I asked the Claude Large Language Model to compare two examples of a person’s off-the-cuff speech, on September 27, 2020 and April 6, 2026. It concluded that a comparison of this 80-year-old American politician’s unscripted and impromptu words showed detectable cognitive decline during those six years.

Claude reported, This is the most striking quantitative finding. The 2026 spontaneous sample has nearly 3× more mid-sentence interruptions relative to word count. This reflects a significantly higher rate of losing the thread, restarting, or abandoning clauses — a known linguistic marker researchers associate with cognitive aging. Scroll down for Project Structure, methods and sources.

There is no part of the process that involves any personal judgment. The prompt cannot be said to be leading. The transcripts cannot be said to be imagined. In this Large Language Model’s assessment, this 79-year-old shows quantifiable early signs of decline.

= = = = =

Claude Analysis

Query: Compare transcripts from September 27, 2020 and April 6, 2026 and see if there are signs of cognitive decline.


Sentence Structure

2021: Average 14.2 words per sentence

2026: Average 13.1 words per sentence

Sentences are slightly shorter in 2026, but more importantly, the structure has degraded. The 2026 speech contains far more fragmented, incomplete constructions and mid-thought restarts.

Self-Interruptions & Fragmentation

2021 em-dashes (thought breaks): 31

2026 em-dashes (thought breaks): 87

This is the most striking quantitative finding. The 2026 speech has nearly 3× more mid-sentence interruptions relative to word count. This reflects a significantly higher rate of losing the thread, restarting, or abandoning clauses — a known linguistic marker researchers associate with cognitive aging.

Example from 2026:

“He was injured quite badly and stranded in an area teeming with terrorists… Many—on top of everything else, they told the communities, actually, with—inside of Iran—the people of Iran…”

Repetition

The 2026 speech shows more within-passage repetition of the same idea:

“They want to always go as far away from the site of the shoot-down. They want to go as far away because they all head right to that site. You want be as far away as you can.”

In 2021, ideas are typically stated once and built upon. In 2026, the same point is re-stated 2–3 times in immediate succession, more frequently.

The superlative “amazing” appears 15 times in 2026 (vs. 0 in 2021), functioning as a low-specificity filler replacing more precise descriptors.

Narrative Coherence

2021: Despite being rambling in style, the speech follows discernible topic threads

2026: The narrative is harder to track. The speaker loops back repeatedly, loses sub-clauses, and at several points the grammatical subject of a sentence shifts mid-thought. There are more moments where meaning must be inferred from context rather than read from the sentence itself.

Qualifications & Caveats

A few important points to keep in mind: [Ed. Note: Shocking modesty from an LLM: an admission of uncertainty.]

* These are both extemporaneous speeches — not edited writing

* I’m not a clinician. Linguistic analysis of speech transcripts is used in cognitive research, but a real assessment would require neuropsychological testing, longitudinal data [ed note: which this query included], and clinical judgment — not a text comparison.

= = = =

Project Structure, methods and sources

Why and how I did this comparison, why I trusted the generally inaccurate and hallucinating Claude LLM engine’s results.

I don’t trust Claude any further than I can throw it, and I firmly believe Large Language Models are incompetent at most tasks they are now asked to perform. But counting words and parsing grammar seem likely to fall within LLMs’ actual area of competence (not corporate strategy, HR, customer service or aiming bombs).

My wife and I wondered if there were any semi-objective way to detect cognitive decline without an in-person test.

Given the difficulty of performing cognitive exams, neurologists have, for years, been looking for behaviors that are quantifiable early signs of decline. A Systematic Review of Natural Language Processing Techniques for Early Detection of Cognitive Impairment (review paper). The science says: unstructured speech is a quantifiable indicator.