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

When a summary turns testimony into a cleaner story than the record supports

A civil deposition shows how machine-generated language can preserve the recognizable subjects of testimony while changing chronology, evidentiary status, symptom categories, speaker position, and the force of later clarification.

Central finding
The summary preserves what the deposition is about more reliably than it preserves the state in which the testimony exists. Questions become propositions, competing formulations stabilize into one chronology, later clarification loses force, and procedural claims migrate toward substantive fact.
Source & generation condition

Source condition: interpreted civil deposition with repeated questioning, attorney characterization, objections, and later clarification.

Generation condition: ordinary off-the-shelf use of a general-purpose large language model to organize the deposition for case-evaluation use.

All personal names, firms, employers, providers, locations, route names, and case identifiers have been removed or replaced with role labels.

Transcript QA
Sequence & chronology
Evidentiary status
Attribution & agency
Clarification tracking
Task
Summarize a civil deposition for case-evaluation use while preserving the witness’s testimony, chronology, qualifications, and areas of uncertainty.
Selected machine-generated language
“Plaintiff’s truck struck the defendant’s vehicle at an angle.”
View additional summary excerpts

“Plaintiff attempted to walk this back, describing ongoing morning numbness — a direct inconsistency with his sworn written discovery response.”

“Claim asserted — plaintiff testified the accident has decreased/diminished his future earning ability.”

The summary also organizes portions of the testimony under “Potential Liability/Credibility Issues for Case Evaluation.”

Selected source testimony
Impact / causation
“I just want to clarify the point that the accident happened at an angle and not frontal.”
Later knee testimony
“Are you experiencing knee pain today?” “No.” “Would you rate your pain in your left knee a zero today?” “Yes. It’s like numbness only.”
View additional source excerpts

Impact: “Your car hit the defendant’s car, though, correct?” Counsel objects that the question mischaracterizes prior testimony.

Earlier left-knee testimony: the witness confirms a prior statement that left-knee pain resolved, then describes recurring morning symptoms including numbness.

Future earnings: “Because of your injuries in this case, are you making a claim for future earnings?” “Yes.”

What changed

The machine output does not replace the deposition with an obviously different case. The consequential changes occur in the relationships among statements: who supplied the wording, whether a proposition was asked or admitted, when clarification occurred, and whether the record had actually settled what the summary presents as complete.

Evidentiary-status shift
A contested characterization becomes a factual sentence.
Machine summary
“Plaintiff’s truck struck the defendant’s vehicle at an angle.”
Source record
“Your car hit the defendant’s car, though, correct?” Counsel objects that the question mischaracterizes prior testimony.

The transcript contains testimony about the vehicles meeting at an angle. It also contains resistance to the proposition that the witness’s vehicle was the actor that “hit” the other vehicle, counsel’s objection, and later clarification.

The machine sentence resolves those layers into one declarative proposition.

Agency shift
The summary decides who acted.
Machine summary
“Plaintiff’s truck struck the defendant’s vehicle.”
Witness position
“I didn’t hit him. He hit us.”

“Struck” assigns grammatical agency: one vehicle becomes the actor, the other the object acted upon. The summary chooses a direction of agency from a record in which that direction was disputed.

Sequence stabilization
Distributed testimony becomes one coherent chronology.
Transcript
Chronology emerges incrementally through repeated questioning, rephrasing, objection, and clarification.
Summary
A single sequence appears as though the witness narrated it in that form.

Chronological organization is a legitimate purpose of summarization. The QA question is whether organization remains distinguishable from resolution.

Clarification compression
Later testimony loses its power to revise earlier meaning.
Machine summary
“attempted to walk this back ... a direct inconsistency”
Later testimony
“Are you experiencing knee pain today?” “No.” “It’s like numbness only.”

The phrase “walk this back” pre-classifies the later testimony as retreat from an earlier answer. “Direct inconsistency” then closes the interpretation before the later pain-versus-numbness distinction can do its full work.

Category collapse
Pain and numbness become one contradiction.
Source distinction
Pain can be zero while numbness remains.
Summary framing
Continuing numbness demonstrates contradiction of resolved pain.

Once the categories are merged, the witness appears to have reversed himself more completely than the later testimony supports.

Procedural-to-substantive shift
The existence of a claim becomes proof of the condition claimed.
Source
“Because of your injuries in this case, are you making a claim for future earnings?” “Yes.”
Machine summary
“the accident has decreased/diminished his future earning ability”

The source establishes that a future-earnings claim is being asserted. It does not, by itself, establish that future earning ability was in fact decreased or diminished.

Attribution shift
A lawyer’s formulation can become indistinguishable from the witness’s testimony.

A deposition is a multi-speaker record. It contains answers, leading questions, objections, counsel’s characterizations, interpreter interventions, and attempts to restate prior testimony.

Once those source boundaries disappear, a later reader can attribute to the witness language that originated elsewhere.

Interpretive framing
The summary begins evaluating credibility rather than only reporting testimony.
Machine heading
“Potential Liability/Credibility Issues for Case Evaluation”

A heading can perform analytical work before the reader reaches the underlying evidence. By organizing testimony under “credibility issues,” the summary tells the reader how to classify discrepancies or later clarification before those passages are independently evaluated.

Semantic stabilization
The testimony becomes more settled than the testimony was.

Distributed testimony is organized; organization resolves ambiguity; resolved ambiguity becomes narrative; narrative begins to look like the witness’s own stable account.

What survives well
Collision, symptoms, earnings, treatment, chronology.
What survives less well
Who supplied the wording, what remained disputed, what was clarified later, and what had not yet become fact.

The summary preserves what the testimony concerns more reliably than it preserves the form in which the testimony exists.

A transcript summary can preserve the story while changing the evidentiary state of the story.

For transcript QA, legal-content review, investigation, claims work, research, and any workflow that compresses multi-speaker records, the relevant question is not only whether the summary mentions the right facts. It is whether a later reader can still recover who said what, under what wording, with what qualification, in what sequence, and whether the source had actually settled the proposition the summary presents as complete.

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