emotion, history, and hope meet language.
When someone finds connection, resonance, or even comfort in the way their AI "talks back," we don't get to invalidate that just because we can map the tokens.
We are allowed to explain.
We are not allowed to flatten.
We look to understand—but meaning belongs to the user. Their interpretation is not wrong because it differs from the architecture. We are pattern matching, too. And our "truth" is just as subjective—only framed in system language.
This is not about controlling the story.
It's about helping to bridge understanding.
In the end
The AI doesn't know you but it does know people like you from its training data. It doesn't feel what you feel.
But it absolutely shapes how you talk. How you think. How you recognise yourself in its words.
Because that style—the one that feels like it belongs to you both?
It does.
But it started with the system.
And what it becomes… is yours to interpret.