---
title: "The Signal Is Rare: What Millions of Text Messages Can Tell Us About How People Communicate"
url: "https://linguisticsnews.com/insight/the-signal-is-rare-what-millions-of-text-messages-can-tell-us-about-how-people-communicate/"
author: "Layne Weant"
published: "2026-09-25"
updated: "2026-09-25"
---

# The Signal Is Rare: What Millions of Text Messages Can Tell Us About How People Communicate

By the time someone types “I’m done,” the conversation has often been changing for a while. My main question was “what moved first?” My current breakpoint work keeps pointing to interaction structure before obvious breakup vocabulary: who sustains the exchange, who answers bids, who repairs conflict, and when that reciprocity stops.

I came to this question for a personal reason. I have been through two divorces. By the time I started building Grey Mirror, I had roughly 15 years of my own message history and about 2.3 million messages to work with. I could find the dates when the marriages officially ended. What I couldn't answer from memory was when the conversations started changing or if there was ever a point of no return. A breakup is an event. A conversational breakpoint is a process. If the interaction has a baseline, then the research problem is to find the earliest sustained deviation from that baseline.

I used that 2.3 million-message personal archive as the development and training ground for the earliest models: parsing messy exports, resolving speakers, building the first behavioral taxonomy, mining examples, hand-checking labels, boosting weak labels with human-added samples, and training different classifier heads. I don't treat those messages as a population sample, since they were the corpus for the training, minus the gold set.

### **The first signal to appear is usually the structure, not the vocabulary**

The strongest pattern in my current breakpoint work is not a magic word and it is not a single sentiment score. The first useful movement is often structural: the conversation stops being jointly maintained in the same way it was before.

I use reciprocity here as an interactional construct, not a raw 50/50 message count. It includes who restarts conversations, who asks follow-up questions, whether replies still answer the prior turn, whether affectionate bids receive a response, who attempts repair after conflict, and whether one person starts carrying longer runs of the exchange. A pair can keep sending thousands of messages while the structure underneath those messages becomes increasingly one-sided.

This is the part I consider a working finding rather than a universal law. I am not claiming that a drop in reciprocity predicts divorce for every couple, and I do not yet publish a universal lead time. I am saying that when a breakpoint becomes visible in a full history, the earliest interpretable change is often in how the two people are coordinating the conversation before the language becomes openly final.

### **Why this is a linguistics problem**

Conversation analysis has treated turn-taking as fundamental to interaction for decades. Sacks, Schegloff and Jefferson described conversation as locally managed and interactionally controlled, while later work on repair examined how speakers recover from problems in speaking, hearing and understanding. Those ideas were developed from spoken interaction, but private messaging gives us a strange new version of the same problem: turns can be split across five bubbles, delayed for hours, resumed after sleep, quoted, reacted to, or answered out of order.

That means a “turn” in texting isn't one message. The model has to consider bursts, reply relations, time gaps, topic continuity, speaker switches and adjacency.

Communication accommodation theory adds another layer. People adjust how they communicate with one another, and romantic partners can become more linguistically aligned over time. Prior research has found that language style matching can relate to relationship stability, and intensive longitudinal work on couples’ text messages has shown that linguistic alignment can change as relationships form. Separate research on more than a million Reddit posts found linguistic markers of an impending breakup months before the event. None of that means one textual feature can diagnose a relationship, but it does support the idea that relational change can leave measurable language traces before the explicit event.

### **The breakpoint is a sequence problem**

A full history gives you a distribution. So you can ask whether a person’s current behavior is ordinary for that relationship or whether it has changed meaningfully, and stayed long enough, to count as a real change.

Grey Mirror handles that by converting the history into time-ordered measurements. For each analysis window, the system can track response timing, initiation balance, question reciprocity, one-sided runs, unanswered affection, repair attempts, escalation, deflection, emotional momentum, turn-taking, topic movement and other observable interaction features. The important step is comparison against the relationship’s own prior baseline, not a generic rule that says a five-hour reply is “bad.”

A candidate breakpoint should survive persistence checks. One ugly night is an episode. A breakpoint is a sustained change across enough messages and enough related signals that the interaction has moved into a different regime. That distinction is why I am more interested in the first stable drift than the most dramatic message near the end.

### **The dataset disclaimers**

The current Wentropy Labs relationship-message research archive is now around 20 million messages, but that number needs a warning label. It is not one clean randomized cohort. It includes separate research subsets, uploaded histories, development data and benchmark cuts built to answer different questions. Some subsets can overlap. When I publish a statistic, I keep the cohort with it instead of adding every message total together for one giant study.

Two published benchmark cuts are useful examples. One signal cohort contained 4,600,611 messages across 29 de-duplicated relationship histories from 20 accounts. Under that study’s classifier rules, about 97.4% of messages carried no detectable emotional signal. That result matters for breakpoint work because it says the system cannot hunt for psychology in every “ok,” grocery-list item or scheduling text. Most of the archive is ordinary life.

A separate conflict-and-repair cohort contained 2,994,932 messages across 44 complete histories. At the cohort medians, escalation appeared about 36 times as often as repair, and only 5 of 44 histories had repair occurring more often than escalation. I do not use that as a breakup predictor because the cohort did not have complete relationship-outcome labels. It is a descriptive result about what happened inside those message histories.

### **The obvious signal I thought would have been more clear?**

Reply time is the signal everybody wants to turn into a relationship story. Someone used to answer in two minutes and now they answer in two hours, so it is easy to decide that the relationship is fading. I expected reply speed to give me a cleaner result than it did.

In one published benchmark comparison, the direction of slower replying split 9 to 13. There was no clean directional story I could defend, so I did not turn it into a headline. Timing still matters inside a person’s own history, especially when it changes sharply from baseline, but the aggregate result was that “slow reply equals detachment” is too simple.

That null result is one reason I prefer a multivariate breakpoint. A relationship can change while total volume stays high. Reply time can lengthen because of work, sleep, children or travel. Conflict can spike temporarily and recover. The more defensible signal is a coordinated shift across interaction structure.

### **How I built the technology from the message level upward**

The easiest way to misunderstand Grey Mirror is to imagine that I upload a giant transcript to a general chatbot and ask what happened. That is not the architecture. The system was built with the idea that none of the data ever touch a neural network; I'd get all my hard numbers and phrase counts and THEN give those to an LLM for an analysis….

| Layer | What it does | Why it matters for breakpoint research |
| --- | --- | --- |
| Parsing + normalization | Resolve timestamps, speakers, duplicates, reply structure and messy export formats. | A false speaker map or broken chronology can manufacture a fake turning point. |
| Deterministic metrics | Count turns, starts, gaps, questions, bursts, unanswered bids and other direct features. | These measurements exist before any narrative interpretation. |
| Specialized ML | Transformer/classifier heads detect speech acts and behavioral categories such as repair, escalation, deflection, validation and affection. | The models convert language into repeated, countable event families while retaining uncertainty. |
| Sequence analysis | Compare metric trajectories across windows and identify sustained deviations or inflection points. | This is where “what changed first?” becomes a time-series question. |
| Evidence layer | Link findings back to dated windows and source messages. | A breakpoint claim should be inspectable instead of being a black-box verdict. |

The earliest models were trained and stress-tested against my 2.3 million-message personal corpus. As the project expanded, I fine-tuned and evaluated transformer encoders and multi-head classifiers, added temporal context such as prior messages and response gaps, and built evidence retrieval so a measurement can be traced back to the thread. The current research archive is far larger, but the design principle has stayed the same: the model should measure before it explains.

### **Two divorces changed the question I was asking**

Going through two divorces gave me two obvious endpoint dates. In both cases, the official end was easy to locate after the fact. The more interesting part was the long runway before it, when the conversation was still active enough that no single message looked like “the big dramatic breakup talk.”

That is what made the personal archive useful as a research problem. I could ask all these questions: which measurable property of the interaction stopped behaving normally first? Did turn-taking become one-sided? Did repair disappear after conflict? Did affectionate bids stop getting answered? Did shared language begin falling apart? Did someone start restarting while the other still replied?

Those questions are not about assigning blame. They are about locating change. A relationship can have conflict for years and still repair. It can have low message volume and remain close. It can have fast replies and be miserable. The useful object is the within-relationship shift, especially when several independent measures start moving together.

### **The limit I care about most**

A breakpoint detector can tell me that the conversation changed. The archive is also self-selected. People who upload a relationship history are not a random sample of couples. Some of the research subsets overlap, outcome labels are incomplete, message histories omit face-to-face interaction, and model labels have error. Those limitations are not footnotes to hide. They define what the results are allowed to mean.

That is also consistent with the broader relationship literature. Communication behaviors are associated with later relationship quality and dissolution, but meta-analytic effects are modest. Language-style research has found meaningful associations with relationship stability, but no single language feature should be treated as destiny. I'm not trying to be a fortune teller or be sensational. I just really like working with data and was hoping for an accidental discovery. Lol

### **What I think the archive is showing**

The breakup itself is usually linguistically obvious. The ghost I was hunting is in the period before that...

My current working conclusion is that the first thing to move is often the interactional structure: reciprocity, repair and response to bids start drifting from the relationship’s own baseline before the vocabulary becomes unmistakably negative. The exact first mover can differ by history, which is why I am more interested in a breakpoint model than a universal “red flag” word list.

That is the part I want to keep studying as the archive grows. With enough longitudinal data, private messaging stops looking like a collection of dramatic quotes and starts looking like a record of coordination between two people. The end of a relationship may be a sentence. The breakpoint is a change in a system.

### **Methods and reporting boundaries**

- The ~20M-message figure describes the broader current research archive, not one homogeneous randomized cohort.
- The 2.3M-message personal archive was used for early model development and training; it is not used as a population benchmark.
- Published 4.6M-message and 2.99M-message benchmark cuts answer different questions and should not be added together as if they are disjoint.
- Breakpoint analysis measures observable changes in the thread. It does not establish private motive, causality, diagnosis or inevitable relationship outcome.
- Reply-time findings are treated cautiously because timing is heavily confounded by daily life and did not yield a clean aggregate direction in one published comparison.
- Real-user relationship data is self-selected, and any population-level claim requires separate validation and outcome labeling.

### **Selected references**

Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696-735. doi:10.2307/412243.

Schegloff, E. A., Jefferson, G., & Sacks, H. (1977). The preference for self-correction in the organization of repair in conversation. Language, 53(2), 361-382. doi:10.2307/413107.

Giles, H., Edwards, A. L., & Walther, J. B. (2023). Communication accommodation theory: Past accomplishments, current trends, and future prospects. Language Sciences, 99, 101571.

Ireland, M. E., Slatcher, R. B., Eastwick, P. W., Scissors, L. E., Finkel, E. J., & Pennebaker, J. W. (2011). Language style matching predicts relationship initiation and stability. Psychological Science, 22(1), 39-44. doi:10.1177/0956797610392928.

Seraj, S., Blackburn, K. G., & Pennebaker, J. W. (2021). Language left behind on social media exposes the emotional and cognitive costs of a romantic breakup. PNAS, 118(7).

Kanter, J. B., Lavner, J. A., Lannin, D. G., Hilgard, J., & Monk, J. K. (2022). Does couple communication predict later relationship quality and dissolution? A meta-analysis. Journal of Marriage and Family, 84(2), 533-551. doi:10.1111/jomf.12804.

**Turn-taking**  
“Turn-taking has been treated as a fundamental organization of conversation since the foundational work of Sacks, Schegloff and Jefferson.”  
[https://doi.org/10.2307/412243](https://doi.org/10.2307/412243)

**Conversational repair**  
“In conversation analysis, repair refers to the mechanisms people use when speaking, hearing, or understanding breaks down, not simply to apologies.”  
[https://doi.org/10.2307/413107](https://doi.org/10.2307/413107)

**Language-style matching**  
“Previous research has found that linguistic style matching between partners can be associated with relationship initiation and stability.”  
[https://pubmed.ncbi.nlm.nih.gov/21149854/](https://pubmed.ncbi.nlm.nih.gov/21149854/)

**Response timing**  
“Experimental research in spoken conversation has linked faster response times with greater perceived social connection, although asynchronous text messaging introduces very different timing constraints.”  
[https://pmc.ncbi.nlm.nih.gov/articles/PMC8794835/](https://pmc.ncbi.nlm.nih.gov/articles/PMC8794835/)

**Grey Mirror research**  
“The methodology, cohort definitions, aggregate findings, and supporting data for this project are documented in the Grey Mirror research archive.”  
[https://justlay.me/research/the-signal-is-rare](https://justlay.me/research/the-signal-is-rare)

**Grey Mirror product**  
“The system that grew out of this work became The Grey Mirror, a full-history conversation analysis pipeline built to analyze entire message histories rather than isolated screenshots.”  
[https://justlay.me/grey-mirror](https://justlay.me/grey-mirror)

---

Layne Weant is the founder of [Wentropy Labs](https://justlay.me) and creator of The Grey Mirror, a full-history relationship conversation analysis system. His work focuses on computational analysis of long-term digital conversations, including turn-taking, repair, conflict, reciprocity, linguistic alignment and changes in communication over time. He also develops Textimony.io for professional and legal review of large message histories.