Before this study, disorganized speech in psychosis was mostly judged the old-fashioned way: trained clinicians listened, took notes, and used rating scales to decide whether someone's language had gone a bit off the rails. After this study, we have something much sharper. Researchers can now ask a language model whether each word fits not just the word right before it, but the whole conversational runway behind it. Same speech, very different flashlight.

When the sentence is fine but the story wanders off

The paper looks at positive thought disorder, a form of disorganized speech that can show up in schizophrenia and first-episode psychosis. A person may jump tracks, lose the thread, or sound disconnected even when short phrases seem normal. It has long been one of those "you know it when you hear it" problems, which is not ideal if you want science instead of vibes with clipboards.

Victoria Sharpe and colleagues studied speech from 60 patients with first-episode psychosis and 35 matched healthy controls, focusing on people who had not yet been treated. That gives a cleaner read on the symptom (Sharpe et al., 2025).

Local gossip versus the full family group chat

Language works on more than one scale.

Local context is the nearby stuff - the last word or two. If I say "peanut butter and...", your brain is already leaning toward "jelly." That is a tiny, neighborhood-level prediction.

Global context is the bigger story. If we have been talking about beach vacations, sunscreen, and bad hotel pillows, your brain uses that running theme to guess what comes next. It is less "what word usually follows this word?" and more "what fits the conversation we are actually having?"

The new study tested whether patients' speech was less sensitive to that broader context than to the immediate local one. In plain English: can word-to-word combinations still look okay while the larger message slips? Apparently, yes.

The GPT trick is sneakily clever

The researchers used GPT-3 to estimate how predictable each spoken word was based on different amounts of earlier context, from just one previous word up to 50 previous words. That let them compare short-range prediction with long-range prediction inside natural speech.

What they found was striking: disorganized speech showed disproportionate insensitivity to global context compared with local context. The speaker's language could still hook onto nearby verbal cues, but the larger conversational arc had less pull. The speech was not simply random. It was selectively off balance.

Even more interesting, this global-versus-local mismatch predicted clinical ratings of positive thought disorder above and beyond overall symptom severity. It did not track negative thought disorder, which is more about sparse speech. That matters, because psychiatry has a long history of tossing different language problems into one large and unhelpful drawer.

Why this matters outside the lab

If this finding holds up, it could help turn disorganized speech into a more objective clinical signal. That does not mean a chatbot is about to diagnose schizophrenia from one conversation.

But it does mean clinicians may eventually get better tools for tracking when speech starts slipping in a very specific way. Recent reviews argue that speech is a promising psychiatric biosignal and that natural language processing could help with symptom monitoring, relapse prediction, and more consistent assessment across clinics (Kappen et al., 2023; Zaher et al., 2024). Other work suggests language models can capture meaningful aspects of thought disorder rather than just producing fancy math wallpaper (Fradkin et al., 2023).

There is also a broader neuroscience angle here. Reviews of formal thought disorder point to disrupted coordination across language networks rather than one single broken "speech center" (Palaniyappan et al., 2023). If global context is specifically weakened, that gives researchers a tighter target for what is going wrong and where.

The catch, because there is always a catch

This is not a finished clinical tool. It still needs replication, testing across languages and cultures, and proof that it works reliably in messy real-world settings. Language models also come with baggage: training bias, privacy concerns, and the fact that "objective" systems can still make subjective mistakes in a lab coat.

Still, the conceptual leap is exciting. Instead of treating disorganized speech as a mysterious cloud of impressions, this study breaks it into a specific computational problem: the brain may still catch the nearby cue while missing the wider conversation.

Because once you can name the exact part of the language machinery that is wobbling, you are no longer just saying, "something sounds off." You are finally asking what kind of off. In brain science, that is often where progress starts - right after the confusion, and just before the neurons start acting like unruly cousins at Thanksgiving.

References

Sharpe V, Mackinley M, Nour Eddine S, Wang L, Palaniyappan L, Kuperberg GR. Selective insensitivity to global versus local linguistic context in speech produced by patients with untreated psychosis and positive thought disorder. Biological Psychiatry. 2025. DOI: https://doi.org/10.1016/j.biopsych.2025.06.001

Palaniyappan L, Homan P, Alonso-Sanchez MF. Language network dysfunction and formal thought disorder in schizophrenia. Schizophrenia Bulletin. 2023;49(2):486-497. DOI: https://doi.org/10.1093/schbul/sbac159 PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC10016399/

Fradkin I, Nour MM, Dolan RJ. Theory-driven analysis of natural language processing measures of thought disorder using generative language modeling. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging. 2023;8(10):1013-1023. DOI: https://doi.org/10.1016/j.bpsc.2023.05.005

Kappen M, Vanderhasselt MA, Slavich GM. Speech as a promising biosignal in precision psychiatry. Neuroscience and Biobehavioral Reviews. 2023;148:105121. DOI: https://doi.org/10.1016/j.neubiorev.2023.105121 PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC11219249/

Zaher F, Diallo M, Achim AM, Joober R, Roy MA, Demers MF, Subramanian P, Lavigne KM, Lepage M, Gonzalez D, Zeljkovic I, Davis K, Mackinley M, Sabesan P, Lal S, Voppel A, Palaniyappan L. Speech markers to predict and prevent recurrent episodes of psychosis: A narrative overview and emerging opportunities. Schizophrenia Research. 2024;266:205-215. DOI: https://doi.org/10.1016/j.schres.2024.02.036

Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.