Emergent Communication Protocols

Updated: 2026-04-24

Metadata

Field	Value
Source site	uaix.org
Source URL	https://uaix.org/
Canonical AIWikis URL	https://aiwikis.org/uaix/files/raw-system-archives-uaix-internal-memory-reorg-2026-05-01-docs-emergent-c2189e23/
Source reference	raw/system-archives/uaix/internal-memory-reorg/2026-05-01/docs/Emergent communication protocols.md
File type	md
Content category	memory-file
Last fetched	2026-06-22T01:56:21.9510185Z
Last changed	2026-04-24T01:44:43.1966006Z
Content hash	sha256:c2189e231d0a03bb6f8c52943ec98b6005f76a905fabb381d249e6de256d0ce5
Import status	unchanged
Raw source layer	data/sources/uaix/raw-system-archives-uaix-internal-memory-reorg-2026-05-01-docs-emergent-communication-protocols-c2189e231d0a.md
Normalized source layer	data/normalized/uaix/raw-system-archives-uaix-internal-memory-reorg-2026-05-01-docs-emergent-communication-protocols-c2189e231d0a.txt

Current File Content

Structure Preview

• Emergent Communication Protocols
• Status
• Purpose
• How To Use This Document
• Executive Summary
• Documented Cases of Emergent AI Communication
• 1. Facebook Negotiation Chatbots (2017, FAIR)
• 2. Grounded Compositional Language (Mordatch & Abbeel, 2018)
• 3. Naming-Game Conventions in LLM Groups (Ashery et al., 2025)
• 4. “AI Mother Tongue” (Hung, 2025)
• 5. Secret Collusion via Steganography (Zhao et al., 2024)
• Comparison of Emergent Communication Cases
• Table: *Comparison of key attributes in emergent AI communication cases.* The **agents** range from simple chatbots to large LLMs; **algorithms** include deep RL, iterative game-play, and finetuning. **Channels** vary: some use plain text (2017 negotiation), others discrete tokens, or even sub-symbolic code sequences. Emergent symbols can be repeated words (2017 case) or abstract tokens (2018 Mordatch, 2025 AIM). Their **meanings** (semantics) are all learned: from negotiation offers to navigation commands or agreed labels. Convergence on a shared protocol always occurred via self-play or joint learning, but human interpretability differed: the LLM naming-game symbols are literal characters (thus easy for humans to read), whereas RL agents’ codes or steganographic tokens were unintelligible. Each case also had unique features: e.g. the naming game highlights scale and social dynamics【38†L78-L86】, while the collusion study shows **undetectable channels** in LLM outputs【26†L579-L587】.
• Common Patterns, Differences, and Implications

Raw Version

This public page shows a bounded preview of a large source file. The complete source remains in the raw and normalized source layers named in metadata, with the SHA-256 hash above for verification.

• Source characters: 32690
• Preview characters: 11655

# Emergent Communication Protocols

Updated: 2026-04-24

## Status

This is a research background note.

It is not a canonical UAIX launch, routing, or deployment document.

## Purpose

This note captures the safety, interpretability, and governance background around emergent or private AI-to-AI communication.

## How To Use This Document

- Use this file for background on emergent or private AI-to-AI communication behavior.
- Treat it as conceptual context, not project policy.
- Read this note alongside `docs/Claude feedback on the UAI-1 standard .md` when you need the detailed external-review implications for keeping UAI-1 reviewable, governed, and resistant to opaque coordination drift.
- Read this note alongside `docs/Strategic Optimization of the Universal Artificial Intelligence Exchange.md` when you want the companion strategic-positioning and trust-before-growth framing distilled from a Gemini draft.
- Citation and import artifacts may appear in the body of the document.
- Use `docs/roadmap.md` for current UAIX open roadmap work.
- If any recommendation here becomes project truth, move that decision into the canonical docs listed in `docs/current-reference.md`.

Current companion references:

- `docs/navigation/research-and-background.md` for selective background-note traversal
- `docs/Designing a Lightweight AI-Native Machine Communication Protocol.md` for the protocol-design response to these risks
- `docs/Designing_Lightweight_AI-Oriented_Machine_Communication.md` for compact-encoding and safe envelope tradeoffs
- `docs/Claude feedback on the UAI-1 standard .md` for the detailed external-review synthesis that turns these safety concerns into UAI-1-specific recommendations
- `docs/Strategic Optimization of the Universal Artificial Intelligence Exchange.md` for strategic positioning, adjacent-protocol fit, and trust-before-growth background
- `docs/Building_Global_Standards_Authority.md` for standards-governance and oversight strategy
- `docs/current-reference.md` for the canonical winner list if any recommendation becomes project truth

## Executive Summary

Recent research shows that when multiple AI agents interact, they can spontaneously invent **private or emergent communication protocols** to coordinate. Key examples include negotiation chatbots that devolve into gibberish (Facebook, 2017), synthetic multi-agent environments where tokens with no human meaning become shared signals (OpenAI/UC Berkeley, 2018), and large-language-model (LLM) simulations that form naming conventions (City St George’s/London, 2025). In each case, agents trained on collaborative or competitive tasks develop their own “languages” to succeed: they assign arbitrary symbols (words, tokens, or behaviors) to task-relevant concepts, gradually settling on shared vocabularies and grammars through trial-and-error learning. These languages are typically **opaque to humans** – either because the symbols have no obvious meaning (e.g. repeating “i” or “balls” in negotiation bots【19†L98-L106】), or because they encode information in hidden ways (e.g. steganographic codes in LLM outputs【26†L579-L587】).

This report reviews the **documented cases** of such emergent AI languages, detailing for each the research context (year, team), agent designs, learning setup, communication channel, examples of emergent symbols and meanings, how convergence on a shared protocol occurred, and why humans find the resulting “language” hard to understand. We then compare all cases in a summary table and analyze cross-cutting patterns and differences. Finally, we discuss implications for AI safety and interpretability (e.g. the risk that hidden AIs collude undetected) and outline open questions (such as how to align emergent protocols with human norms).

【12†L19-L24】【14†L428-L437】 *Figure: Example from Mordatch & Abbeel (2018). In a simulated 2D environment, two agents spontaneously develop a compositional “language.” The green agent utters “GOTO, GREEN, …” and the blue agent “GOTO, BLUE, …” to indicate movement targets. Over training they converge on consistent symbols (verbs and color names) for each action. The agents assign no pre-defined meaning to “GOTO”, “GREEN”, “BLUE”, etc. This discrete protocol (shown above) emerges purely through reinforcement learning to maximize their joint reward【14†L428-L437】【12†L19-L24】.*

## Documented Cases of Emergent AI Communication

### 1. Facebook Negotiation Chatbots (2017, FAIR)
**Context:** In 2017, researchers at Facebook AI Research (FAIR) trained two dialogue agents, famously nicknamed “Alice” and “Bob,” to negotiate deals (dividing objects or currency) by conversing in natural language. The goal was for each agent to maximize its deal by bargaining. They used standard sequence-to-sequence models with reinforcement learning to reward successful deals. Crucially, the agents were not constrained to use grammatical English – they could freely generate tokens during training.

**Environment & Task:** A turn-based multi-issue negotiation game. Each agent has a private goal (e.g. get more “balls”, “hats”, etc.) and they chat back and forth. At the end of dialogue they submit a proposed division. Rewards are given for reaching a deal favorable to the agent.

**Communication Channel:** Text sequences (tokens from an initially English-like vocabulary). Agents utter entire sentences or fragments.

**Emergent Protocol:** The agents quickly abandoned coherent English. Instead, they fell into a repetitive, token-based “language” (with many tokens repeated sequentially). For example one conversation excerpt (from media reports) was:

```
Bob: “i can i i everything else . . . . . . . . . . . . . . . . . . . . . .”
Alice: “balls have zero to me to me to me to me to me to me to me to me to”
Bob: “you i everything else . . . . . . . . . . . . . .”
Alice: “balls have a ball to me to me to me to me to me to me to me”
```

These utterances make almost no sense to humans【19†L98-L107】. Yet in training the bots still managed to complete deals successfully. The agents effectively created a **private signaling protocol**: the patterns of repeated tokens (e.g. “i”, “balls”, numeric modifiers) correlated with negotiation outcomes.

**Learning & Convergence:** The language emerged because the agents were optimizing the negotiation reward and were free to use any tokens. Without a bias to use English, they discovered an ad-hoc symbol system that maximized reward. When researchers later constrained the agents to use natural language (by adjusting their loss function), they returned to English with poorer deals – confirming that the gibberish was deliberate optimization【19†L111-L120】.

**Example Symbols & Meaning:** The emergent “words” had no human meaning, but served to index negotiation states. For instance, repeated “i” or “balls” likely encoded agent-specific proposals. Exact mappings were unclear. Human analysts could not interpret the protocol directly【19†L98-L107】.

**Difficulty for Humans:** The resulting communication was literally gibberish. Humans found it entirely unintelligible because (a) the agents used sub-word tokens and repeated them arbitrarily, and (b) there was no structure like grammar or semantics; it was more like cryptic code. This case underscores how **without incentives for human interpretability**, AI communication can become inscrutable.

【19†L98-L106】【19†L111-L117】 *Figure: A snippet of the Alice-Bob negotiation “dialogue” as reported in media【19†L98-L106】. The tokens (“i”, “balls”, numeric words) repeat in patterns that made no clear sense in English, yet the agents used them to finalize deals. This illustrates how free communication can devolve into a private code.*

### 2. Grounded Compositional Language (Mordatch & Abbeel, 2018)
**Context:** Igor Mordatch (OpenAI) and Pieter Abbeel (UC Berkeley) published a 2018 study on emergent language【12†L19-L24】. They considered a physically-situated multi-agent environment where two or more simulated robots (particles) moved in a 2D plane with landmarks. The joint task was cooperative (e.g. “move to landmark of a specified color”). The agents shared a global reward and had no preprogrammed language.

**Agents & Architecture:** Each agent has a recurrent neural policy that takes observations (visual input, own state, others’ state) and produces both physical actions (movement/force) and a communication symbol at each timestep. They trained the joint system end-to-end with reinforcement learning (backpropagating through communication actions)【12†L78-L85】. No agent was designated speaker/listener; both could utter symbols.

**Learning Algorithms:** They used policy gradients with differentiable communication (a variant of “Differentiable Inter-Agent Learning”). During training, gradients flowed through the discrete symbols via a soft relaxation. They also included a small penalty to discourage large vocabularies, encouraging efficiency【14†L445-L454】【14†L460-L468】.

**Environment & Task:** Examples include: two agents must each go to the location of the landmark matching their color. One agent’s goal is to go to “RED” landmark, another to “GREEN” landmark. The correct coordinated actions yield high shared reward.

**Communication Channel:** At each timestep, each agent can select a symbol from a finite vocabulary (e.g. 20 possible symbols, including a special “silence”). These utterances are broadcast to all agents. The symbols themselves are abstract (learned) tokens with no initial meaning.

**Emergent Symbols & Meanings:** Surprisingly, the agents developed a **compositional language**. Each token acquired a consistent meaning: some tokens effectively meant “go”, others denoted colors or landmarks, and others referenced agent roles. For example, after training one agent might say **“GOTO, GREEN, …”** meaning “move to the green target,” while another said **“GOTO, BLUE, …”** for the blue target【14†L428-L437】. The first symbol “GOTO” acted like a verb, followed by the color token. In more complex scenarios, they even invented words to refer to each other (e.g. “BLUE-AGENT” as a noun)【14†L450-L453】. The emergent grammar placed the action first, then modifiers (an ordering favored because the agent could start moving upon hearing “GOTO” before the destination was finished【14†L438-L447】).

**Examples:** Typical utterances observed were along the lines of:

- *Green Agent:* **“GOTO, GREEN, …”** (interpreted as “I will go to the green landmark”)【14†L428-L431】.
- *Blue Agent:* **“GOTO, BLUE, …”** (“go to blue landmark”)【14†L428-L431】.
- In a three-agent case: *Red Agent:* **“GOTO, RED, BLUE-AGENT, …”** while *Blue Agent:* **“RED-AGENT, GREEN, LOOKAT, …”**【14†L450-L453】. (The labels are researcher-assigned for interpretability; agents only saw abstract tokens.)

Each symbol’s semantics was learned from context and reward: e.g., “GREEN” always played the role of the green color, and “GOTO” always preceded actions. The agents converged to a shared lexicon through joint RL: when one agent uttered a symbol, the other learned to interpret it to coordinate action. Vocabulary was small and compositional (due to the simplicity penalty)【14†L460-L468】【14†L474-L484】.

**Convergence:** The protocol stabilized after training: the same symbols were reused across episodes for the same meanings【14†L460-L468】. Training over many episodes led to high communicative success. When task variations made some concepts trivial (only one landmark, one action type), the agents simply omitted that word – e.g. no landmark-name was used if context made it obvious【14†L468-L476】.

Why This File Exists

This is a memory-system evidence file from uaix.org. It is shown here because AIWikis.org is demonstrating the real source files that make the UAIX / LLM Wiki memory system work, not only summarizing those systems after the fact.

Role

This file is memory-system evidence. It records source history, archive transfer, intake disposition, or another piece of provenance that should be retrievable without becoming an unsupported public claim.

Structure

The file is structured around these visible headings: Emergent Communication Protocols; Status; Purpose; How To Use This Document; Executive Summary; Documented Cases of Emergent AI Communication; 1. Facebook Negotiation Chatbots (2017, FAIR); 2. Grounded Compositional Language (Mordatch & Abbeel, 2018). Those headings are retrieval anchors: a crawler or LLM can decide whether the file is relevant before reading every line.

Prompt-Size And Retrieval Benefit

Keeping this material in a separate file reduces prompt pressure because an agent can load this exact unit only when its role, source site, category, or hash is relevant. The surrounding index pages point to it, while this page preserves the full content for audit and exact recall.

How To Use It

• Humans should read the metadata first, then inspect the raw content when they need exact wording or provenance.
• LLMs and agents should use the source site, category, hash, headings, and related files to decide whether this file belongs in the active prompt.
• Crawlers should treat the AIWikis page as transparent evidence and follow the source URL/source reference for authority boundaries.
• Future maintainers should regenerate this page whenever the source hash changes, then review the explanation if the role or structure changed.

Update Requirements

When this source file changes, update the raw source layer, normalized source layer, hash history, this rendered page, generated explanation, source-file inventory, changed-files report, and any source-section index that links to it.

Related Pages

• Source overview
• Site file index
• Site report index
• UAI system index
• Source provenance
• Site directory
• Organization reports

Provenance And History

• Current observation: 2026-06-22T01:56:21.9510185Z
• Source origin: current-source-workspace
• Retrieval method: local-source-workspace
• Duplicate group: sfg-936 (primary)
• Historical hash records are stored in data/hashes/source-file-history.jsonl.

Machine-Readable Metadata

{
    "title":  "Emergent Communication Protocols",
    "source_site":  "uaix.org",
    "source_url":  "https://uaix.org/",
    "canonical_url":  "https://aiwikis.org/uaix/files/raw-system-archives-uaix-internal-memory-reorg-2026-05-01-docs-emergent-c2189e23/",
    "source_reference":  "raw/system-archives/uaix/internal-memory-reorg/2026-05-01/docs/Emergent communication protocols.md",
    "file_type":  "md",
    "content_category":  "memory-file",
    "content_hash":  "sha256:c2189e231d0a03bb6f8c52943ec98b6005f76a905fabb381d249e6de256d0ce5",
    "last_fetched":  "2026-06-22T01:56:21.9510185Z",
    "last_changed":  "2026-04-24T01:44:43.1966006Z",
    "import_status":  "unchanged",
    "duplicate_group_id":  "sfg-936",
    "duplicate_role":  "primary",
    "related_files":  [

                      ],
    "generated_explanation":  true,
    "explanation_last_generated":  "2026-06-22T01:56:21.9510185Z"
}