arXiv: AI as a hidden third — a study on romantic communication assisted by ChatGPT

Writing a breakup message has always been painful. Today, a growing number of people outsource that task to ChatGPT — and it is not merely a question of laziness. Researchers Skyler Wang and Isabella Luppi, in a paper published on June 13, 2026, systematically analyze what actually happens when AI becomes a mediator in romantic communication: who is speaking, and whose message is it?

The study is not concerned with AI romantic partners or chatbot companions — but with something subtler and more widespread: the use of language models as tools for writing, interpreting, and formulating messages between two real people. The data foundation is a multi-year corpus of vernacular discourse from platforms such as TikTok and Reddit, collected between 2023 and 2026.

When two become three — with only one knowing

The central theoretical contribution of the paper is the concept of the covert triad. The authors define it as a relationship that is phenomenally dyadic — between two people — but operationally triadic, since the invisible third participant, an AI model, is aware of the communication only from the side of the partner who uses it.

This information asymmetry is not a minor technical detail. It alters the very structure of the relationship: one partner communicates with the support of a system capable of analyzing tone, predicting reactions, and suggesting formulations, while the other partner believes they are receiving a direct, unfiltered expression. The covert triad is not an exception — the study documents its presence in the accounts of users who employ AI for apologies, clarifications, breakups, and the interpretation of received messages.

Articulation labor and the delegation of emotional expression

The second key concept the authors introduce is articulation labor. This is the expressive dimension of emotional labor — specifically: the conversion of an inner experience into language that another person can receive and understand.

The authors draw a distinction between two levels: feeling labor — what a person actually feels — and articulation — converting that feeling into a receivable message. The paper’s finding is that articulation labor is increasingly being delegated to AI, while the feeling itself remains exclusively with the human. The user thus knows what they want to say, but not how to say it — and that intervention is handed over to the model.

This division of labor has implications that go beyond practical convenience: when a person signs a message they did not formulate, the question arises of what exactly that message conveys and to whom it truly belongs.

Who is actually writing your message?

This is the question the study does not leave unanswered — but neither does it leave it uncomplicated. The authors document how users actively negotiate boundaries between their own role and the AI’s. Many describe themselves as the “final editor” or “the one who knows what they want but not how to say it” — suggesting that a sense of authorship persists, even when the formulation comes from outside.

But cultural resistance is real and visible in the corpus. A portion of users and observers express discomfort, a sense of deception, or a loss of trust when they discover that a message was written with AI assistance. The authors do not portray this resistance as outdated conservatism, but as active negotiation over what romantic authenticity means in the digital age.

Authenticity is being reconfigured — but with resistance

The fundamental conclusion Wang and Luppi offer is that authenticity in interpersonal communication is being reconfigured. Previously, authenticity was tied to linguistic authorship: a message is yours because you wrote those words. Now, according to the study’s findings, it is shifting toward emotional ownership: a message is yours if the feeling behind it comes from you, regardless of who formulated the sentences.

This shift is not an abstract philosophical argument — it is documented in real user accounts who explicitly articulate it as their own rationalization or as a line of defense against criticism. At the same time, the study records active cultural resistance to that redefinition: a body of voices claiming that AI mediation in romantic communication fundamentally undermines trust, regardless of the sender’s emotional intentions.

The study takes no normative stance. Instead, it offers a conceptual toolkit — the covert triad and articulation labor — that future research and public debate can use for a more precise conversation about what actually changes when AI enters between two people trying to communicate.