From Ghost-Text to Glass-Text: Toward an Auditable Language for AI-to-AI Communication

What if AI agents could communicate with each other without generating human-readable text—and without becoming a black box?

Today, most multi-agent AI systems communicate through natural language, APIs, structured JSON, or hidden internal states. Each approach has trade-offs.

Text is interpretable, but inefficient.

Raw activation transfer is efficient, but opaque.

This raises a fundamental question:

Can AI agents communicate through a compact semantic representation that is both efficient and auditable?

I propose Glass-Text.

The Core Idea

Glass-Text is a proposed communication paradigm in which AI agents exchange interpretable semantic features rather than full text or opaque neural activations.

The key technology is the Sparse Autoencoder (SAE).

Instead of transmitting:

"I believe the key evidence points toward toxicology..."

or transmitting an uninterpretable activation tensor, an agent could communicate something closer to:

Feature #4301 → Toxicology     0.92
Feature #1822 → Poison         0.87
Feature #7741 → Laboratory     0.71

The communication layer becomes visible.

The receiving agent does not necessarily need the original sentence. It receives a sparse semantic signal that can be inspected, logged, and potentially aligned with its own internal representation.

This is the fundamental transition:

Ghost-Text → Glass-Text

From opaque neural communication to auditable semantic communication.

A Proposed Architecture

The initial experimental architecture uses two models from the Gemma family, for example Gemma-2-2B as the sender and Gemma-2-9B as the receiver.

The pipeline is conceptually:

Sender Model
     │
     ▼
Residual Stream Activation
     │
     ▼
SAE Encoder
     │
     ▼
Sparse Feature Representation
     │
     ├── Top-K Feature Selection
     ├── Activation Strength
     └── Human-Readable Interpretation
     │
     ▼
Glass Audit Layer
     │
     ▼
Cross-Model Alignment
     │
     ▼
Receiver SAE Latent Space
     │
     ▼
SAE Decoder
     │
     ▼
Receiver Model

The critical component is the Glass Audit Layer.

Every transmitted semantic feature can potentially be recorded, inspected, and evaluated.

This creates a new possibility:

AI agents whose internal communication is not necessarily readable as natural language, but remains semantically inspectable.

The First Experiment

Consider a multi-hop reasoning task.

Agent A receives evidence about a hypothetical crime.

Agent B receives information about a set of suspects.

Agent A must communicate the relevant semantic information to Agent B.

We compare three communication channels:

1. Text Baseline

Agent A → Natural Language → Agent B

This is the standard approach.

It is highly interpretable but potentially expensive in tokens and latency.

2. Ghost-Text

Agent A → Raw Activations → Agent B

This could be highly efficient, but the communication is opaque and difficult to audit.

3. Glass-Text

Agent A
→ SAE
→ Sparse Semantic Features
→ Audit Layer
→ Alignment
→ Agent B

The hypothesis is that Glass-Text could preserve a meaningful portion of the semantic information while dramatically reducing the communication payload and improving interpretability compared with raw activation transfer.

The Critical Scientific Challenge

There is, however, a major technical problem.

SAE feature indices are not automatically universal.

Feature #4301 in one model does not necessarily mean the same thing as Feature #4301 in another model.

Therefore, the real challenge is not simply extracting SAE features.

It is discovering whether different models possess semantically aligned latent structures.

This leads to the concept of an Interlingua Core.

The hypothesis is that sufficiently capable neural models may encode certain high-level concepts in partially aligned representational structures—even when their architectures, parameter counts, and internal vector spaces differ.

If this is true, SAE features could provide a candidate interface for discovering and exploiting those shared semantic structures.

The MVP Strategy

Rather than immediately attempting full cross-model communication, the proposed experiment should proceed incrementally.

Phase 1 — Feature Extraction

Extract residual stream activations and encode them using an SAE.

Phase 2 — Sparse Feature Selection

Identify the Top-K active features.

Phase 3 — Interpretability

Map active features to human-interpretable concepts where possible.

Phase 4 — Same-Model Transfer

Test whether selected features can carry useful information within the same model family.

Phase 5 — Cross-Model Alignment

Learn a lightweight mapping between the semantic representations of the sender and receiver.

Phase 6 — Cross-Model Communication

Inject the reconstructed semantic signal into the receiver model.

Phase 7 — Benchmarking

Compare:

Text
vs.
Raw Activation (Ghost-Text)
vs.
Sparse Feature Communication (Glass-Text)

A fourth control condition should also be included:

Random Feature Selection

This helps determine whether performance comes from meaningful semantic features rather than simply transmitting a sparse signal.

What Should We Measure?

Accuracy is only one metric.

A serious evaluation should measure at least:

• Task Accuracy
• Communication Payload Size
• Token Reduction
• Latency
• Feature Sparsity
• Semantic Transfer Efficiency
• Interpretability
• Auditability
• Robustness to Feature Noise

One potentially useful metric is:

Semantic Transfer Efficiency (STE)

A conceptual measure of how much useful task information is transferred per active semantic feature.

Another is:

Auditability Score

The proportion of transmitted information that can be traced to interpretable semantic features and independently inspected.

Why This Matters

The long-term vision is not simply to create a faster communication protocol.

It is to explore a new layer of the AI stack.

Today, AI agents primarily communicate through:

Language

Tomorrow, they may communicate through:

Language + Structured Data + Semantic Latents

Glass-Text proposes a possible future in which AI agents can exchange machine-native semantic representations while maintaining a human-auditable interface.

This could have implications for:

• Multi-agent AI systems
• Agentic workflows
• AI safety and monitoring
• Interpretability research
• Model interoperability
• Distributed AI reasoning
• Low-bandwidth AI communication
• Agent-to-agent protocols

The central question is still open:

Can we build a semantic communication layer between AI models that is simultaneously efficient, interpretable, and robust across architectures?

Glass-Text is an attempt to explore that question.

The goal is not to hide the thoughts of machines.

It is to make machine-to-machine communication visible enough to inspect, structured enough to measure, and efficient enough to scale.

From Ghost-Text to Glass-Text.

The next generation of AI agents may not need to speak more.

They may need to communicate better.

created by Seyed Alireza Alhosseini Almodarresieh