A question of generative speech
The topic evolves around the issue of what are described as Large language models (LLMs). Although linguistics, philosophy and sociology have been studied as part of my university degree, I find language is still a complete mystery. Even the topic of LLMs is quite hard to follow. However based upon what has been said by academics in a number of videos this year, there’s no doubt there’s a likely revolution in terms of especially how spoken language is generated because LLMs have demonstrated there’s no need for a brain to be actively involved in the process.
Evidently there’s no doubt communication as viewed by many is more about the emotive side. It is rarely the words uttered but its more about the voice and how it resonates. Excitement, reassurance, persuasion, pauses in speaking, plus variations in tone, pitch and rhythm all add to the comprehensive communication a speaker can devolve. The surprise these days is speech is being seen as a communication medium that doesn’t even need a brain!
The experts admit they had once thought that language was impressionably tied to thinking. With LLMs that has been shown to be wrong. Take away language and one still can think. Those who lost their ability to speak could still solve math problems, understand written instructions, and continued to be able to sense other peoples’ emotions. Its said ‘Language is primarily a tool for communication rather than thought’.
Language is a defining characteristic of our species, but the function, or functions, that it serves has been debated for centuries. Here we bring recent evidence from neuroscience and allied disciplines to argue that in modern humans, language is a tool for communication, contrary to a prominent view that we use language for thinking. We begin by introducing the brain network that supports linguistic ability in humans. We then review evidence for a double dissociation between language and thought, and discuss several properties of language that suggest that it is optimized for communication. We conclude that although the emergence of language has unquestionably transformed human culture, language does not appear to be a prerequisite for complex thought, including symbolic thought. Instead, language is a powerful tool for the transmission of cultural knowledge; it plausibly co-evolved with our thinking and reasoning capacities, and only reflects, rather than gives rise to, the signature sophistication of human cognition.1
LLMs have shown that thinking isn’t necessary for speaking. The LLM models almost never give up even if they produce illogical strings or nonsense and that is much like what humans do too when arguing. That because if there is any sign of weakness people would leave AI in droves because there is an expectation AI knows everything, when in fact it doesn’t. Its why one practically never sees AI saying ‘I don’t know’.
Large Language Models
I first learnt about LLMs through Curt Jaimungal. The fall out from two of Curt’s videos is that ‘many of the firm views of psychology and neuroscience are now shaking under the impact of AI models and new data.’ Not only that, ‘the discovery of language models as mind mirrors is tearing open our ideas about knowledge itself.’ The real discovery is not the models, but a property of language scientists had not seen clearly before.
Most people now know the basic idea behind large language models like ChatGPT, Gemini, and Claude. These read huge amounts of text, then learn to predict the next word in a sequence. For example, with ‘The cat sat on the…’ the model guesses ‘mat’ or another likely continuation. It would for example be ‘the cat sat on the mat.’ It is clear language is auto-generative. The ‘corpus of language contains the structure needed to generate itself.’ The models do not invent this structure but uncover it. What these models are doing is learning the predictive web that’s already present in human language.
LLMs offers a very different way of looking at tools like GPT and LLMs reveal the hidden logic language already has. The onset of AI has given many hints that language is an autonomous informational system. Its one that runs on top of both silicon and human brains. In other words it needs no other input than its own and does not depend on a human mind to utter what appear to be long strings of thoughtful speaking.
Autoregressive LLMs generate text by predicting one token (a word or part of a word) at a time, using the previously generated tokens as input for the next prediction. This sequential process allows models like GPT to generate coherent text by conditioning the prediction of each new token on the sequence that came before it. The field is rapidly advancing to incorporate human speech (which is no doubt largely audist/oralist based), leading to the development of Speech Language Models. By training on large datasets of spoken conversation, those Speech Language Models aim to generate more natural and contextually appropriate spoken dialogue, including using conversational markers, like ‘yes’, ‘erm’, ‘aha’ and ‘okay’, whilst also handling turn-taking (a sort of exchange of banter) much like speaking/hearing people do in real-time conversations.
A brief overview of how LLMs work
Since I am not an expert in LLMs (or the programming of AI either) I had to ask Chat GPT for some guidance on how AI learns communication as I felt these explanations needed to be more precise. I did not want to write stuff that could be wrong in their explanations hence the next section was done with the help of Chat GPT.
Large Language Models (LLMs) have a huge database of words, even books, articles, documents. LLMs learns the following in order to speak fluently:
- How humans chat
- How questions work
- How to sound polite, funny, or serious
- How to connect ideas
Prediction is a very important component of LLMs. When one converses with the AI engine, it responds accordingly – there’s no doubt about that – but it doesn’t actually ‘think’ like a human. Instead its largely a guessing game based off numerous clues.
One asks/says/types something. The Chat AI scours its huge database and guesses the next word that makes the most sense. (This in LLM parlance is called a token). It then deducts the next word, and the one after and so on in the same way. In a sense its like a magician pulling a rabbit out of a hat, but in this case its a whole paragraph.
The LLM keeps a short term context window which records everything that has been asked of it for every session. Whilst talking the AI keeps its current context window (or memory bubble) active. It contains everything one has said (or typed) recently and also everything it has said back. This helps the Chat to stay on topic and not wander off it, not only that it avoids repetition.
The most important aspect in terms of the above is the LLM doesn’t actually ‘think’ like a human. To reiterate, the AI relies on words, stories, dialogues, explanations, questions and answers, learned patterns of grammar, associations between words, structures of reasoning, and common ways humans express ideas. These AI models converts patterns into meaningful responses. Its one word -> then the next -> then the next – it looks like the AI is driving a full conversation.
The above is very important in how humans are not exactly the generator of the conversations they are saying. The argument among the experts in terms of communication is it doesn’t actually ‘think’ like a human AND it looks like the AI is driving a full conversation – which quite demonstrably shows how humans communicate. They don’t actually think the strings of words that are used in conversations. What LLMs are doing is exactly the same as what humans do – this being that speech and linguistic input gets done without thinking.
Next is a video that gives a basic (and non-academic) overview of what is being espoused in terms of communication.
Chase Hughes – Language Has a VERY DARK Secret (23rd November 2025). ‘You’ve been walking around thinking you’re seeing the world, but really you’re just seeing whatever your language lets you have. Once you catch it in the act, it feels like finding a glitch in your own brain. Until this hits you, you don’t realize how much of your life has been pre-chewed, pre-labeled, and handed to you like baby food.’ (That is exactly the essence of it! Those who thought speech was irretrievably linked to the mind can no longer be relied upon to assert its superiority – especially the audists and oralists).
Its why academics are saying human speech is essentially its own agent, a thing with a mind of its own. When the specialists refer to what are conversations and speeches, this is of course a specialist area only hearing people could do – which is to investigate, and then acknowledge that much of what comes from conversation isn’t the work of the brain. Its why, in 2025/2026, the incidence of LLMs is now helping academics to understand something that has been debated since the time of the Greek philosophers.
What it boils down to is Language behaves like an organism that lives in peoples’ heads.
Language is an autonomous informational system and it runs in people’s brains, but its not the same thing as the person’s sensory or emotional life. This language feeds on patterns of symbols and produces streams of words, both in speech and inner monologue. It is ‘downloaded’ into people from infancy. Nobody asks for consent. Long before anyone can read a waiver or privacy policy, the language of one’s culture has already installed itself in a person’s nervous system. It means a kind of internal language model is running and it uses a person’s brain as a substrate, but it is its own thing and does what it is specialized to do – which is generate the next string of words.
Ncuti Gatwa (the 15th Doctor Who) is absolutely right when he says ‘So I got the chance to learn BSL, which was amazing! And I think its just such a more effective, beautiful, intentful, way of communicating. You have to really commit yourself to who you’re talking to… This is such a different level of communication…’ Essentially he describes the extra layer of thought and cognition that’s necessary to learn sign language. In other words it can’t simply be ‘downloaded’ – it essentially needs an extra layer of thought/cognizance which isn’t the case with naturally acquired speech. Clip is from Doctor Who Unleashed which looks at the making of the Doctor Who episode ‘The Well’ which also stars Rose Ayling-Ellis. Note: BSL means British Sign Language.
There’s also the question of acquiring second (or more) spoken languages. If the original is ‘downloaded’ what are the others? Are they ‘downloaded’ too? Or is there a different process at work? (As Ncuti Gatwa stressed in the above video). One thing is greatly evident from this and its no matter what other languages anyone could speak, almost everyone is comfortable speaking their own native tongue – however that of speaking another takes quite a bit of effort.
Curt Jaimungal – Language Without Meaning: How LLMs Exposed Our Biggest Illusion (11th June 2025).
In the next post The (Terrifying) Theory That Your Thoughts Were Never Your Own (Youtube 21st July 2025) we find that ‘Professors Elan Barenholtz and William Hahn propose that language is not a tool we use but a self-generating organism that uses us.’
(Updated 10th December 2025 to include an extra post which it is hoped will give those still wanting to learn more about how LLMs work some clearer idea of the mechanics that are used to generate language, and indeed the impressive responses that one receives when one asks Chat GPT, or Gemini, or Perplexity etc a question or questions. or even seeks a mere chat.)
Originally it had been said: In the next post The (Terrifying) Theory That Your Thoughts Were Never Your Own (Youtube 21st July 2025) we find that ‘Professors Elan Barenholtz and William Hahn propose that language is not a tool we use but a self-generating organism that uses us.’
However that is now the fourth post in this series. The next (the third) is Basic Large Language Model understanding – which is an attempt to give those, who do not know the mechanics of LLMs very well, an opportunity to get to know the LLMs’ systematics better. It will also help with the fourth post (The (Terrifying) Theory That Your Thoughts Were Never Your Own) because that post and the video included involve considerably more knowledge and insight.
