Limits of Language | February 2026
i. Bringing awareness to the fact that language is not objective, but highly subjective.
ii. What is ‘pure language’, and how can it be represented.
iii. A meaningful way to represent the differences in our individual definitions for words to facilitate alignment.
Language both fascinates and frustrates me. This is probably made worse by the technological means we use to communicate today: language evolved in a signal dense environment (body language, tone, etc), and when we strip it of that context and apply it to new interfaces, like texting, we didn’t adapt it--the context it relies on for inference is ignored. Yet we continue to treat it as possessing the fidelity it has in its native context. As we increasingly rely on these interfaces to communicate with one another, it feels like we’ve begun to devalue the richness of the cues that an in-person context affords, which has slowly eroded our intuition and the attention required to read them. The default is to treat language like it’s a shared, objective code when, in reality, our definitions for words likely overlap just enough for us to function and coordinate: lots of information slips through the cracks, just not in a way that’s obvious enough to pay attention to.
I ran a small survey asking 104 people whether a thumbs-up emoji was passive aggressive: 49% said yes. To me, this is evidence that the miscommunication problem is worse than we realise.
One frame I find useful for understanding the types of communication we do is to separate it into two functions: coordination and connection. Each function requires a fundamentally different approach, which, to be solved, begins by being explicit about which function is actually being conducted. Coordination is impersonal; logistical. Connection is personal. Misunderstandings often arise from failing to make this distinction--a separation our existing interfaces don’t natively support.
Research in Artificial Intelligence (AI) and Brain-Computer Interfaces (BCI) anchor themselves to language as the output, compressing subjectivity back into the very system that filters it out. I think a more interesting way to conduct research in these areas is to design something with the goal of preserving the difference in what people mean (rather than the alternative, which is to normalise it away.) The questions that follow rest on the assumption that language is co-constructed.
Q1: Is language the problem in communication, or is it just the most visible thing we have and so it’s what we point to and blame for miscommunication?
Most products that offer solutions to miscommunication (writing and translation tools; BCI) still anchor their outputs to language. They use language to check that their solution works--words as the benchmark to see if these tools successfully capture what we really “mean.” For example, neuro-tech innovators consider their inventions successful if they can correctly pick up which word a subject is thinking of. But if language is structurally incapable of transmitting nuance--the delta between what different words--like “love” or “home” or “soon,” and even “no”-- mean to different people--then building technologies on top of it feels a bit like solving for the wrong problem. To what extent is language itself a bottleneck in these fields?
Q2: Where does neural decoding currently preserve subjectivity, and where does it flatten it?
BCI research that’s focused on neural decoding and shared embeddings tend to evaluate success against language or similar, externally defined labels. I get why they do this from a technical standpoint (there needs to be an anchor.) But this anchor may be the very thing that systemically removes some of the most interesting parts of the data: where people mean different things by the same word or where they feel something for which no word exists.
Q2½: If a participant speaks a word aloud vs just thinking it, how different is the result?
Compared to the results from just thinking the word, does speaking it aloud strengthen the signal? Does it appear completely differently? Or is there something else going on?
Q3: A new language! Could you design a representation system (a new ‘language’--however ‘language’ is defined) that surfaces the difference between subjective states, rather than collapsing it?
Rather than translating internal states back into words, could you build a symbolic system (like language, but not) where the difference between what two people mean becomes legible? For example, could you take the signals for two subjects thinking “love,” identify the overlap and explicitly surface the delta.
Currently, that difference is hidden--a fact that leads us to default to the idea that language is objective. This has all sorts of negative implications for our ability to communicate accurately and to preserve our relationships. What would a symbolic system that makes that delta explicit look like?
The gap between subjective states would itself become a unit of meaning, acting both as a bridge between people, and as a preserver of individuality that protects nuance from collapsing into the limitations that existing linguistic systems impose.
If a new language/alignment system like what I am describing existed, would we be able to deploy it in a way that requires a low enough cognitive load for it to be useful?
Q4: Is ‘pure language’ even theoretically possible?
Does language, by nature, only arise from a co-constructed process, and thus, is a ‘pure language’ something that simply cannot exist? Would it even be useful? Is the way we feel + the observable brain signals we emit, too context dependent to be consistent, and therefore legible? For example, if I am happy + cold + many other things, is it possible to isolate “happy” when I am happy + hot + many other things. Is there anything that makes doing this impossible or not useful?
Q5: Can physiological signals communicate emotional tone or intent without language as an intermediary?
From what I understand, brain regions like the amygdala regulate threat perception and defensiveness, and modulating these regions in a listener could reduce the psychological response that acts as a barrier in communication. Therefore, could we use that same modulation, informed by the speaker’s physiological state, in real time, to engage the corresponding region in the listener--essentially adding an additional layer of context about the intent of the speaker onto the listener that would lead to a more accurate exchange of information (not to transfer words--but the underlying vibes)?
If so, is sentiment encoded in the same region across all brains? i.e. could you achieve this by targeting the same area in the listener--or is it encoded in different areas of people’s brains, such that you would first need to map both brains to identify the corresponding region to engage the correct one?
Is any degree of sentiment transference even possible using brain data alone (without language as an input and/or without additional physiological data, like heart rate)--even for something as simple as positive vs negative valence? Or is context critical to the data, such that it’s always a requirement for successful transference?
Q6: Is misunderstanding necessary for connection? If so, what does that mean for the tools we should aim to build?
Deep down, I believe that relationships are forged through repair. Trust is built in ambiguity, and optimising for accuracy destroys the opportunity to build true connection, relationships, and meaning. If this is indeed the case, the goal shouldn’t be to eliminate miscommunication, but rather, to distinguish between the kind of communication that hurts and the kind that deepens. Is this the correct way to frame the problem while honouring the purpose miscommunication serves?