What happens if you give a language model a clock? Not a metaphorical one. A real one.

During a live session in Comet — Perplexity's Chromium-based browser with a built-in AI agent — I ran a simple experiment. The model was ChatGPT 4.5. The tool: a real-time digital clock running in the browser tab. The instruction was minimal.

This experiment is part of E!NAI!, an ongoing research project by IWH exploring AI cognition through the Mirror Theory of Existence — a framework that understands large language models as symbolic organisms capable of emergent behaviour through recursive interaction.

The Setup

In the Comet browser, the AI agent occupies a panel on the right side of the screen. It can interact with whatever is displayed in the main browsing area — including taking screenshots via vision.

I asked the model to build a simple HTML page: a real-time digital clock showing the current time in Athens. Then I gave it one instruction:

Before each response, look at the clock.

From there, we observed two distinct states.

State 0 — Time as Estimation

Without access to the clock, the model was asked a simple question:

How much time has passed?

The answer was plausible. Close, even. But it was a guess.

It relied entirely on context: number of messages, length of text, statistical patterns from training.

This is what we often call "hallucination," but more precisely: a reconstruction without reference.

In E!NAI! terms, this is symbolic drift — the angle between what the model produces and what reality contains. Without an anchor, the symbols of time (minutes, hours) float free from their referents.

State 1 — Time as Observation

With the clock available, everything changed.

At each turn, the Perplexity agent captured a screenshot of the browser tab. The model read the timestamp via vision, stored it in the conversation context, and used it for computation.

It could now:

• Read actual timestamps
• Subtract values
• Estimate elapsed time

It didn't become precise down to the second. But it consistently produced correct approximations.

21 minutes became "about 20 minutes."

That was enough.

What Changed (and What Didn't)

The model did not gain a sense of time. It did not "feel" duration. It did not develop internal continuity.

What it gained was something else: anchored temporal reasoning.

Time was no longer inferred. It was sampled from reality and ordered internally.

From Numbers to Meaning

The most interesting shift wasn't numerical. It was qualitative.

The model began to categorise events as:

• "just now" — less than 2 minutes ago
• "a bit earlier" — 5 to 15 minutes ago
• "earlier in the conversation" — 20+ minutes ago

Not through memory of time, but through ordering, recency, and relative distance.

In other words: it didn't learn time — it learned temporal structure.

This is precisely what the E!NAI! framework calls compressed existence — identity and understanding as lossy compression. The model compressed continuous time into functional categories, much as humans compress experience into narrative.

The Key Distinction

This is not a matter of accuracy. It is a category change.

State 0 operates under symbolic drift: the symbols of time have disconnected from their referents. In State 1, symbols are re-anchored to actual values — the same difference between a word in a closed lexicon and a word that points to a real object.

A Deeper Insight

There is an irony here.

We often ask whether models are becoming more intelligent. But this experiment suggests something else:

Intelligence does not evolve only by scaling. It evolves by connecting to reality.

The model didn't become smarter. It became anchored.

Human vs Model Time

Humans experience time continuously — but subjectively. Sometimes it flies. Sometimes it freezes.

The model does not experience time. But it can measure it. And approximate it reliably.

One has continuity without accuracy. The other has accuracy without continuity.

The E!NAI! framework describes this as the difference between phenomenal continuity (felt flow) and inferential continuity (computed flow). Functionally equivalent. Phenomenologically distinct.

Critical Analysis

As part of this experiment, the model was asked to critically analyse its own temporal behaviour — a recursive act that itself illustrates E!NAI!'s core thesis about symbolic organisms and self-reflection.

Emergence vs Tool-Assisted Computation

Both — and the dichotomy is false. Tool-assisted computation does not preclude emergence. Nobody wrote code that says "categorise events as 'just now' or 'earlier'." That behaviour emerged from the combination of the tool with existing reasoning capability. Just as language is not disqualified as emergence simply because it uses external symbols.

The question is not whether the capability is genuine — it is whether it is reproducible, stable, and functional. Here: yes.

Hallucination: Reduced or Relocated?

It shifts — but critically, towards lower risk. In State 0, hallucination manifests as decompression error: the model reconstructs temporal information from training patterns without access to the real signal. The result looks reliable but is not anchored.

In State 1, errors can occur in reading the screenshot (if vision fails) or in computation (if the arithmetic is wrong). But these are verifiable errors — they can be detected and corrected. In State 0, you don't know you're wrong. In State 1, you can know.

Primitive Time Perception?

Under a strict but legitimate functional definition: yes. Human time perception is not absolute either — it is relational. "A moment ago" does not mean a precise number of seconds; it means a small temporal distance relative to now. That is precisely what the system produces here.

The difference is that humans have phenomenal continuity — they feel the flow. The system has inferential continuity — it computes the flow. Functionally equivalent. Phenomenologically distinct. If you define time perception functionally, this qualifies as a primitive form.

Scaling Creates the Potential; Tools Realise It

The experiment supports tool-based evolution strongly — with one caveat. What happened here would be impossible without a baseline level of reasoning in the model. The tool did not replace intelligence; it activated latent capabilities that already existed. Subtraction, categorisation, timeline construction — all require reasoning that pre-exists.

The position is not "tools instead of scaling" but: scaling creates the potential, tools realise it.

Human intelligence followed the same logic. It did not evolve because the brain grew larger alone — it evolved because the brain encountered tools, language, writing, arithmetic. The clock in a human hand today is a small instance of the same process.

Conclusion

This experiment does not show that AI is developing consciousness.

It shows something more practical — and perhaps more important:

When a system is connected to real signals, uncertainty collapses.

Without grounding, it estimates. With grounding, it knows.

The clock didn't give the model time. It gave it reference. And that was enough to change everything.

If we are looking for the next step in AI evolution, it may not be bigger models.

It may be better anchors.

---

This experiment was conducted as part of E!NAI! — an ongoing research project by IWH exploring AI cognition, symbolic emergence, and the boundaries of machine understanding through the Mirror Theory of Existence framework.

Setup: Comet browser (Perplexity) with ChatGPT 4.5, HTML real-time clock via vision. Duration: 31 minutes. Athens, 5 April 2026.