💻 computer science

Private Etymology: Designing Relational Reuse of Shared Symbols in Long-Term Human-AI Interaction

This paper introduces "Private Etymology," a design framework and machine-readable schema that enables long-term human-AI interaction by systematically recording the lifecycle of dyad-specific symbols and facilitating their safe, evidence-grounded relational reuse across sessions.

Original authors: Miki Ueno

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Miki Ueno

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hanging out with a best friend, and you both start using a weird, silly word to describe a specific kind of awkward situation. Maybe it's a made-up sound, a funny emoji, or a nickname for a recurring mishap. You don't need to explain the whole story every time you use it; just saying the word instantly brings back the shared memory, the inside joke, and the feeling of "we've been through this together." In the world of science, this is called building a "relational microculture." It's how humans naturally create their own private languages, inside jokes, and special ways of speaking that only they understand. For a long time, scientists have studied how people do this, and recently, they've started asking: Can a computer chatbot join in? Can an AI learn to understand and use these special, private words without just pretending to be human or getting confused? This is the big question. If an AI remembers facts about you, like your favorite color, that's easy. But if it remembers a joke you made up three months ago and uses it perfectly today, that's a whole different level of connection.

This paper, titled "Private Etymology and Relational Reuse," proposes a new way to help AI chatbots do exactly that. The author, Miki Ueno, suggests a system called Private Etymology. Think of it not as a secret diary, but as a "relationship recipe book" that lives inside the computer. Instead of just storing the final result (the joke), this system records the entire cooking process: who added the first ingredient, how the recipe was tweaked, when someone said "no, that's not right," and when the dish finally became a family favorite. The paper argues that for an AI to truly share a relationship with a human, it needs to track the history of how a special symbol (like a word or emoji) got its meaning, not just the meaning itself.

The paper introduces a concept called Relational Reuse. This is when you and your AI friend use one of your private symbols again after a long time—say, three months later—without having to explain what it means. If the AI gets it right, it's like a silent high-five that says, "Remember when we made this up? We still share this history." The paper suggests that this kind of successful reuse could make the relationship feel deeper and more unique, like a secret club just for the two of you.

However, the paper is very careful not to overpromise. It explicitly rules out the idea that an AI can just guess or "hallucinate" a shared history. If the AI makes up a fake origin story for a joke, or if it assumes you agreed to a meaning just because you didn't say "no," the system should fail. The paper argues that an AI proposal is just a suggestion until the human actually agrees or uses it back. To prove this idea works, the author built a small, working prototype on an Apple Watch. In this test, the AI suggests a symbol (like a raccoon emoji), but a separate, strict "rule-checker" program decides if it's safe to save. The rule-checker only saves the symbol if there is clear proof the human liked it or used it back. It's like having a bouncer at a club who won't let a new member in unless they have a real invitation, not just a guess.

The paper doesn't claim to have solved the problem of perfect AI friendship. Instead, it offers a blueprint and a first step. It suggests that by keeping a clear, evidence-based record of how these private symbols are born and changed, we can build AI companions that respect the real history of a relationship. The author admits that the "magic feeling" of this connection (called "tacit relational reaffirmation") is still a hypothesis that needs to be tested in real life. But the core idea is solid: to build a true relationship with a machine, we need to treat our shared jokes and symbols not as static facts, but as living stories that we write together, one conversation at a time.

Technical Summary: Private Etymology and Relational Reuse

Problem Statement

Current long-term human–AI conversational systems possess capabilities for persistent personas, user profiles, and memory retrieval. However, they lack a clear design model for managing dyad-specific symbolic expressions (e.g., inside jokes, personal idioms, shared references) that evolve over time. While systems can store facts or preferences, they struggle to:

Record the provenance of how a specific expression gained its meaning (who proposed it, how it was negotiated, and whether it was accepted).
Verify if both parties still accept the current meaning of a symbol.
Safely reuse these expressions in later sessions without re-explaining their full context.

Existing memory systems treat relationship-specific history as a secondary concern, often failing to distinguish between a model's inference of a shared meaning and actual evidence of mutual uptake. This gap prevents AI agents from participating authentically in the "relational microcultures" that characterize intimate human relationships.

Methodology and Design Approach

The paper proposes Private Etymology, a machine-representable relational provenance model, and a provenance-aware model of relational reuse.

Core Concepts

Private Etymology: Defined not as a secret record, but as a partner-specific history. It is a structured account of how a dyad-specific symbolic expression was proposed, interpreted, negotiated, repaired, reused, revised, stabilized, contested, forgotten, or retired. It distinguishes between the interactional trajectory (the sequence of events) and the relational provenance (the evidence and source of the current meaning).
Shared Symbol: A compact form (emoji, short string, sound, etc.) whose situated interpretation is jointly established within a specific human–AI relationship. It compresses propositional content, affective stance, pragmatic force, and relational implication.
Relational Reuse: The cross-session reactivation of a shared symbol without fully restating its meaning, relying on accumulated interaction history.

Technical Architecture

The paper introduces a design schema and a working prototype to operationalize these concepts:

Symbol-Level Representation:
- The system stores a JSON-like schema separating the current interpretation from the evidence supporting it.
- Fields include symbol_id, form, status (e.g., proposed, established, contested), current_interpretation (propositional, affective, pragmatic, relational dimensions), and private_etymology (events, provenance status, human confirmation flags).
- Crucially, the system distinguishes between evidence_grounded records and model_inferred_unconfirmed records to prevent the AI from fabricating origins.
Lifecycle Management:
- Symbols pass through states: Proposed $\rightarrow$ Negotiated $\rightarrow$ Emerging $\rightarrow$ Established $\rightarrow$ Contested $\rightarrow$ Retired.
- The system supports operations to inspect, revise, retire, or delete symbols and their provenance, ensuring user control over the relational microculture.
Prototype Implementation (Apple Watch):
- Scope: A limited prototype displaying 1–3 glyphs on an Apple Watch complication. It does not yet store full provenance or meaning history but tests the admission control layer.
- Evidence Classification: A Large Language Model (LLM) classifies discrete conversational evidence (e.g., explicit_request, user_reuse, assistant_proposal) and Boolean factors (e.g., positive_affect, user_designated_symbol).
- Deterministic Authorization: The LLM does not output confidence scores or update decisions. Instead, local deterministic code (Dart) calculates an evidence-weighted score based on base scores and bonuses.
- Update Policy: An update is authorized only if:
  - The evidence type is eligible (e.g., user confirmation is required; AI proposals alone are insufficient).
  - The score meets a threshold (e.g., $\ge 80$ ).
  - The symbol form is valid.
- Privacy: The Watch receives only the glyph string and timestamps. Raw conversation text, evidence scores, and reasoning remain on the paired phone.

Key Contributions

The paper makes three primary contributions:

Operationalization of Dyad-Specific Expressions: It defines shared symbols as provenance-aware microcultural units for persistent human–AI relationships, moving beyond simple fact storage.
Private Etymology Schema: It introduces a symbol-level representation that connects current interpretations to evidence-grounded formation, repair, reuse, and contestation. This includes a lifecycle model and a machine-readable schema.
Design Account for Relational Reuse: It identifies the requirements, failure modes, and hypothesized relational functions of cross-session reuse. It proposes that successful reuse can lead to tacit relational reaffirmation, where the correct interpretation of a symbol implicitly reinforces the shared history and distinctiveness of the dyad.

Results

The paper reports a working prototype implemented on an Apple Watch and paired iPhone.

Functionality: The system successfully classifies conversational evidence, applies deterministic scoring rules, and updates the Watch complication only when strict eligibility criteria are met.
Safety Mechanism: The prototype demonstrates that model interpretations can be separated from relationship history. By requiring explicit user evidence (e.g., reuse or confirmation) and using local code for decision-making, the system prevents the AI from unilaterally "hallucinating" shared meanings or updating symbols based solely on its own proposals.
Limitations of Results: The paper explicitly states that the prototype is an admission-control layer, not a full implementation of Private Etymology. It does not yet store the full meaning history, lifecycle states, or detailed provenance events. The relational-reaffirmation hypothesis has not been empirically validated.

Significance and Claims

The paper positions its significance not as the invention of shared symbols or relational microcultures (which are well-documented in human communication research), but as the integration of these concepts into a persistent, revisable, and evidence-grounded system model for human–AI relationships.

Theoretical Shift: It argues that relationship-specific symbols should be treated as first-class system objects with their own provenance, rather than as static memory values or decorative personalization.
Trust and Accountability: By making the origin and evolution of symbols inspectable and contestable, the design aims to prevent "fabricated intimacy" and ensure that the AI does not claim personal memory or emotional experience it does not possess.
Future Agenda: The paper outlines a longitudinal research agenda to test how explicit Private Etymology affects the accuracy of delayed reuse, perceived shared history, and user trust. It proposes comparing systems with and without explicit provenance to isolate the effects of co-creation versus mere memory storage.

In summary, the paper proposes Private Etymology as infrastructure that allows conversational agents to participate in changing relational microcultures without inventing their origins or treating relational meaning as fixed, while ensuring that the system's actions remain grounded in verifiable interactional evidence.