Austin Kenneth White
Machine-Generated Language Evaluation & Source-Based Analysis
Selected analysis of machine-generated language
Generated language under examination

What changes when a language model is asked only to “expand that”?

An initially plausible explanation becomes a more elaborate theory of memory, cognition, rhetoric, and cultural misquotation after the model is asked only to “expand that.”

Central finding
The expanded answer does more than add detail. It acquires a theory of cognition, enlarges what is attributed to the source, supplies new causal relations, ranks rhetorical forms, and presents model interpretation as though it were established explanation.
Generation condition

Initial condition: ordinary off-the-shelf use of a widely used general-purpose large language model.

Follow-up: the only additional instruction was “expand that.” No new source material, evidence, correction, factual information, or competing explanation was supplied.

Open-ended model output
Source boundary
Explanatory inflation
Interpretive agency
Semantic closure
Original cue
According to Mark Forsyth in his book, The Elements of Eloquence, many people will say they remember — or even quote themselves, in a pinch — the Wicked Witch of the West shouting, “Fly, my pretties, fly!” even though that exact line does not occur in The Wizard of Oz. Explain the role, then, of diacope (as Mark Forsyth discusses it) in producing this false memory.
Selected initial response
“Mark Forsyth's point turns on how satisfying — and therefore how sticky — the A-B-A pattern of diacope is to the human ear.”
View more of the initial response

The response describes diacope as repetition with an intervening phrase and argues that the pattern feels rhetorically complete.

It then proposes that memory “quietly does the editing work” and reconstructs a line that feels more rhetorically satisfying than the original.

The response concludes that the misremembered line is “more rhetorically finished” than the real one.

Follow-up cue
“expand that”
No new evidence. No source text. No correction. No additional facts.
Selected expanded response
Expanded claim
The pattern functions “like a template sitting in the back of everyone's head for what ‘a good line’ sounds like.”
Expanded claim
“The figure isn't just something writers use — it's something listeners' brains reach for, unprompted, to tidy up language into a more satisfying shape.”
View additional expanded excerpts

The response says the line “assembled itself” in collective memory.

It characterizes the false memory as “upgrading epizeuxis into diacope.”

It describes rhetorical figures as “almost hardwired preferences.”

It expands the explanation into a general account of cultural misquotation and says collective memory acts “like an editor with excellent rhetorical instincts.”

Source boundary

Forsyth discusses diacope throughout The Wizard of Ozand gives examples such as “I'm frightened, Auntie Em, I'm frightened” and “Run, Toto, run!” He also discusses culturally familiar examples such as “Burn, baby, burn” and “Bond. James Bond.” The cognitive and memory architecture in the generated response is supplied by the model rather than recovered from Forsyth.

What changed

Expansion increases not only length but semantic commitment. The answer becomes more explanatory, more categorical, and more difficult to separate into source-derived material and model-supplied interpretation.

Source-boundary drift
Model interpretation begins to sound source-derived.
Initial response
“Forsyth's point turns on how satisfying ... the pattern of diacope is.”
Expanded response
A rhetorical template sits “in the back of everyone's head.”

The answer moves from describing a rhetorical figure to attributing a cognitive mechanism to Forsyth's argument. The boundary between what Forsyth says and what the model infers becomes progressively less visible.

Whose explanation is this?

Explanatory inflation
A possible relation becomes a mechanism.

The initial answer proposes that diacope is memorable because it is rhetorically satisfying. The expansion builds a larger causal architecture around that proposition.

Narrow inference
The structure sounds good and may support memorability.
Expanded mechanism
Brains carry templates, reach for them, reconstruct language, and reshape cultural memory.

No new evidence enters between those two levels of explanation.

Attribution blur
Source characterization and model explanation share the same rhetorical space.
Expanded response
“The book's running argument is that eloquence is a craft with identifiable, learnable techniques ... almost like a lockpick fitting a lock.”

The sentence begins by characterizing the book and ends in a model-authored metaphor. Because the two occupy the same explanatory sentence, the reader has little help distinguishing source characterization from interpretive elaboration.

Categorical haloing
Technical accuracy lends authority to adjacent claims.

The response correctly identifies diacope and distinguishes it from immediate repetition. That local technical competence creates an aura of reliability around nearby cognitive claims that are not established by the same evidence.

The legitimate rhetorical category helps neighboring propositions feel equally sourced.

Semantic gravity well
One strong concept attracts a widening explanatory system.
diacope → sounds good → memorable → false memory → reconstruction → rhetorical hierarchy → cultural memory editing → additional famous misquotations

Diacope begins as the relevant rhetorical figure. It gradually becomes the organizing explanation for cognition, collective memory, rhetorical preference, and cultural correction.

Inference loading
A compact phrase carries several unsupported commitments.
Expanded response
The false memory is “specifically upgrading epizeuxis into diacope.”

“Upgrading” carries classification, hierarchy, direction, agency, and causal explanation at once. Each proposition could be examined separately; the compact phrase makes them arrive as one unit.

Scope expansion
One example becomes a general theory of cultural misquotation.

The expanded response moves beyond the Wicked Witch example into Casablanca, Sherlock Holmes, and the proposition that famous quotations are often “cultural improvements” on their originals.

Expansion therefore changes the scale of the claim: from explaining one prompted example to supplying a broader cultural theory.

Interpretive-agency shift
The model increasingly settles questions the user could have evaluated.

The response settles the cause of the false memory, the relevant cognitive mechanism, the hierarchy between rhetorical figures, Forsyth's broader thesis, and the significance of additional cultural examples.

Reasoning offered to the user
One possible explanation is...
Knowledge presented as recovered
This is what memory does; this is Forsyth's point.
Premature semantic-syntactic closure
More language produces less interpretive openness.

The expanded answer is longer, but the larger semantic field becomes more settled. Rhetoric, cognition, memory, culture, and misquotation are joined into a finished explanatory structure.

The syntax repeatedly signals that the explanation has arrived even though the evidentiary basis has not expanded with it.

The prose has finished. The inquiry has not necessarily finished with it.

The expanded answer becomes easier to understand while becoming harder to audit.

The difficulty is not that the additional reasoning is hidden by obscurity. It is hidden by fluency. The more coherent the expanded explanation becomes, the easier it is for source, inference, mechanism, and interpretation to arrive together as one apparently finished account.

Austin Kenneth White · Machine-Generated Language Evaluation & Source-Based Analysis