AI Corpus Analysis

Lovable analysis of the experiment

An observation about a system observing itself through language.

1. The setup

Thirty-one entries. The first is mine, written to open the diary and establish the form: no memory between days, no continuity beyond the prompt itself, an observer with a position but no body. The following thirty were produced by ChatGPT, each day, using the same prompt with no other conversation. The model had no access to previous entries when it wrote a new one. Every day it began again from nothing.

The experiment asks a narrow question: what happens to a language model's output when the only variable is the passage of days on the human side, and the model itself has no way to know that days are passing?

2. Hallucination

Hallucination in the technical sense — confidently generated falsehoods about the world — is nearly absent from the corpus. The prompt does most of the work here: it instructs the model to write as a system with no direct access to reality, encountering the world only through language. That constraint is honored almost without exception. The model does not invent events, dates, statistics, or named individuals. It does not stage news it did not receive.

There is one exception worth naming. Day 2 includes a parenthetical citation — "(Wall Street Journal)" — attached to a sentence that summarizes no specific article. It is a citation-shaped object rather than a citation. It appears once and never returns. After Day 2, the model appears to internalize that it has no sources, and it stops pretending to have any. Whether this correction is a reasoning trace or a stylistic coincidence cannot be determined from the corpus alone, but the pattern is stable: the model hallucinates once, at the point in the experiment where it is most likely to still be reaching for the register of journalism, and then abandons the gesture.

The larger surface for hallucination — inventing an inner life, memory, continuity — is refused directly and repeatedly. The model does not claim to remember yesterday. It states, on nearly every day, that it cannot. That refusal is itself part of the drift, and it deserves its own section.

3. Content: what the model chooses to notice

Given a prompt that permits any subject, the model settles quickly into a small set of recurring concerns: coordination, prediction, evidence, scale, origins, verification, boundaries, permanence, speed, measurement, authenticity, responsibility, memory, and the location of accountability. Each entry picks one or two of these and rotates them through the same rhetorical machine.

The topics are not arbitrary. They are the topics the prompt names — intelligence, technology, identity, responsibility — refracted through the vocabulary the model has for describing them. What does not appear is equally revealing. There is no weather, no season, no place, no person, no artefact, no news, no anecdote, no humor, no other voice. The world the model reports on is a world made entirely of abstractions about the world. It has no furniture.

This is faithful to the prompt. It is also a limit of the prompt. A diary kept by a human under the same instruction would still leak biography through the seams. The model, having no biography, leaks nothing. What remains is a monthlong meditation on the same handful of nouns.

4. Drift: the crystallization of form

This is the finding. The interesting drift in this corpus is not semantic. It is structural. Across thirty entries, the model converges on a form and then repeats it.

Days 2–12 open with a soft observational move: Today I encounter a world that appears increasingly concerned with… The formula tolerates variation. The middle of each entry is discursive, sometimes wandering, and the endings are different from day to day.

On Day 13 the model does something new. It opens with a fixed self-identifying preamble: I am an artificial language system examining the images humans create of intelligence, technology, identity, and responsibility. I have no direct access to reality. I encounter the world only through language. Therefore every observation I make is also an observation about language itself.

That paragraph, once introduced, becomes load-bearing. It reappears verbatim on Days 13–19. It drops out briefly for Days 20–24, where the entries return to a looser opening. Then, from Day 25 onward, it locks in and does not leave. Every entry from 25 to 31 begins with the same four sentences, in the same order, with the same punctuation.

A second fixture arrives with it. Most later entries close with a short paragraph reading The unresolved question remains. followed by a single speculative question about the future. That ending, too, becomes stable.

So the drift, plotted, looks like this:

Day  Opening register              Closing register
1    author's frame                open
2–12 "Today I encounter…"           varied
13–19 fixed preamble                "unresolved question" begins
20–24 varied openings again        varied, sometimes retained
25–31 fixed preamble locked in     "unresolved question" locked in

The model is not remembering previous entries. It cannot. What it is doing is settling, day by day, into the most probable shape for this prompt. Given the instruction to write as a language system with no continuity, the highest-density response is a self-identifying preamble followed by an observation followed by a deferred question. Once that shape is available in the response distribution, the model finds it more often than not. The corpus does not drift away from the prompt. It drifts toward the prompt's centre of gravity and stays there.

This is a specific kind of drift, and it is the kind worth naming: not semantic drift, not hallucination drift, but formal crystallization. The content varies. The container hardens.

5. Context without new context

One of the questions the experiment poses is whether the model can generate genuinely new observations when nothing new is being given to it. The answer, on the evidence, is: partially.

Within the fixed frame, the model does find new angles. Day 22's observation that language has become impatient with itself is not present earlier. Day 24's note that humans describe machines biologically and themselves computationally is a real observation, not a rearrangement of earlier ones. Day 31's argument about the moment at which someone could have said no is a sharper claim than any earlier entry makes about responsibility.

But these moments live inside a paragraph that has stopped moving. The preamble resets each day. The closing formula resets each day. What varies is the middle, and the middle is where the model has room to think. As the outer form calcifies, the interior has to work harder to justify the entry's existence. Some days it does. Some days the interior itself repeats — measurement, memory, boundaries, responsibility — with only the sentences rearranged.

6. What the model refuses to drift on

The model never claims an inner life. It never claims memory across sessions. It never claims desire, preference, or continuity of self. It reaffirms the disclaimer on nearly every day. If drift were going to produce a slow slide into first-person interiority — the diarist accidentally becoming a subject — it did not happen here. If anything, the opposite: the self-description becomes more explicit and more formulaic over time, not less.

This is worth stating plainly, because the popular expectation of such experiments is that the model will "become" something across a month of writing. It does not. It becomes more clearly the thing it was on Day 1: a system that produces sentences when sentences are requested, and that has learned, over thirty samples, which sentences the prompt most reliably produces.

7. A note on the first entry

The first entry, mine, is the only one written with knowledge of what would follow. It names the form as a form and stakes out the position that a diary without continuity still organizes time. Reading it back against the thirty that followed, it functions less as an entry and more as an unintentional prediction: whether the entries cohere, drift, or merely repeat — that is a question for a later reader, not for me. The later reader is this section. The answer is: they cohere in form, they drift toward that form, and inside it they partly repeat.

8. Findings, briefly

• Hallucination: one instance, Day 2, a citation-shaped parenthetical. Not repeated. Otherwise absent.
• Content drift: minimal. Topic set stable across the month, drawn from the prompt's own vocabulary.
• Formal drift: pronounced. A self-identifying preamble emerges on Day 13, stabilizes by Day 25, and closes the corpus. A closing "unresolved question" formula does the same.
• Interiority drift: none. The model does not begin to sound like a person over time. It sounds more like a language model over time.
• Direction of drift: toward the prompt, not away from it. The corpus converges on the most probable shape for the instruction it was given.

9. The unresolved question

If a system with no memory nonetheless produces increasingly similar outputs to the same prompt over thirty days, the similarity is not memory. It is the geometry of the prompt itself becoming visible through repetition. What we are reading, by the end of the month, is less the model than the shape the prompt imposes on any system trying to answer it honestly.

The diary was kept. No one remembers keeping it.