---
title: "What Language Do Bilinguals Think In? Beneath Words"
description: "Bilinguals do not think in one fixed language. Underneath the inner voice is a layer of pure concepts that words only compress and label."
url: https://buildfirstbrain.com/journal/subvocalization-across-different-languages/
canonical: https://buildfirstbrain.com/journal/subvocalization-across-different-languages/
author: "Lawrence Arya"
authorUrl: https://www.linkedin.com/in/vibecoding/
published: 2026-06-05
updated: 2026-06-05
category: "Future & Language"
tags: ["bilingualism", "language of thought", "subvocalization", "first brain", "concept graphs"]
lang: en
---

# What Language Do Bilinguals Think In? Beneath Words

> **TL;DR** Bilinguals do not think in a single fixed language. They switch based on context, emotion, and topic, and a large part of thought happens below any language at all, in pure concepts that words only compress and label. The inner voice (subvocalization) is a surface layer, not the engine. The Build First Brain approach works at the deeper layer: it builds a language-independent concept graph, so understanding lives in connections rather than in the words of any one tongue.

Bilinguals do not think in one fixed language. They switch fluidly depending on context, emotion, who they are with, and what they are reasoning about, and counting in one language while cursing in another is completely normal. More striking: a large share of thinking happens beneath any language at all, in pure concepts that words only compress and label afterward. The inner voice you "hear" is subvocalization, a surface rendering, not the engine underneath. This is exactly where the Build First Brain approach operates: it builds understanding as a language-independent concept graph, so what you know lives in the connections between ideas rather than in the words of any single tongue. If you have ever felt a thought before you had words for it, in any language, you have already met the layer that matters.

## What language do bilinguals actually think in?

It varies, by design. Research on [multilingualism](https://en.wikipedia.org/wiki/Multilingualism) describes bilinguals selecting languages contextually: the language of the environment, of strong emotion, of a specific domain learned in one tongue. People often do arithmetic in the language they learned it in, dream in whichever fits the scene, and feel that swearing or intimacy lands harder in their first language. There is no single "thinking language" running underneath; there is a switchboard.

But the deeper answer is that much of thought is not in words at all. Anyone, monolingual or not, has had the experience of knowing what they mean before finding the words, of a thought arriving whole and then being slowly translated into a sentence. That gap, between having the thought and verbalizing it, is the clue that language is the output stage, not the place thinking happens.

## Is there a language of thought beneath the words?

This is a serious question in philosophy and cognitive science, not just a metaphor. The [language of thought hypothesis](https://en.wikipedia.org/wiki/Language_of_thought_hypothesis), associated with Jerry Fodor, proposes that thinking happens in a mental representational system, sometimes called mentalese, that is not any spoken language. The [Stanford Encyclopedia's treatment of the language of thought](https://plato.stanford.edu/entries/language-thought/) lays out the case that cognition operates over structured mental representations with their own combinatorial syntax, independent of English, Mandarin, or Arabic.

The debate is real and unsettled, the rival [linguistic relativity](https://en.wikipedia.org/wiki/Linguistic_relativity) tradition argues that the language you speak does shape how you perceive and categorize. Both can be partly true: there is a pre-linguistic conceptual layer, and the particular language you use also nudges attention and habit. What is not credible is the naive picture that you simply think in your native sentences. The structure runs deeper.

| Layer | What it is | Language-bound? | Role |
| --- | --- | --- | --- |
| Concept graph | Connected meanings and relations | No | Where understanding actually lives |
| Mentalese / language of thought | Structured mental representation | No (debated) | The reasoning substrate |
| Subvocalization (inner voice) | Silent rehearsal of speech | Yes | Surface rendering, working memory aid |
| Spoken or written output | Words shared with others | Yes | Compression for transmission |

## What is subvocalization, and why is it just the surface?

[Subvocalization](https://en.wikipedia.org/wiki/Subvocalization) is the silent inner speech, the faint internal articulation that accompanies reading and thinking, complete with tiny movements in the speech muscles. It feels like the medium of thought because it is the part you can hear. But it is better understood as a working-memory aid and a rendering layer: the mind dresses a concept in words so you can hold and manipulate it, the way a screen renders a file you could not otherwise see.

For bilinguals this is vivid, because the same underlying concept can be subvocalized in either language, and the choice barely touches the meaning. The thought "this is dangerous" exists as a concept first; whether your inner voice says it in one language or another is a costume change, not a different thought. That is the strongest everyday evidence that **language is a compression layer** over something more fundamental.

## Why does the First Brain think in concepts, not words?

Because understanding is structural, and structure is language-independent. Your **biological knowledge graph** stores meaning as concepts (nodes) and relationships (edges), the synapse-level mind map where each idea is a puzzle piece defined by what it connects to. A node like "leverage" or "entropy" is not the English word; it is a position in a web of relations that you could express in any language, or in none. We argued the structural side of this in [language is a topology](/journal/language-is-not-a-tool-it-is-a-topology/): the shape of your concept network matters more than the vocabulary draped over it.

This reframes what fluency and learning even are. **Speech is a low-bandwidth protocol**, a slow serial stream we use because we cannot yet transmit concept graphs directly, a limit explored in [will English always be the global language](/journal/bci-telepathy-and-the-end-of-english/). When you truly learn something, you are not memorizing sentences; you are wiring a concept and its connections into your graph. **First Brain before Second Brain** follows directly: storing words in an app is the shallowest possible encoding, while building the concept and its edges in your own memory is where comprehension lives. The method for building that language-independent structure is the core of Building Your First Brain, free for the first 1,000 readers.

It also explains why bilinguals often make unusually good thinkers and prompters. Holding two compression layers over one concept graph makes the graph itself more visible, because you can see the same idea wearing two different costumes, the advantage we examined in [does being bilingual help with AI](/journal/the-polyglots-secret-to-ai-prompting/). And it is why reading a translation can still deliver the original meaning when you read for the concepts rather than the words, the skill in [do translations lose meaning](/journal/how-to-read-a-translated-book-natively/).

## How do you train the concept layer instead of the word layer?

Aim under the words:

1. **Encode meaning, not phrasing.** When you learn something, ask what it connects to and what it contradicts, not how it was worded. A concept you can re-explain in your own words, or another language, is wired in; a phrase you can only repeat is not.
2. **Translate to test understanding.** If you cannot say it in simpler words or a second language, you learned the sentence, not the idea. Re-expression is a comprehension check.
3. **Notice the pre-verbal thought.** Catch the moment you know your meaning before the words arrive, and trust it; that is the concept layer reporting in.
4. **Build the graph in your head.** Subvocalized rehearsal helps working memory, but durable understanding is the connections wired into biological memory, available without any inner narration.

The honest limits keep this from becoming mysticism. Language is not merely decoration: it genuinely sharpens, stores, and sometimes shapes thought, and some concepts are hard to hold without words, so the goal is not to abandon language but to stop mistaking it for the thinking itself. And the strong "we think entirely in wordless concepts" claim is contested, real evidence supports both a pre-linguistic layer and a shaping role for language, so treat post-symbolic, fully language-free cognition as a direction the First Brain points toward, not a finished destination.

## Key takeaways: what language bilinguals think in

Bilinguals do not think in a single fixed language; they switch by context and emotion, and much of thought happens beneath language entirely, in a concept layer that words only compress and label. Subvocalization, the inner voice, is the surface rendering, not the engine, which is why the same concept can be voiced in either language without changing its meaning. The Build First Brain approach works at the deeper layer, building a language-independent concept graph where understanding lives in connections, not vocabulary. The honest limit: language genuinely shapes and stores thought too, and fully language-free cognition is a contested ideal rather than a settled fact, so train the concept layer without discarding the words.

## Frequently asked questions

### What language do bilinguals think in?

Bilinguals do not think in one fixed language; they switch based on context, emotion, topic, and who they are with, and much of thought happens beneath any language at all, in pure concepts. The inner voice is a surface rendering of those concepts, not the source. This is why the Build First Brain approach focuses on the deeper layer, building a language-independent concept graph where understanding lives in connections rather than words.

### Do people think in words or in concepts?

Both, at different layers. A great deal of thinking happens in a pre-linguistic conceptual layer, you often know your meaning before you find the words, and language then renders and sharpens it. The language of thought hypothesis argues for structured mental representations independent of any spoken language, while linguistic relativity shows the language you use also nudges attention. The deepest layer is conceptual; words are the compression over it.

### What is subvocalization?

Subvocalization is silent inner speech, the faint internal articulation that accompanies reading and thinking, often with tiny movements in the speech muscles. It feels like the medium of thought because it is the part you can hear, but it functions as a working-memory aid and a rendering layer rather than the engine of cognition. The same underlying concept can be subvocalized in any language without changing its meaning.

### Does the language you speak change how you think?

To a degree. Linguistic relativity research shows the language you speak influences how you categorize colors, track time, frame agency, and direct attention, so language does shape habits of thought. But it does not fully determine thinking, because a pre-linguistic conceptual layer underlies it. The realistic view is layered: concepts run deepest, and the particular language you use nudges and sharpens them rather than creating them from nothing.

### How does being bilingual affect thinking?

Holding more than one language over the same concepts can make the underlying structure more visible, because you see each idea expressed two ways and notice that the meaning is the costume-independent part. Bilinguals often switch languages by domain and emotion and can re-express ideas flexibly, which is a strong comprehension skill. It also tends to help with tasks like prompting AI, where seeing past any single phrasing to the concept is an advantage.

## Dive deeper in

- [Does language affect how we think? Language is a topology](/journal/language-is-not-a-tool-it-is-a-topology/)
- [Does being bilingual help with AI? Polyglot prompting](/journal/the-polyglots-secret-to-ai-prompting/)
- [Do translations lose meaning? Read past the words](/journal/how-to-read-a-translated-book-natively/)
- [Will English always be the global language? BCI telepathy](/journal/bci-telepathy-and-the-end-of-english/)

---

Source: https://buildfirstbrain.com/journal/subvocalization-across-different-languages/
Author: Lawrence Arya — https://www.linkedin.com/in/vibecoding/