Using Artificial Intelligence on Language for Diagnosing Alzheimer’s Disease Language contains signals about our health and disease. The words we choose when describing the same picture can differ depending on whether we are in pain or feeling anxious. This research seeks to uncover the state of mental health from language using artificial intelligence (AI), with a focus on detecting and predicting Alzheimer’s disease (AD). The idea that language can reveal indicators of AD has precedence in the literature. In a study by Snowdon et al. (1996, JAMA), nuns were followed for nearly 60 years. The study found that those who later developed AD had written diary entries with lower idea density—measured by the number of verbs, adjectives, adverbs, and prepositional phrases per 10 words. Building on this promise, we employed a feature-engineered approach, extracting linguistic markers such as the ratio of pronouns to nouns (e.g., saying “she” instead of “Nada” or “my friend” without specifying the referent), sentence length, and word complexity. Using this approach, we classified AD from healthy controls (HC) with high accuracy. Remarkably, applying the same feature set to Persian—a language from a different branch of the Indo-European family—yielded 90% accuracy. While this approach is highly effective, it is inherently limited by human-selected features. To overcome this, we use large language models (LLMs) for AD detection. Trained on labeled language samples, these models predict the probability of AD in unseen language samples. Explainability algorithms then reveal the most predictive linguistic features. This approach has shown accuracies higher than 90% in detecting the clinical syndrome and underlying pathology. Our results demonstrate that the sequence of words carries valuable information about both our current and future mental health. AI offers an accessible and highly sensitive tool to decode these hidden linguistic markers, enabling early diagnosis and improved management of neuropsychiatric disorders.