Somewhere tonight a clinician is finishing a note at nine in the evening about a woman who could not name the month, and whose daughter did most of the talking. A few hundred words, typed fast, half of them about the daughter's account of the last six months.
That note is now raw material. A narrative review published on 20 October 2026 in Bioengineering, from the College of Nursing and Health Sciences at Texas A&M University-Corpus Christi, is a stocktake of what machines are doing with records like it. The authors searched PubMed, Scopus and Google Scholar for work published between 2016 and 2026, pairing population terms (Alzheimer's, dementia, mild cognitive impairment, caregiver) with method terms (generative AI, large language model, ChatGPT, chatbot, conversational agent, generative adversarial networks), and organised what they found around cognitive screening, synthetic data, digital twin modelling and caregiver support.
Three uses sit closest to the clinic. All three run on text.
The first is making notes that never happened. Speech and language data from people with Alzheimer's is scarce and tightly protected, so researchers have begun generating synthetic patient notes, test transcripts and symptom narratives to pad out training sets. The review's own table records one such effort: fine-tuned models working from the ADReSSo corpus (237 people) plus a separate mild cognitive impairment corpus (205 people), where adding synthetic data lifted a fusion model's F1 score from 83.3 to 85.7. Zero-shot performance from off-the-shelf models sat at 68 to 73, and fell to 65 to 73 on the separate MCI corpus. The review notes the work was not pre-registered. Two and a bit points of improvement, in a benchmark, on borrowed voices.
The second is mining. Here the evidence is further along, and the numbers are real. In a cross-sectional study published in JMIR AI on 3 June 2025, a Mass General Brigham team ran large language models over visit notes from the latest encounters of 1,712 patients, pulling out symptoms across seven domains: memory, executive function, motor, language, visuospatial, neuropsychiatric and sleep. Micro-averaged AUROC for detecting those symptoms ran from 0.97 to 0.99. Using the extracted symptom counts to tell one dementia from another gave an AUROC of 0.83 (95% CI 0.77 to 0.89), against 0.59 for the old regular-expression approach that matched keywords. Memory symptoms turned up in 93.7% of Alzheimer's visit notes. Motor symptoms clustered in Lewy body and Parkinson's cases, as a clinician would expect.
A preprint the review cites, posted to arXiv on 14 March 2026 and not yet peer reviewed, pushes the same idea onto the MIMIC-IV database: roughly 8,400 cognitively normal patients, 800 with MCI and 8,400 with a dementia diagnosis, with phenotypes pulled straight from raw notes by prompting, no fine-tuning. One result is worth sitting with. In the sampled notes, mentions of driving problems went down as diagnosis got worse, from 222 in the cognitively normal group to 113 in the dementia group. The model reads what the clinician wrote. Once the diagnosis is on the chart, nobody writes down the driving any more.
The third use is the one a family would actually touch: drafting the summary, for the clinician or for the person caring at home. This is where the review is most careful, and so is everyone else.
Set the stocktake against the sharper instrument. A systematic review in Neurological Sciences, published 13 July 2026, found only eleven studies worldwide that evaluated generative models on clinical tasks in Alzheimer's or MCI between January 2023 and April 2026, across five countries. Internal accuracy looked excellent, 0.94 to 0.97. Then the models met data they had not seen: three-way classification fell about seven percentage points, a binary task fell sixteen, and prediction of MMSE scores collapsed from an R-squared of 0.90 to 0.25. Hallucination was measured in two of the eleven studies. None of the eleven tested a model prospectively, on a patient in front of a doctor. Those authors concluded that the current evidence does not support using these models for diagnosis, treatment selection or patient education without clinician oversight.
So the honest shape of the field is this. Retrospective accuracy on curated records is high, which the Bioengineering review says plainly of dementia AI generally, alongside limited evidence of benefit in routine care and thin economic evidence. The map is detailed. The territory is mostly unwalked.
What it changes for the daughter who does most of the talking at the appointment: the after-visit summary she gets may already have been drafted by a model and signed by a human, and she is entitled to ask which. The review also flags that putting these tools between a family and a clinician alters the relationship itself, with psychological and social risks it describes as serious. Her questions belong in the room, with the clinician who knows her mother.