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 were thinking about Trump’s obvious cognitive decline—obvious to us anyway—and wondered if there were any semi-objective way to detect it.

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).

I considered using tweets as the source material, but realized that they were not spontaneous, but rather created with some time and thought.

Also, they might be written by different authors. The President says he writes all his own tweets, but he says lots of things that turn out not to be true.

However, it is absolutely true he speaks without a script at news conferences. So, I downloaded the transcripts of news conferences on September 27, 2020 and April 6, 2026 from Presidential News Conference Transcripts (transcripts). They were edited only to remove the names of the reporters and their questions.

I uploaded the two clean files, and copied all the tests suggested by the the review paper. Then I asked Claude to “compare these two documents for evidence of changed cognitive ability. Both prompts provided the same results, although it is possible that Claude merely copied criteria from the same academic paper.

Because of the danger of AI slop, I checked to make sure the review was real. It is cited in multiple sites of peer-reviewed papers that were written before the common use of LLMs to hallucinate research. The transcripts were checked by finding reports from actual media outlets on those dates, indicating the content was accurate.

This is not Trump Derangement Syndrome nor is it confirmation bias. 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.