Indirected Reality
We are redesigning the Queue website. Take a look and let us know what you think.
August 2, 2026 Volume 24, issue 3
The Coding Philosopher
Indirected Reality
Language, truth, and LLMs
Jim Waldo
How can LLMs appear to know facts about the world when they are trained only on patterns in language rather than direct experience of reality? Hallucinations shouldn’t surprise us. The real mystery is how these systems get anything right at all.
---
Plato’s Cave. Jan Pieterszoon Saenredam. 1604
How much can we know about the world by simply examining language? We use language to convey what we know about the world, but can we find out things about the world just by looking at the structure and nature of language itself? This is a question that philosophers have been thinking about for the past century or more, starting with Frege and Russell and continuing through such thinkers as Wittgenstein and Kripke. While this discussion has stayed mostly within the realm of philosophy (sometimes spilling over into linguistics), it is suddenly relevant to computer science as we try to understand how the AI systems built on large language models work.
One of the constant problems with systems built around LLMs is the phenomenon of hallucination, which is really just a kind way of saying that these systems can get things wrong. There are lots of papers about how to minimize these hallucinations and lots of ways of constructing prompts that minimize the problem, but the fact is that from offering nonexistent court decisions to citing papers that have never appeared (sometimes by authors who don’t exist), hallucinations are part of working with these systems.
Lots of work has gone into trying to explain why hallucinations happen. But if you understand how these models work, a more appropriate question is, how do they ever get anything right? Large language models, as the name indicates, are just about language, not the world. How is it that programs trained on one thing (language) can tell us about something else (the world)? This reduces to a deeper philosophical question regarding the relationship between language and reality that has been debated since well before LLMs in particular and AI in general brought the same question in a different guise.
Just to review, LLMs are based on training using examples of language; the bigger the training set, the better. Training consists of building a probability space, populated by the tokens in the language, structured around the likelihood that the tokens will appear together. This space is then used to generate new sequences of tokens based on the context (that is, the set of tokens fed into the system as the prompt). Conditioned on the most likely next token, the prompt forms a vector through a high-dimensional space, giving both a location and a direction. The result sounds like something you would get talking to another person and often contains correct information as well. But understand that the model itself is built around the statistical co-occurrence of words or tokens in the training set, a concept that is much more connected to the structure of language than to the world itself.
Of course, this is just how the foundation model is created. After that, all kinds of reinforcement learning happens, along with the building of guardrails and other filters. The base, however, is built just on the statistics of language as drawn from as large a training set as can be assembled.
The statistical co-occurrence has much more to do with the syntax of a language than it does with the semantics. Where words occur in a sentence (and therefore how often they occur close together) is determined by the grammatical form of the language, not with anything that has to do with the world. Different languages have different syntactic structures (at least at the level we perceive), and so you would expect that the probability space of those languages would also be different (which it appears they are). But syntax isn’t about truth; that is the realm of semantics.
So, why do the outputs ever get anything right? One possibility is that LLMs are simply a complex form of crowdsourcing. Crowdsourcing involves asking a large number of people to give an answer, then taking the most common of the answers as correct. LLMs can be thought of as being trained on the answers to all of the questions that could have been asked to produce the training set. We do generally use language to say things that we at least think are true, so most examples of language, in fact, describe the world. An LLM trained on, say, only Harry Potter fan fiction would have a different set of responses than the ones we tend to use, which are trained on whatever the companies building the models can find to feed in.
It is also possible that the distinction between syntax and semantics is not as straightforward as we might like to think. While semantics is often characterized as the study of meaning, it isn’t clear what is meant by meaning (which is itself a bit ironic). One overly simple definition was given by Wittgenstein: “The meaning of a word is its use in the language” (Philosophical Investigations, §43). And the probability space of the occurrence of terms is a reasonable mathematical abstraction for the use of those terms.
If we were simply doing philosophy, we could argue about this to try and find the solution. But since we are also computer scientists, we could think of some experiments to see if we can confirm this idea. If the answer we get from LLM-based programs is really dependent on the co-occurrence of terms, then we might expect that hallucinations happen more often when there are more problematic correlations in the use—for example, when topics are obscure (not talked about as much, so not as much usage for data) or controversial (where there is less regularity in the usage patterns). There is some evidence that this is the case, but more tests are needed.
Another experiment would be to see if we get different answers to the same question when the question is asked in different languages. If LLM-based systems are really just a reflection of the co-occurrence of terms in the language, and given that languages have different syntax rules, we should see variation in the answers provided by LLMs in different languages. In fact, there are some indications that the answers will differ in different languages, but this is an experiment that is fraught with other questions.
For example, this would assume that languages have different syntactic structures. While this seems obvious, it flies in the face of the theories put forward by Chomsky and other linguists in the 1960s that claim there is variation on the surface of languages but a single universal grammar for all languages at the level of deep structure. This is still an open question in linguistic circles, although not as widely believed as it once was.
But we don’t have to solve that problem to wonder about this experiment. We would have to know what is meant by “the same question” in multiple languages; sameness across translation is a difficult problem. And then there is the worry of whether differences are caused by different worldviews, different training sets, or any of a host of other possible variables. The experiments may be suggestive, but they are hardly definitive.
Nor should we expect them to be. No matter how we think LLMs are connected to the world, we end up with at least a level of indirection between what is in the LLM and the world we are trying to think about since whatever the LLM tells us is mediated by language. And this brings us back to Plato, as do most questions in philosophy.
In Book VII of the Republic, Plato gives the allegory of the cave. Imagine, he says, that we have all been chained in a cave so that we can look in only one direction. Behind us is a fire, and the real world of people, objects, and other things occurs between those of us in chains and the fire. All we can see are the shadows that are cast by these objects, which we take to be the real world. But what we see is just a reflection of the truth.
In the same way, what we get from LLM-based systems are the shadows of the real world, reflected by language. But these reflections are not exact, so we need to be cautious about what we believe when these systems confidently tell us about the world. Perhaps we can break out of this cave by trying to train our models—not on language, but on the world itself; this is the sort of endeavor that some like Yann LeCun are attempting to build.
Even in these models, there is some level of abstraction and indirection caused by the model (as British statistician George Box said, “All models are wrong, but some are useful…”), so we are just replacing one level of abstraction with a different one. That replacement might be worth it, as an abstraction built more precisely for reflecting the world might be better at the job than human language. This may be an explanation for why LLM-based systems are so much better at coding than they are at some other tasks—the language used to train them is much more precise (and restricted) than general human language. But that is a topic for a different day.
No matter our approach, there will always be some level of indirection between the way in which we think about the world and the world itself. Even if we were not trying to represent things in digital hardware, we as humans view the world through our own levels of indirection. That was the problem Plato was addressing when he first talked about the cave—our human perceptions are just reflections of what he took to be a more real world that held the truth. This moves us into pure philosophy, where there are very few answers and many, many questions. If it hasn’t been solved since the time of Plato, we too are unlikely to solve it.
Related Material
General Semantics David K. Lewis <https://philpapers.org/rec/LEWGS>
Solomon’s Sword Beats Occam’s Razor Choosing your best hypothesis Stan Kelly-Bootle <https://queue.acm.org/detail.cfm?id=1365508>
AI: It’s All About Inference Now Model inference has become the critical driver for model performance. Michael Gschwind <https://queue.acm.org/detail.cfm?id=3733701>
KV the Apostate Faith-based computing versus the unnatural science Kode Vicious <https://queue.acm.org/detail.cfm?id=3806209>
Jim Waldo is a Gordon McKay Professor of the Practice of Computer Science in the John A. Paulson School of Engineering and Applied Sciences at Harvard, where he teaches courses in distributed systems and privacy. He is also the chief technology officer for SEAS, and a professor of technology policy in the Harvard Kennedy School, where he is associated with the Belfer Center, the Shorenstein Center, and the Carr Center. Prior to being the CTO at SEAS, Waldo was the first chief technology officer for Harvard University. He received his Ph.D. in Philosophy.
Originally published in Queue vol. 24, no. 3
---
More related articles:
Neil J. Gunther \- Titan Transients and LLM Scalability While training LLMs, OpenAI observed what it termed the compute-efficient frontier (CEF), which characterizes a lower bound on token-prediction error or test loss. This limit has raised concerns that it might constitute a universal constraint on the scalability of GenAI systems. This article presents a model of LLM computational dynamics, based on the universal scalability law (USL), that provides a framework for understanding the CEF. The USL defines bistable minima in the tokenized neural net landscape. The deeper minimum determines the lowest achievable loss of an LLM instance. Successively larger LLMs possess deeper global minima that align with the power-law slope of the CEF.
Erik Meijer \- In Code They Think; In Proof We Trust AI agents that use tools can be hijacked by prompt injection to exfiltrate sensitive data. Runtime defenses such as model alignment, output scanning, and content classifiers are fundamentally reactive: By the time they detect an attack, irreversible actions may already have been taken, and sophisticated encodings such as steganography, encryption, and chunking can evade any content-based check. We propose a preemptive alternative: Constrain the agent to express its plan as a Kotlin script, then statically verify the script against a security policy before any tool executes. The verification uses taint analysis, a well-established static analysis technique. A fact extractor walks the Kotlin AST to identify sources (sensitive data origins), sinks (network exit points), and data flows between them.
Brian Houck, Travis Lowdermilk, Cody Beyer, Steven Clarke, Benjamin V. Hanrahan \- The SPACE of AI As AI tools become increasingly embedded in software development workflows, there is a growing need to understand their true impact on developer productivity and experience. This article presents findings from a mixed-methods study examining how developers perceive AI’s influence across the dimensions of the SPACE framework: Satisfaction, Performance, Activity, Collaboration, and Efficiency. Drawing on survey responses from more than 500 developers and qualitative insights from interviews and observational studies, we find that AI is broadly adopted and widely seen as enhancing productivity, particularly for routine tasks. The benefits vary, however, depending on task complexity, individual usage patterns, and team-level adoption. Developers report increased efficiency and satisfaction, with less evidence of impact on collaboration.
Margaret-Anne Storey \- From Technical Debt to Cognitive and Intent Debt Generative AI is dramatically accelerating software development, allowing teams to generate and modify code faster than ever before. For decades, software engineering has focused on managing technical debt—how code structure and implementation make systems harder to change. But in the age of AI, technical debt might no longer be the most important constraint. This article argues that the real risks are shifting toward two less visible forms of debt: cognitive debt and intent debt. Cognitive debt is the erosion of shared understanding across a team where no one can confidently explain how a system works or predict the impact of a change.
---
---
© ACM, Inc. All Rights Reserved.