AI Knows Nothing: A Collingwoodian Exploration of Machine-Generated Truth

Theory

Apr 23

Written by Tuin T. O. Scheffer

Tuin is currently doing the Research Master’s in Modern History & International Relations at the RUG. She specialises in European Union affairs and policy formation. This essay was submitted in January 2025 as part of the course on Theory of Modern History and IR.

AI Knows Nothing: A Collingwoodian Exploration of Machine-Generated Truth

Above my desk, there is a small, hand-painted portrait of R. G. Collingwood. He sits atop a pile of books on the European Union and International Relations Theory, and next to him is a cat figurine. Every day, as I sit down to write or study, he reminds me of the value of curiosity, hard work, and human touch.

While Collingwood’s portrait is a new addition to my set-up, what he represents is not. The rise of Artificial Intelligence (AI) in the last few years and the growing discussion surrounding it, have made it clear to me that being aware of these human qualities is becoming increasingly important.

Therefore, this essay is meant to explore (not explain) R. G. Collingwood's theory on question and answer and connect it to the added value of AI. It attempts to place the theory in a wider academic debate on modernity and the assumptions which Collingwood finds to be the foundation of questioning.

In his 1939 book An Autobiography, Collingwood explores the concept of truth and establishes that there is no such thing as assumptionless knowledge. Assumptions are the conscious ideas we have about something before it unfolds, and these ideas shape what we ask about them and, consequently, what we learn. But what and how we ask, as well as how questions are answered, has changed significantly in recent years as a result of the introduction of Artificial Intelligence and Large Language Models (LLMs).

They have introduced a conceptually different way of answering questions and “understanding” context which goes directly against Collingwood’s theory of Truth (Teubner et al., 2023; Shanahan, 2024). Yet this theory, as I argued in previous essays, is essential to opening up an incredibly wide and valuable thinking space which moves research beyond fact and instead focuses it on the act of researching.

Using AI to generate truth in our modern world has the potential to make it much more limiting, I fear. So if AI and its machine-generated truth are becoming so prevalent in our everyday lives, can we still say that Collingwood is relevant? Or, alternatively, is he more significant than ever?

Given that all questions are asked in a context in which we then create assumptions to explore the topic we investigate, I posit that the way AI handles questions is insufficient for talking about truth in modern society. Therefore, if we wish to approach an understanding of the role AI plays in our idea of truth today, we need to explore the context in which we as individuals exist, and how this is used and duplicated by AI.

This exploration is structured as follows: first I will briefly introduce Collingwood and attempt to explain his theory on question and answer as he developed it in An Autobiography. Then I will talk about AI and LLMs, to explain how they work and what this means for truth. The section after that will reflect on how applicable Collingwood is in our current age and thinking about machine-generated truth, before heading to the conclusion which finds that Collingwood is more relevant than in today’s society which increasingly relies on Artificial Intelligence for truth-generating.

The Theorist

Robin George Collingwood (1889-1943) was best known as a philosopher and archaeologist who worked on aesthetics and the philosophy of history as a subject and academic field (Leach, 2009; Inglis, 2011; Stanford Encyclopedia of Philosophy, 2020). Posthumous he became an influential figure in debates about the nature of social versus natural sciences. But even when alive, Collingwood was a fascinating figure.

He was made a fellow at Oxford while still taking his final examinations and became an authority figure in almost every field he tried his hand at, despite maintaining that many of these fields were more passion projects than anything else. Furthermore, he read in English, French, Spanish, Italian, German, Dutch, Latin, and Greek, and published on history, archaeology, philosophy, ethics, and aesthetics.

In the last decade of his life, Collingwood’s serious affinity for overworking himself (as well as his suffering from insomnia) caught up with him and took a serious toll on his health. However, this pushed him to rush the publication of more work, and the cycle only ended when he passed away in 1943 (Stanford Encyclopedia of Philosophy, 2020).

As mentioned, in An Autobiography, Collingwood poses that there is no such thing as assumptionless knowledge. He instead finds, as all knowledge takes the form of answers to questions, and all questions depend on assumptions, that all knowledge is inherently assumptive. His critique of realism states that there is such a thing as pure, assumptionless knowledge which we as people can access.

This goes beyond Kant’s critique, as he still maintained that assumptionless knowledge existed, but only God could access it (Stanford Encyclopedia of Philosophy, 2020).

Collingwood finds knowledge is inherently subjective and interpretative. This means that, if one does not know the context of an original question and can instead only rely on a completely contextless answer, the answer cannot be truthful. If, for example, I simply stated “the egg”, but did not elaborate on it, instead leaving you to fend off the mystery contained in the statement all by yourself, you would have no way of knowing if it was possibly true or false. Without the context of a question, knowledge is delegated to the unknowable and undesirable sphere of statements.

Answers can therefore be, at the same time, true and untrue. And so answers alone are not something you can object to. This might lead us to believe that in Collingwood's logic, there is no objective truth. This is however not the case. It is imperative that we separate the idea of knowledge being the same as truth. Thinking is fleeting, thought is permanent. And with historical and ontological investigation, we can retrace what people intended to ask. This forms the ontological foundation of Collingwood, but it is also often overlooked by critics, despite its crucial role in his overall theory.

Collingwood’s theory is well accompanied by Foucault’s regimes of truth, despite the understandable tendency to think that Foucault is Collingwood’s complete opposite, his theoretical antithesis of sort. However, I pose that Foucault is crucial to understanding the context of Truth which Collingwood laments over. Foucault confronts how, when, and why truth has been inscribed in the individual throughout the history of Western society, which in turn has given rise to a peculiar form of subjectivity built on the space of ‘interiority’ (Lorenzini, 2016).

Foucault (Lorenzini, 2016) made the context of an asker to be the desires and feelings (the field of thought) experienced by the individual, which they have to interpret and scrutinise as subjective data. Through this process, the truth of an asker can be discovered. He also talks about how it has been a historical development that the obligation to find the truth of ourselves in ourselves, as well as the manifest in through a discourse of avowal, emerged. He states that this has nothing but a series of techniques of power and of the self-inscribed within a regime of truth that digs in ourselves the very space in which it produces the truth we are asked to discover and manifest.

Foucault defines his regime of truth as “the types of discourse [a society] harbours and causes to function as true; the mechanisms and instances which enable one to distinguish true from false statements, how each is sanctioned the techniques and procedures which are valorised for obtaining truth; the status of those who are charged with saying what counts as true” (Foucault, 1977; translation provided by Lorenzini, 2016). A regime of truth thus refers to the well-known circularity and essential link between power and knowledge. In Lorenzini’s (2016) words, a regime of truth is the strategic field within which truth is produced and becomes a tactical element necessary for the functioning of several power relations within a given society.

When combining Collingwood’s theory and Foucault’s regimes of truth, everything we know becomes contextual. The questions we ask, as per Foucault, are dependent on personal circumstances and exist within a frame of human power relations. The answers, as per Collingwood, are fully dependent on the assumptions we make about the world and therefore also completely reliant on situation and context. As a consequence of using these theories, we can come to see that there is not one set reality, but rather that it is in constant conjecture with other people and events.

The Challenger

Artificial Intelligence has taken the world by storm since the public release of ChatGPT on November 30th, 2022 (Teubner et al., 2023).

In the discussion surrounding it, people talk about both AI and LLMs, and for philosophical clarity’s sake, it is important to first address the difference between the two, before explaining what they do and how they might interact with the aforementioned theories of Collingwood and Foucault. Generative AI in general, can create any “new” content ranging from pictures to text. Large Language Models are a specific type of generative AI which can only generate new text (Teubner et al., 2023; Shanahan, 2024). The following exploration focuses specifically on the function of LLMs in the creation of truth.

Large Language Models such as ChatGPT and Gemini (previously known as Google’s Bard) are learning algorithms that can process language. They function with the use of neutral transformer models. These models are capable of processing massive amounts of text, by treating them as data (transforming them into tokens) and using those to focus on different parts of the input. By combining large-scale transformer architectures with enormous amounts of textual training, LLMs have scaled up and can now understand, translate, predict, and generate text at a level comparable to that of people (Teubner et al., 2023; Shanahan, 2024).

And this was the intention; LLMs are designed to look and work like the human brain and are trained using any human-made text available online. Their datasets span a wide range, from news articles and books to Wikipedia entries and blogs. By consuming all this data, AI can be trained to recognise and generate similar texts, and that is why LLMs can respond in such a humanlike fashion (Teubner et al., 2023; Shanahan, 2024).

However, there are downsides to using LLMs. Shortly after its release, ChatGPT was banned in multiple locations as its answers had “a high rate of being incorrect”, and became a contested topic in universities because it is terrible at citing and generally creates a high potential for plagiarism (Teubner et al., 2023). Teubner et al. (2023) state that the key challenge to generative LLMs is determining what is a “good” text, as this is dependent on the type of text which needs to be generated. They reflect on how devastating it is for a company’s net-worth if their LLM gives wrong answers to questions.

That is why, they state, training AI still involves a lot of human feedback which corrects and steers AI to align with human preferences and context (Shanahan, 2024).

Because LLMs are completely trained using human material they can respond like a real person would. But this is also dangerous, Shanahan (2024) notes. As LLMs become more powerful and integrated into our everyday lives and work, it becomes increasingly tempting to anthropomorphise them and overestimate their capabilities. After all, the systems are trained on human thinking and coded to respond in a similar matter, so it feels only natural to refer to them as people.

But they are not.

LLMs and other generative AI are purely robotic systems, which have been coded to recognise patterns and fill in the most likely response. These systems do not “understand”, “believe”, or “know” anything.

They are predictors of human behaviour.

Does this mean that LLMs are not “aware” of the context, despite answering our contextual answers? And if this is the case, following Collingwood’s logic, can answers generated by LLMs be true?

The Reflections

The question is, when Collingwood posed that there is no such thing as assumptionless knowledge as all knowledge takes the form of answers to questions which are in turn all founded on assumptions we people make about the world around us, is this still applicable when the answers are no longer human-made?

As just mentioned, while we build systems whose capabilities more and more resemble those of humans, it becomes increasingly tempting to anthropomorphise those systems, even though they work in ways that are fundamentally different from the way humans work. Humans have evolved to co-exist, and human culture has evolved to facilitate this co-existence, which ensures a degree of mutual understanding. Human language is an aspect of our collective behaviour, and it only makes sense in the wider context of the human social activity of which it forms a part. Language-involving activities make sense because we inhabit a world that we share with other language users (Shanahan, 2024).

But it is a serious mistake to unreflectingly apply to AI systems the same intuitions that we deploy in our dealings with each other, especially when those systems are so profoundly different from humans in their underlying operation (Shanahan, 2024). AI (and consequently LLMs) is generative, so it means we can ask them questions. But only very specific questions. They mathematically model the statistical distribution of tokens in the vast public corpus of humangenerated text.

Shanahan (2024, 70) explains it well: we might give an LLM the prompt “Who was the first person to walk on the moon?” And it is likely to answer (correctly) “Neil Armstrong”. While on one level, yes, the model gave the correct answer to a question, but what is really happening is that we asked the model “Given the statistical distribution of words in the public corpus, what words are most likely to follow the sequence “Who was the first person to walk on the moon?””, and the most frequently flagged answer would be “Neil Armstrong”. Give the model any other prompt, and the same process will happen.

Where a person would be able to distinguish between different questions and the spheres of each of them, an LLM only sees numbers. The system lacks the capacity to engage in self-aware inquiry as it does not have access to the external reality in which people exist. This means that for AI, there is no context, the way that there is for us. LLMs only function on surface-level language patterns and thus cannot reflect on why we ask questions nor shape their answers according to a specific context. It is, after all, a tool, similar to an encyclopedia (Shanahan, 2024).

One could argue that as these models are trained with man-made sources, they are inherently steeped in the external reality we inhabit. Maybe even more so than we are, as AI can access almost all digital sources at any time, while people cannot possibly know everything.

But Shanahan (2024) argues that the external reality goes beyond mere facts. It also includes the interpersonal contexts; our awareness of other people, ourselves, and sources of information available to us. This is something AI can impossibly access. LLMs are “aware” of this fact also, and this is what I refer to when talking about LLMs not having access to the external reality of people.

Furthermore, while people understand the concept of truth and have a sense of it as relational, LLMs consistently fail to grasp the dynamic interplay between question and answer.

The questions we ask AI are not even real questions, but rather prompts, which, according to Shanahan (2024) there is an art to asking. Finally, AI never answers our prompts. Not really. It generates probable continuations of the words (data) we fed it. It does not engage with us and any sense we get of having a conversation or discussion with the interface is really just its design.

This means that when we analyse LLMs and AI-generated answers with a Collingwoodian definition of truth, the system becomes inherently flawed and meaningless. Everything that Collingwood deemed essential to creating truth and understanding the world, are things that LLMs cannot do.

It cannot reflect on questions which leaves it unable to create complete truth, it cannot see the relationship between question and answer, and cannot even really answer because it generates statements of fact instead of answers to questions.

So, if Collingwood poses that context is crucial to the formation of an answer and Foucault states that it is also crucial to questions, but LLMs operate and generate information outside of human context, are Collingwood and Foucault still relevant to the question of questions in today’s technological society?

According to Collingwood, the truth of an answer is dependent on the presupposition it was meant to answer. And to find its relative truth, we have to establish what the original question asked. Following this logic, the true value of (academic) research is not in the answers it finds, but in the questions it asks.

This is an incredibly liberating position to take with regard to research. It frees us of the banality of right and wrong and allows us to focus on the actual act of research. With answers in a secondary position compared to questions, research can focus on the purpose of a discipline instead of the processes within it. This encourages critical reflection on the questions which are asked, and in turn opens up a wide, valuable, and critical thinking space.

LLMs on the other hand, with their efficient and increasingly accurate text generation, may shift us away from writing and researching, to instead focus on the content of the ideas which are communicated. (Teubner et al., 2023). AI may actually be leading us back to the limited sphere of facts. Yet, as I established, LLMs are inherently unable to interact with the context of our questions and thus cannot deal in truths. All AI’s answers and “facts” are mere statements.

If we were therefore to allow machine-generated truth to have a bigger role in research and societal knowledge production, I think that our world and worldview would get much smaller. Collingwood welcomes the richness of exploration in research. His work encourages us to wonder, ponder, and reflect. When we use AI for research and let machine-generated truth dictate what is true or false, in my opinion, we are stripped of opportunity and whimsy.

But is there a way in which we can use LLMs to still approximate our own Collingwoodian truth? This was a question I played with while writing the majority of this exploration and I think it is worthwhile to reflect on it here also.

AI and LLMs do not, technically speaking, have beliefs, or, as they generate statistically likely sequences, know anything. This raises the additional question, is knowing necessary for answering questions truthfully? Is believing? According to Collingwood and Foucault, the context makes an answer truthful or not. What therefore matters, when answering a question, is the degree of philosophical engagement, which we, people, are capable of because of our communicative intent (Shanahan, 2024).

This means that we are aware of the context of those who ask the question and the question itself, our context of knowledge, and the context of information available to us. This leads me to believe that “knowing” and “believing” are necessary qualities for answering questions and generating truth. But not just “knowing” or “believing”, I also think that “understanding” is crucial to properly generating truth, as this is the key to reshaping the context into a coherent answer.

However, LLMs, which operate on superficial levels only, are incapable of this. They, as purely statistical models, do not “know”, “believe”, or “understand” anything. Recognising patterns is not enough to deal with the intricate and varied human nature and history of language which informs our context, as per Foucault, which in turn informs our questions. This further strengthens me in my conviction that LLMs are incapable of answering questions in a way that we can understand as “truthful”.

At the same time, it could be argued that because LLMs are trained with exclusively humanmade sources (for now), so inherently they possess a human-like quality. I am aware of that, and so, I went to the source. Annex I shows a “conversation” between ChatGPT and me on the topic of Collingwood and AI.

I acknowledge that there is a paradox in arguing that AI is incapable of telling the truth, and then asking it for its opinion on that statement. But in that “conversation”, I was made hyper-aware that I was interrogating an algorithm. ChatGPT and other LLMs like it are trained using human-made sources and designed to respond like a person would, but it is pliable.

Shanahan (2024) mentions that AI is a tool. And it behaves like it. It is there to serve whomever is asking it a question. If you ask it to change anything, it will do so immediately. But what is really important to take away from the “conversation” I had with ChatGPT, is its explanation of how it creates “answers”. I mentioned it earlier, but LLMs do not answer, they predict the next most likely sequence of words. They can do so because they have processed so many human sources, but they are still fundamentally different. A prediction is truthful in the way a human-made answer is.

If we treat it like that, I think that it can help us in achieving truth. For this, we must understand that, as I have argued, AI does not generate truth or answers. It purely creates statements, that, as I mentioned in the beginning, are unknowable and undesirable. But what if we then treat these statements as our context?

What if we take what interfaces such as ChatGPT and Gemini generate based on the entire human body of knowledge, and use that to base our new questions on? I think that when we see LLMs not as the final destination but instead as the starting point, they can still be used for generating truth in a way that Collingwood would approve of.

When working with AI, what is important is the question you ask, although, really, it is about how you ask the question. I established that AI does not register context, and only deals with text when it is reshaped into numerical data it can analyse. Therefore it is important to remind ourselves of what LLMs are, and what they are capable of. And it is even more important not to assign them human qualities.

Based on this, I do not think that we should be including AI-generated answers in our domain of truth, because as Collingwood found, truth is completely dependent on context (question), and this is the one thing AI cannot account for.

Of course, with the immense popularity of AI, the ever-growing online data-base which is available to it, and its design which allows it to constantly enhance itself, means that while AI is currently unable to access external reality and therefore in a Collingwoodian perspective unable to tell the truth, this might change in the future. And before this happens we must ask ourselves if we want to live in a world of machine-generated truth.

That is why it is so important to continue to remind ourselves of what AI is and what it can do. But more importantly, we must continue to remind ourselves of what truth is and what it takes to generate it.

Whatever conclusion we draw from these questions, I am convinced Collingwood (and by extension, Foucault) remains essential to these philosophical explorations and that he therefore may be even more relevant today than he has ever been.

References:

Collingwood, R. G. (1939). An Autobiography. London and New York: Oxford University Press.

Collingwood, R. G. (2001). An essay on metaphysics. Oxford University Press.

Dharamsi, K., et al. (2018). Collingwood on philosophical methodology. Palgrave Macmillan. https://search.ebscohost.com/login.aspx?direct=true&scope=site&db=nlebk&db=nlabk&AN=2002415

D’Oro, G. (2003). Collingwood and the metaphysics of experience. Routledge. https://search.ebscohost.com/login.aspx?direct=true&scope=site&db=nlebk&db=nlabk&AN=92995

Inglis, F. (2011). History man: the life of R.G. Collingwood. Princeton Univ Press. http://site.ebrary.com/id/10478263

Leach, S. (2009). An Appreciation of R. G. Collingwood as an Archaeologist. Bulletin of the History of Archaeology, 19(1), 14–20. https://doi.org/10.5334/bha.19103

Lorenzini, D. (2016). Foucault, Regimes of Truth and the Making of the Subject. Foucault and the Making of Subjects, 63-75.

Robin George Collingwood (Stanford Encyclopedia of Philosophy). (2020). https://plato.stanford.edu/ENTRIES/collingwood/

Shanahan, M. (2024). Talking about Large Language Models. Communications of the ACM, 67(2), 68–79. https://doi.org/10.1145/3624724

Teubner, T., et al. (2023). Welcome to the Era of ChatGPT et al. Business & Information Systems Engineering, 65(2), 95–101. https://doi.org/10.1007/s12599-023-00795-x