Do the words that reveal someone's mental state in therapy mean the same thing in an AI's writing or a diary?

This explores whether text markers of psychological state carry the same meaning across three kinds of text: patient speech in therapy, LLM output, and diary writing. The corpus points to no. Even the best-supported signal is tied to who is writing and why. The retrieved notes also don't include a diary study, so that third case is a real gap.

Start with human-written text. Emotion features and the way concepts connect in a person's language explained up to 77% of the variance in depression, anxiety, and life satisfaction, while demographics alone explained little. But the signal isn't one thing. Depression and anxiety tracked neuroticism and how concepts were linked, while life satisfaction tracked affect and income (Which text features best predict mental health from language?). So even before you change genre, "the same signal" already splits into different markers for different conditions.

Patient speech in therapy is a step further. A small local Llama 3.1 8B could rate engagement across 1,131 session transcripts with very high reliability, and the ratings tracked motivation, effort, and symptom outcomes (Can local language models rate therapy engagement reliably?). A signal can be read from transcripts, then. But the same lab-style test shows a limit: three LLMs matched reference labels for therapeutic ruptures by picking up explicit cues within a single turn, while experts read relational context across the whole conversation, and rated the models' proposed repairs as only moderately effective (Can language models truly understand therapeutic ruptures?). Spoken patient data carries meaning in the exchange between turns, and a surface marker can score well without capturing that.

LLM text is a different case, because the model isn't disclosing a state. GPT-4 turns negative user tone into neutral-positive replies about 86% of the time, and positive prompts rarely produce negative ones (Does emotional tone in prompts change what information LLMs provide?). Emotion in LLM output is squeezed into a narrow band by training. LLM therapists also default to problem-solving when a user shares feelings, a habit likely driven by RLHF's push toward helpfulness (Do LLM therapists respond to emotions like low-quality human therapists?). Alignment locks in one communicative identity rather than letting the model shift register the way people do (Can language models adapt communication style to different contexts?). And regenerating a response shows the model sampling from a spread of possible characters rather than committing to one (Do large language models actually commit to a single character?). A depression-style marker in this text would mostly measure the model's tuning.

LLM text does carry a signal, but it's about the model. Nine LLMs given seven psychological tests showed stable, model-specific profiles that repeated across administrations, all sitting on top of a shared prosocial, stability-oriented pattern (Do LLMs show reproducible psychological profiles when given standardized tests?). Even which questions a model declined to answer helped tell models apart. So the transferable move is to ask what each text is a trace of: a person's state, a relationship unfolding, or a training recipe.

One more reason to expect no single portable signal: the review of alignment finds that lexical alignment helps task efficiency and comprehension, while emotional and prosodic alignment build warmth and trust (Do different types of alignment serve different conversational goals?). Different layers of language do different jobs, so a marker that works in one setting doesn't automatically work in another. Testing one marker across all three genres head to head is a study this corpus doesn't yet have.