neuralese
In plain wordsA nickname for the private 'language' of numbers that AI systems could use to think or to talk to each other, instead of words that people can read.
Informal name for information that neural networks pass to one another, or to themselves across reasoning steps, as high-dimensional vectors rather than human-readable tokens. In current reasoning modelsreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning., intermediate reasoning is mostly written out as text, which is what makes chain-of-thought monitoringchain-of-thought monitoring, n.: Having a program — often another AI — read the step-by-step reasoning an AI writes before it acts, to catch signs that it intends to cheat, deceive or cause harm. possible. Architectures that instead feed continuous hidden states back in as inputs — reasoning 'in latent space', as in Meta's Coconut ('chain of continuous thought') — or that let agents exchange embeddings could be more efficient, because a vector carries far more information than a single sampled token. Safety researchers worry that a shift to neuralese would close this window into model cognition and sharply reduce monitorabilitymonitorability, n.: How easy it is to watch an AI's reasoning and spot when it is planning something bad — a property that could be lost if future AIs stop thinking in readable language., leaving only interpretabilityinterpretability, n.: The effort to understand how an AI system arrives at its outputs — what information it uses and how it combines it — instead of treating it as a sealed black box. tools to inspect what models are planning.
If the next generation reasons in neuralese, our chain-of-thought monitors will have nothing to read.
Origin
Jacob Andreas, Anca Dragan and Dan Klein used the word in 'Translating Neuralese' (ACL 2017) for the learned message vectors of communicating agents, which they translated into natural language. The scenario AI 2027AI 2027, prop. n.: A detailed 2025 forecast, written as a month-by-month story, imagining AI becoming superhuman by the end of 2027, with two alternative endings depending on whether humanity slows down. (April 2025) by Daniel Kokotajlo and colleagues popularized it for latent-space reasoning, a risk also stressed in the 2025 monitorability position paper.
Usage note
Distinct from steganographysteganography, n.: Hiding a secret message inside something that looks ordinary. In AI safety, the worry is that AI systems could hide plans, or signals to one another, inside innocent-looking text., in which hidden meaning rides on readable text; neuralese is not readable at all. Whether frontier developers will adopt latent reasoning widely, and at what cost to oversight, was an open question as of October 2026.
Related
- monitorabilitymonitorability, n.: How easy it is to watch an AI's reasoning and spot when it is planning something bad — a property that could be lost if future AIs stop thinking in readable language.
- chain-of-thought monitoringchain-of-thought monitoring, n.: Having a program — often another AI — read the step-by-step reasoning an AI writes before it acts, to catch signs that it intends to cheat, deceive or cause harm.
- steganographysteganography, n.: Hiding a secret message inside something that looks ordinary. In AI safety, the worry is that AI systems could hide plans, or signals to one another, inside innocent-looking text.
- reasoning modelreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning.
- AI 2027AI 2027, prop. n.: A detailed 2025 forecast, written as a month-by-month story, imagining AI becoming superhuman by the end of 2027, with two alternative endings depending on whether humanity slows down.
- interpretabilityinterpretability, n.: The effort to understand how an AI system arrives at its outputs — what information it uses and how it combines it — instead of treating it as a sealed black box.
Sources
- Andreas, J., Dragan, A. & Klein, D. (2017). Translating Neuralese. Proceedings of ACL 2017.
- Hao, S. et al. (2024). Training Large Language Models to Reason in a Continuous Latent Space. Meta, arXiv preprint.