The aide writes "resistive to care." The daughter, calling that night, says her mother was fine at lunch and then would not let anyone near her left shoulder. The night nurse thinks the shoulder started Tuesday. The chart holds a pain score of zero and a dose of paracetamol timed 06:40. Four accounts of one afternoon, and no agreement about whether anything hurt.
Scenes with that shape are the ordinary texture of pain in dementia, and a Perspective paper published this week in Frontiers in Pain Research is about what a computer should be allowed to do with them.
The authors, Anitha Saravanan, Will Lewis, Jinbing Bai and Staja Booker, working at US universities including Northern Illinois University and Illinois State University, put the difficulty in the information rather than the instruments. Validated observational tools, caregiver and staff knowledge, clinical examination and repeated within-person assessment remain the foundation of care, they write. The trouble is that what those methods produce may be incomplete, conflicting, or documented across different settings, which leaves the synthesis sitting on an already stretched clinician.
Their line on generative AI is drawn early and drawn hard: "it should not independently detect, diagnose, or determine treatment for pain."
What it may do, in their framework, is organize. Seven testable use cases sit under three core functions: capacity-sensitive elicitation, source-attributed synthesis, and longitudinal support and communication. The seven include adaptive pain interviews, spotting possible pain-related content in everyday language, structuring behavioral observations, multi-informant synthesis, synthesis of validated nonverbal indicators, longitudinal pain-related diaries, and caregiver communication support.
Three safeguards bound all of it. Only appropriate, source-identified information goes in. Consent or assent is required when the system interacts directly with the person. And the output has to separate observation from interpretation, show agreement and disagreement, identify missing information, and present uncertainty along with other possible causes of distress. Every use case ends the same way, with mandatory clinician verification before anything is acted on or written into the record.
Read that third safeguard slowly, because it is the whole design. "She pushed the aide away" belongs on one line. "She was agitated" belongs on another, marked as somebody's reading of the first. If you have ever watched a single word in a chart harden into a fact that follows a parent from shift to shift, you already know why that separation matters more than any clever model.
The paper's own worked example is deliberately unglamorous. Behavioral narrative structuring, for a nurse or direct-care worker in long-term or acute care: source-tagged notes and items from validated observational tools go in, a structured summary comes out, and that summary stays linked to the original observations so the nurse can check it against the source. The link back is the product. A summary you cannot trace is a rumour with better formatting.
The framework also refuses one size. A person with mild impairment may still give a meaningful self-report and might use an adaptive interview and a longitudinal diary. A person with advanced dementia may need synthesis of observational, caregiver and clinical information with no direct AI interaction at all. The three functions are complementary, and the authors are explicit that they need not run in sequence; one, two or all three may apply.
The longitudinal function is the one with the sharpest promise and the sharpest risk. Its job is to notice changes from the person's usual pattern over time, which is exactly the signal that gets lost across a weekend of agency staff. It is also the job most likely to fail by succeeding too often. Anyone who has worked a ward knows what happens to a system that flags everything: it gets muted, and then it is furniture. The paper's answer is governance, a clinician at every gate, which is a real answer and an incomplete one until somebody builds the thing and counts.
Which brings the honest accounting. This is a Perspective article. No patients were assessed, no system was built, nothing was trialled. As one summary of the work put it, "its contribution is a proposed structure for future research and governance", with adoption contingent on prospective studies.
So what does a daughter do with it today. Nothing to a medicine cabinet, and nothing to a care plan; that conversation belongs with the clinician who knows her mother. What it hands her is a vocabulary for the next care conference. Who observed that, and when. What does the validated tool score say, and who scored it. What changed from her usual pattern. What do we still not know. Those are now someone's published design requirements, which makes them easier to ask out loud.