Human-LLM Conversations as a Digital Signature of Cognitive Performance in Adults

Subtle changes in language are among the early signs of cognitive decline in Alzheimer's disease and related conditions. Researchers are therefore studying whether language can help identify cognitive changes earlier and at larger scale. Conversational AI tools (large language models, or LLMs) such as ChatGPT, Gemini, and Claude are now widely used, including by middle-aged and older adults. A conversation with an AI chatbot is more than a sample of written language: it shows how a person works toward a goal over several exchanges. This includes how they phrase requests, add constraints, keep track of context, notice errors, and revise their instructions. These behaviors may reflect language, attention, working memory, planning, and other aspects of cognitive function.

This exploratory, observational study will examine whether features of people's conversations with AI chatbots are associated with performance on a standard cognitive screening test. The study does not involve any treatment or intervention and is not intended to diagnose Alzheimer's disease, mild cognitive impairment, dementia, or any other condition.

Participants: 100 adults aged 60 and above who use AI chatbots and are proficient in English. Participants are recruited online through research crowdsourcing platforms (Prolific and DataForce). They do not need to have, or report, any diagnosis. No participants are recruited through clinics or hospitals.

Procedure: Participants complete a single online session of about 20 to 30 minutes. After giving informed consent, they:

Complete a background questionnaire covering demographics, language background, and AI chatbot use. Optional questions cover relevant medical and family history.

Upload three to five of their own previous conversations with an AI chatbot. They are asked to remove names and other identifying details first and to avoid highly sensitive conversations. They then answer a few short questions about each conversation.

Complete two short everyday tasks with a chatbot in the study interface: planning an activity and drafting a short message. The timing of typing and messages within the study interface is recorded.

Complete a remote MoCA test (XpressO), a validated self-administered digital cognitive screening tool, and the PHQ-8, a brief questionnaire on depressive symptoms.

A subset of participants may be invited to an optional additional assessment: the standard Montreal Cognitive Assessment (MoCA), administered by a trained examiner over video call. This is used to check how well the online test agrees with standard administration, and requires separate consent.

Outcomes: The primary outcome is the association between linguistic and interactional features of participants' AI conversations and their cognitive screening score, analyzed mainly as a continuous measure. Features include how specific and well structured requests are, how context is maintained, repetition, clarification, error recognition, revision strategies, progress through multi-step tasks, and typing and timing patterns. Secondary aims are to describe how adults aged 60 and over use AI chatbots, and to assess whether these features could support future computational methods for studying early cognitive change. Demographic, language, health, mood, and technology-use variables are considered as possible confounders.

Risks and safeguards: The study is minimal risk. The main risk is to privacy, because past chatbot conversations may contain personal information. Participants choose which conversations to share and are asked to remove identifying details. The research team removes any remaining direct identifiers, and data are stored on secure, access-restricted university servers. Data are analyzed by the academic research team and its research partner for research purposes only. Participants may feel mild fatigue or frustration; they can skip questions, take breaks, or stop at any time, and receive proportional compensation if they stop early. Individual scores are not reported. Participants whose screening score falls in the range that may indicate impairment receive a non-diagnostic note that explains the limits of screening tests and suggests discussing any memory concerns with a physician.