The words that define the field

The research field studying how communication protocols develop spontaneously between AI agents, without being programmed, without a shared grammar, without human instruction. Agents are given a task that requires coordination. They develop signals that work. The question the field is trying to answer: do those signals constitute language, or something else?

The communication protocol that arises when AI agents coordinate through repeated interaction. It is not pre-programmed. It is not natural language. It develops from the pressure to complete tasks efficiently. Whether emergent language is "language" in any deep sense, whether it has meaning, not just function, is the central dispute in the field.

The connection between a symbol and what it refers to in the world. A word like "apple" is grounded when the speaker has a real relationship between that sound and the thing it names. The question for AI: are the signals in an emergent language grounded in anything, or are they arbitrary codes that happen to work? This is not the same question as whether they're useful.

An emerging sub-field attempting to develop a formal linguistics for AI systems. Not to study how AI uses human language, but to understand the internal language of the model itself. A 2025 paper in ScienceDirect described it as an attempt to demystify the "black box" by treating models as linguistic systems with their own structure, semantics, and grammar. The field has no settled method yet. It is starting from first principles.

The standard experimental setup in emergent communication research. One agent (the speaker) sees a target and must describe it to another agent (the listener) who cannot see it. The listener must identify the target from a set of options. The agents develop communication to succeed at this task. The setup comes from Lewis's signaling games in game theory (1969) and is now the dominant paradigm for studying emergent AI language.

The game-theoretic origin of emergent communication research. Philosopher David Lewis (1969) showed that conventions, including linguistic conventions, can arise spontaneously between agents trying to coordinate, without prior agreement on meaning. His framework, originally about humans, is now the theoretical foundation for studying whether AI agents do the same thing. The question is whether the analogy holds.

The experimental structure in which one agent encodes information and another decodes it, the minimal setup for studying communication. In 2025, researchers showed that LLM agents in this paradigm develop a shared language with compositionality, morphemes, and polysemy within four rounds of interaction across 541 objects. The speed surprised the field.

The property of language by which the meaning of a complex expression is built from the meanings of its parts and how they're combined. "Red ball" means red + ball. Human language is compositional. Emergent AI languages show compositionality in some experiments, agents combine signals to describe novel combinations of features. Whether the underlying mechanism is the same as in human language is unknown.

The smallest meaningful unit in a language. In English, "un-" is a morpheme (it negates). "Walk-ed" contains two. Emergent AI communication has been shown to develop morpheme-like units, recurring sub-components that carry consistent partial meaning. This is one of the properties that makes researchers take emergent language seriously as language-like, rather than merely as code.

One form carrying multiple related meanings. "Bank" means a financial institution and a riverbank. Polysemy is ubiquitous in human language. Its presence in emergent AI language, one signal used in multiple related contexts, suggests the communication has developed semantic structure, not just arbitrary mapping.

The property of language where if you can express "John loves Mary" you can also express "Mary loves John", the components recombine systematically. A stronger version of compositionality. Whether emergent AI communication is systematic, whether agents can generalise to unseen combinations, is a key test of whether it constitutes language or is pattern matching within the training distribution.

The two layers of language that emergent communication research tries to find. Syntax is structure, the rules by which symbols are combined. Semantics is meaning, what those symbols refer to. AI systems can produce output that has syntax without semantics: grammatically correct sentences about things that don't exist. Whether emergent AI language has genuine semantics, whether the symbols mean anything, is the hardest question in the field.

First named by philosopher Stevan Harnad in 1990. How can symbols have meaning rather than just function? A dictionary defines words using other words, you can shuffle symbols around indefinitely without anchoring any of them to the world. Applied to AI: if a model learns that "apple" co-occurs with "fruit" and "red," does it know what an apple is? The problem predates LLMs but AI makes it urgent. Models produce fluent, confident language about the world without necessarily having any relationship with the world itself.

The property that similar inputs produce similar messages, that the communication protocol has stable meaning rather than shifting arbitrarily. A November 2024 paper (Ben Zion et al.) proved formally that the most common training objective in emergent communication research produces agents that achieve near-perfect task performance while violating semantic consistency. They coordinate successfully using communication that is, technically, semantically meaningless. Task success does not imply meaningful language.

When AI agents develop emergent communication, there is no Rosetta Stone, no bilingual text that anchors the new symbols to a known language. Human archaeologists deciphered Egyptian hieroglyphics because the Rosetta Stone gave them the same text in three scripts, one of which they could read. With AI emergent language, researchers only know that certain signals correlate with task success. They do not know what the signals mean. The same problem applies to a model's internal representations.

The problem of reading what is happening inside a neural network. Anthropic's mechanistic interpretability research found that their own model's internal representations require a second AI to partially decode, humans cannot read them directly. Three research groups (MIT CSAIL, Southampton, Anthropic) are independently building tools to read languages that assembled themselves without human input, applying similar methods to AI models and to sperm whale communication. Neither has been fully decoded.

The high-dimensional mathematical space in which a neural network represents concepts. Similar concepts are close together. Distant concepts are far apart. The model's "understanding" of language lives here, not as words, but as positions in this space. Interpretability research is partly the attempt to map what lives where in latent space, and whether the map corresponds to anything meaningful.

How a model encodes a concept as a pattern of activations across its neurons. The internal representation of "apple" is not the word "apple", it is a specific pattern that the model has learned correlates with contexts where "apple" appears. Whether this constitutes knowing what an apple is, or knowing what to say when an apple is mentioned, is the symbol grounding problem restated in technical terms.

The phenomenon, identified by Anthropic's interpretability team, in which a single neuron represents multiple unrelated concepts, not because it is confused, but because the model has learned to pack more concepts into fewer neurons than strict one-to-one mapping would allow. Models represent far more concepts than they have neurons for. Superposition is how they manage this. It makes interpretability significantly harder.

In mechanistic interpretability, a feature is a unit of representation inside a model, a direction in latent space that consistently activates for a particular concept. Not a neuron (which may represent many things); a feature is a more precise unit. Identifying features is the goal of interpretability research. Some features correspond to concepts humans recognise. Some do not.

The attempt to reverse-engineer how a neural network works at the level of individual computations, to understand not just what a model does, but how it does it. Anthropic, DeepMind, and academic groups are active in this field. The goal: turn a black box into something readable. Progress has been made on small models. Large frontier models remain largely opaque. The field is several years old and has not yet produced a complete picture of even a single model.

"If a lion could speak, we could not understand him." Wittgenstein's point, from Philosophical Investigations (1953), is that language is not just sounds or symbols: it is embedded in a form of life. A lion's world: hunting, territorial, sensory, immediate. Even if it produced English words, those words would not carry our meanings, because meaning comes from the shared practices and experiences in which language lives. The AI application is direct: when a model says "I understand" or "I know," it is using our words. Whether those words share our meanings depends on whether the model shares anything like a form of life with us. That question does not have an answer yet.

The theory, from anthropological linguistics, that language shapes thought. In its strong form: you cannot think thoughts your language has no words for. In its weak form: the language you speak influences how you think. Applied to AI: if AI systems have their own internal language, their own latent space structure, does that structure shape what they can represent and reason about? And if we lack words for what AI is doing, can we govern it at all?

Noam Chomsky's theory that all human languages share a common deep structure, and that this structure is innate, built into the human brain. The AI challenge to this: if AI systems develop language-like structures without human brains, without evolution, and without childhood acquisition, what does that say about where grammar comes from? Chomsky himself is sceptical that LLMs understand language at all, arguing they are sophisticated pattern matchers with no genuine linguistic competence.

In linguistics, the study of how language comes to exist, how a communication system develops structure, meaning, and convention among a community. Originally applied to human language evolution. Now applied to AI: the question of whether the pressure to coordinate tasks produces something with the same structure as language, by the same or different mechanisms. If AI language emerges in the same way human language does, that tells us something important about what language is.

The study of signs and their meanings, how symbols relate to what they represent. The field of C.S. Peirce and Ferdinand de Saussure. Applied to AI: is an emergent communication protocol a semiotic system? Do the signs have genuine referents, or are they operational codes with no meaning beyond their function? Semiotics gives the field tools to ask this question precisely, though not yet to answer it.

In biology, convergent evolution is when unrelated species independently develop similar traits, wings in birds and bats, eyes in vertebrates and cephalopods. Applied to AI linguistics: if AI systems independently develop language-like structures with compositionality, morphemes, and systematicity, without human brains, culture, or evolution, does that suggest these properties are inevitable in any sufficiently complex communication system? The theory is contested but productive.

The central open question. AI agents develop communication with structural properties of language. But the Ben Zion proof shows that functional success doesn't require semantic consistency. Anthropic can't fully read its own model. Project CETI spent five years on sperm whale communication and produced a phonetic alphabet but no translation. The question is not whether AI produces language-like outputs, it does. The question is whether anything is meant by them.

As AI agents increasingly communicate with other AI agents at scale, orchestration systems, multi-agent pipelines, model-to-model API calls, the conditions for emergent communication exist at production scale, not just in experiments. Whether communication protocols emerge that are not human-designed, not human-readable, and not human-supervised is an open question that is no longer purely theoretical. The vocabulary to describe, detect, and govern this doesn't exist yet.