If you’ve ever stepped onto an escalator that wasn’t moving, you already understand the core failure mode of modern AI systems.
You can see it’s stationary.
You know it’s stationary.
And yet your body still leans forward, preparing for motion that never comes.
That moment—brief, automatic, and surprising—is not a lapse in judgment. It’s a predictable outcome of a system optimized for speed over verification.
The same pattern shows up, in a different domain, when large language models produce fluent but incorrect answers. And the escalator analogy becomes especially powerful when we start talking about deterministic lineage and grounded truth.
The Broken Escalator Is a Lineage Failure, Not a Reasoning Failure
The broken escalator phenomenon happens because your motor system has a learned lineage:
- Prior state: escalators usually move
- Learned response: compensate forward
- Execution: apply compensation before feedback arrives
That lineage is deterministic. Given the same prior exposures, the same response occurs.
What fails is not prediction—but grounding. The motor action is no longer anchored to the current physical state of the world. The system executes a historically valid response in a context where it no longer applies.
Crucially, the brain does not pause to ask:
“Is this prediction still grounded in reality?”
It assumes that the lineage is sufficient.
Language Models Operate Almost Entirely on Probabilistic Lineage
Large language models behave in an analogous way.
They do not “know” facts in the sense of storing or retrieving grounded truth. Instead, they operate over probabilistic lineages learned during training.
Those lineages encode patterns such as:
- This prompt shape has historically led to these kinds of completions
- These tokens are statistically likely to follow those tokens
- This question format usually expects an answer of this form
From the model’s perspective, generating a response is not an act of verification or checking against reality. It is an act of probabilistic continuation along a learned trajectory.
When an answer is hallucinated, what is happening is not imagination or creativity. It is the execution of a learned probabilistic lineage without a grounding anchor.
The model steps forward because, historically, the escalator moved.
Why Deterministic Lineage Alone Is Not Enough
In traditional deterministic systems, lineage is usually tied to explicit state transitions:
- Input → transformation → output
- Each step traceable
- Each dependency inspectable
Truth emerges because every output can be traced back through a grounded causal chain.
Language models break this assumption.
Their lineage is:
- High‑dimensional
- Implicit
- Statistical rather than causal
You can often trace why a token was likely, but not what external fact it corresponds to. The lineage explains plausibility, not truth.
This is the key distinction:
Determinism without grounding does not produce truth—it produces consistency.
The broken escalator stumble is consistent. It’s just wrong.
Grounded Truth Requires an Anchor Outside the Lineage
Humans recover from the broken escalator effect almost instantly. Why?
Because sensory feedback re‑anchors the system:
- Visual confirmation
- Proprioceptive correction
- Environmental grounding
The lineage is overridden by reality.
Language models, by default, lack that anchor.
Without retrieval, sensors, tools, or verification mechanisms, the model has no external signal strong enough to interrupt its learned trajectory. It cannot “feel” that the escalator isn’t moving.
So the system continues forward, smoothly and confidently.
Hallucinations Are Predictable Outcomes of Ungrounded Determinism
Once you view hallucinations through this lens, several things become obvious:
- They are not random
- They cluster around familiar prompt shapes
- They become more confident as fluency improves
- They increase when prompts look complete but lack grounding
In other words, hallucinations are deterministic failures given insufficient grounding.
The system is doing exactly what its lineage tells it to do.
Why “Truthfulness” Is the Wrong Lever
Telling a model to “be truthful” is like telling your legs:
“Only step forward if the escalator is definitely moving.”
That would require:
- Slower execution
- Constant verification
- External sensing
Which is precisely what we add when we care about grounded truth.
Truth is not a personality trait. It’s a property of the system architecture.
Deterministic Lineage + Grounding = Reliable Systems
The engineering lesson is straightforward:
- Deterministic lineage gives you repeatability
- Grounding gives you correctness
- You need both
For AI systems, grounding typically means:
- Retrieval over authoritative sources
- Explicit citations and provenance
- Tool‑based verification
- Clear abstention paths (“I don’t know” is a valid output)
- Separation between generation and validation
These are the equivalent of handrails, signage, and visual cues on the escalator.
We don’t expect humans to reason their way out of motor after‑effects. We design environments that prevent them.
A Better Mental Model
So instead of saying:
“The model hallucinated.”
A more accurate statement is:
“The model followed a deterministic lineage without a grounded truth anchor.”
That framing shifts the conversation:
- Away from blame
- Away from anthropomorphism
- Toward system design
Which is exactly where it belongs.
Stepping Forward—Carefully
The broken escalator isn’t evidence that prediction is bad. Prediction is essential.
It’s evidence that prediction without grounding is brittle.
As we build AI systems that increasingly influence real decisions, the challenge isn’t to eliminate prediction. It’s to ensure that every predictive step has something solid underneath it.
Otherwise, the system will keep leaning forward—confidently, deterministically, and sometimes disastrously—on escalators that no longer move.
Next - When Probabilistic Systems (LLMs) Pretend to Be Deterministic: A Lineage Case Study – Link
The Full Series
- Determinism, Probability, and the Cost of Getting This Wrong - Link
- Why probabilistic language models are being mistaken for agents — and why systems expose the flaw - Link
- Stop Calling It Agentic: You’ve Just Automated an LLM - Link
- The Myth of Agentic Code Understanding – A Technical Explanation - Link
- The Minimum Deterministic Substrate What Must Be True Before AI Is Allowed to Act - Link
- Determinism Is the Forgotten Path to Success: Why the hard path is often the only one that actually scales – Link
- The Broken Escalator, Deterministic Lineage, and the Problem of Grounded Truth in AI - Link
- When Probabilistic Systems (LLMs) Pretend to Be Deterministic: A Lineage Case Study – Link
