Abstract

Where a relationship is heading often shows in how people interact before it shows in outcomes. This article reviews four bodies of evidence. First, observed interaction, even its first three minutes, predicts later relationship outcomes, and relational variables predict breakup better than personality does (Gottman & Levenson, 1992; Carrère & Gottman, 1999; Le et al., 2010). Second, brief samples of behaviour carry information about consequential outcomes (Ambady & Rosenthal, 1992; Ambady et al., 2002). Third, prediction has firm limits: change in relationship quality is largely unpredictable from self-reports, and automated alerts that fire falsely lead people to stop relying on them (Joel et al., 2020; Parasuraman & Riley, 1997; Dixon et al., 2007). Fourth, people misjudge automated judgements in both directions: they abandon algorithms after seeing them err, yet accept recommendations, explained ones included, even when they are wrong (Dietvorst et al., 2015; Parasuraman & Manzey, 2010; Bansal et al., 2021). On this evidence, Mars draws opportunities, risks and key signals from your conversations as insight cards, each backed by an evidence chain to the original quotes and the source conversation, and shows only active insights by default. The evidence supports this design direction; it is not a test of Mars itself and does not show that any insight is correct.

Keywords: relationship dissolution · thin slices · interpersonal prediction · trust in automation · algorithm aversion · automation bias · relationship insights · evidence chain

1. Introduction

The way people interact carries early information about where a relationship is going. In a four-year study of married couples, the pattern of positive and negative behaviour in their interactions was associated with lower satisfaction, poorer health and a greater risk of separation (Gottman & Levenson, 1992). Outside the family, ratings of 10-second clips of surgeons’ voices during routine office visits distinguished surgeons who had malpractice claims from those who had none (Ambady et al., 2002). Opportunities and risks in working relationships plausibly leave similar traces in what people say to one another.

This article asks two questions. Can such signals be read early enough to matter, and how far can that reading be trusted? And if a machine surfaces them, how should it present its judgements so that people neither ignore them nor follow them blindly?

Method: this is a narrative literature review. We favoured peer-reviewed work accessible through a DOI, and preferred meta-analyses, longitudinal studies, systematic reviews and controlled experiments. We deliberately included work that limits the popular claims in this area: that relationships can be forecast, and that explaining an AI’s output makes people trust it appropriately. Every factual statement carries its source, and the full reference list with original links follows the article.

2. Interaction signals foreshadow where relationships go

Gottman and Levenson studied 73 married couples in 1983 and again in 1987. A single balance-theory variable derived from their interactions divided couples into regulated and nonregulated groups. Compared with regulated couples, nonregulated couples showed more negative and less positive emotional expression, more defensiveness, stubbornness and withdrawal, lower marital satisfaction, poorer health at the second assessment, and a greater risk of dissolution, including more frequent consideration of dissolution and actual separation (Gottman & Levenson, 1992). In a later study of 124 newlywed couples, coding only the first three minutes of a conflict discussion made it possible to predict marital outcome over the following six years, for husbands and wives alike (Carrère & Gottman, 1999).

A meta-analysis of nonmarital romantic relationships, drawing on 137 studies and 37,761 participants over 33 years, points the same way. Commitment, love, inclusion of the other in the self and dependence were among the strongest predictors of breakup; satisfaction, perceived alternatives and investments were modest predictors; social network support was a robust predictor; and personality measures had limited predictive utility (Le et al., 2010). What happens between two people, in other words, tells more about the relationship’s course than who each person is.

Building on this evidence, Mars looks for signals in the conversations themselves: as you chat or upload meeting transcripts, it picks out opportunities, risks and important signals for the people involved and shows them as insight cards on each person’s Insights tab (Gottman & Levenson, 1992; Le et al., 2010).

3. Brief samples carry information

Ambady and Rosenthal meta-analysed the accuracy of predictions of objective outcomes in social and clinical psychology from observations of expressive behaviour shorter than five minutes. Across 38 results the overall effect size was r = .39. Longer observation did not yield greater accuracy, predictions from clips under half a minute did not differ significantly from those based on four to five minutes, and accuracy did not depend on the channel rated (face, speech, body or tone of voice) or on whether the behaviour occurred in a laboratory or more naturally (Ambady & Rosenthal, 1992).

A field example comes from medicine. Coders blind to surgeons’ claims status rated two 10-second clips per surgeon, taken from the first and last minute of office visits with two different patients. Controlling for content, voice tones rated as more dominant identified surgeons with previous malpractice claims (odds ratio 2.74, 95% confidence interval 1.16–6.43) (Ambady et al., 2002).

These studies rated nonverbal and vocal behaviour, often by trained coders. What transfers to written conversation is the narrower point that a short excerpt can carry real information about a relationship, not the specific cues those studies measured.

This is why a Mars insight card keeps its source: expanding the evidence chain shows the original quotes behind it, and you can jump back to the source conversation, so the passing remark that prompted a judgement stays next to the judgement itself (Ambady & Rosenthal, 1992).

4. The limits: relationships resist prediction, and false alarms are costly

The most comprehensive test of relationship prediction to date applied machine learning (random forests) to 43 dyadic longitudinal datasets from 29 laboratories. The strongest relationship-specific predictors of relationship quality were perceived-partner commitment, appreciation, sexual satisfaction, perceived-partner satisfaction and conflict. These variables predicted up to 45% of the variance at baseline but only up to 18% at the end of each study, and change in relationship quality over the course of a study was largely unpredictable from any combination of self-report variables (Joel et al., 2020). A relationship’s current state can be read in part; where it will go is far harder to forecast.

Research on automation adds a second caution. Parasuraman and Riley distinguished misuse, overreliance that leads to monitoring failures and decision biases, from disuse, the neglect of automation, which is commonly caused by alarms that activate falsely; this often happens because the base rate of the condition to be detected is not considered when setting the trade-off between false alarms and misses (Parasuraman & Riley, 1997). In an experiment with 32 undergraduate students, false-alarm-prone automation hurt overall performance more than miss-prone automation and affected both compliance and reliance (Dixon et al., 2007).

Following this evidence, Mars presents insights as judgements with their grounds, not as forecasts of where a relationship will go. On a contact’s Insights tab they can be filtered by type (opportunity, risk, insight, need, budget, decision) and by status (active, retired, all). Only active insights show by default, so judgements that no longer hold stay out of the way, and when you choose Retired to look back, those cards carry a Retired tag (Joel et al., 2020; Parasuraman & Riley, 1997).

5. Trusting automated judgements: between aversion and overreliance

Automation is often problematic because people fail to rely on it appropriately. Lee and See’s review argues that trust guides reliance especially when complexity makes a complete understanding of the automation impractical, and its conceptual model includes the influence of display characteristics on trust (Lee & See, 2004). A systematic review of 127 studies in 101 papers published from 2002 to 2013 organised trust in automation into dispositional, situational and learned layers and proposed design recommendations that encourage appropriate trust (Hoff & Bashir, 2015).

Reliance can fail in the direction of aversion. In five studies, people who saw an algorithm forecast were less confident in it and less likely to choose it over an inferior human forecaster, even when they had seen it outperform the human, because they lost confidence in the algorithm faster after seeing it make the same mistake (Dietvorst et al., 2015). In three further studies, participants were considerably more likely to use an imperfect algorithm when they could modify its forecasts, even slightly, and performed better as a result (Dietvorst et al., 2018). A systematic review of 61 peer-reviewed articles grouped the proposed causes and remedies of algorithm aversion into five themes, among them expectations and expertise and decision autonomy (Burton et al., 2020).

Reliance can also fail in the opposite direction. Automation bias produces both omission and commission errors when decision aids are imperfect; it occurs in naive and expert participants alike and cannot be prevented by training or instructions (Parasuraman & Manzey, 2010). Explanations are not a simple remedy: in user studies on three datasets, explanations did not increase the gains from AI assistance and instead increased the chance that people accepted the AI’s recommendation regardless of its correctness (Bansal et al., 2021). In an experiment with 199 participants, cognitive forcing designs that made people engage with the explanation reduced overreliance compared with simple explanations, but people gave those designs the least favourable ratings (Buçinca et al., 2021).

On this evidence, Mars does not ask for trust in a bare conclusion. Each insight card is backed by an evidence chain: you can read the original quotes and jump back to the source conversation, so a judgement can be checked against what was actually said (Lee & See, 2004; Hoff & Bashir, 2015). Reviewing the cards remains your step, by type and by status (Dietvorst et al., 2018; Burton et al., 2020).

6. From evidence to design

Taken together, the findings set out what a tool that surfaces relationship signals should do: read signals in the interaction itself, keep each judgement tied to the words it came from, treat judgements as revisable rather than as forecasts, and leave the decision to the person. Mars follows each point as follows:

1. Interaction signals foreshadow relationship outcomes → insights drawn from everyday conversationAs you chat or upload meeting transcripts, Mars picks out opportunities, risks and important signals for the people involved and shows them as insight cards on that person’s Insights tab (Gottman & Levenson, 1992; Carrère & Gottman, 1999; Le et al., 2010).
2. Brief excerpts carry information → an evidence chain back to the original wordsEach insight card is backed by an evidence chain: expand it to read the original quotes, or jump back to the source conversation (Ambady & Rosenthal, 1992; Ambady et al., 2002).
3. Prediction is limited and false alarms cost reliance → types, statuses and retired judgementsInsights can be sorted by time and filtered by six types and by status; only active insights show by default, so judgements that no longer hold stay out of the way, and retired cards are tagged when you look back (Joel et al., 2020; Parasuraman & Riley, 1997; Dixon et al., 2007).
4. Appropriate reliance needs visible grounds and the person’s own judgement → grounds instead of a bare conclusionEvery judgement shows its grounds instead of a bare conclusion, and you review the cards by type and status; before a call or a coffee, a pre-meeting brief brings recent insights together with the person’s profile groups, open reminders and past conversations (Lee & See, 2004; Dietvorst et al., 2018; Burton et al., 2020).

7. Limitations

First, the studies cited here test how people interact, judge and rely on automation; none tests Mars. They support the design direction but say nothing about how accurate any individual Mars insight is, and this article makes no claim that Mars insights improve relationship outcomes; that would need its own evaluation.

Second, the samples are specific: married and newlywed couples (Gottman & Levenson, 1992; Carrère & Gottman, 1999), nonmarital romantic partners (Le et al., 2010), surgeons in office visits (Ambady et al., 2002), undergraduate students in a simulated monitoring task (Dixon et al., 2007) and participants in incentivized forecasting tasks (Dietvorst et al., 2015; Dietvorst et al., 2018). Extending them to business and social relationships, and to other cultures, calls for care. The surgeon study is associational and concerns claims history, not future claims (Ambady et al., 2002).

Third, the evidence on prediction is sobering. Self-reports predicted relationship quality only in part and its change hardly at all (Joel et al., 2020), and thin-slice accuracy was established for nonverbal and vocal channels rather than for written text (Ambady & Rosenthal, 1992). Signals in a conversation are therefore cues worth attention, not forecasts.

Fourth, showing evidence does not guarantee appropriate trust. Explanations can increase acceptance of wrong recommendations (Bansal et al., 2021), the designs that reduce overreliance are the least liked (Buçinca et al., 2021), and automation bias affects experts too (Parasuraman & Manzey, 2010). An evidence chain helps only when it is read. Mars therefore puts the original quotes one step away from each judgement, keeps judgements that no longer hold out of the default view, and leaves the decision with you.

8. Conclusion

Opportunities and risks in a relationship often leave early traces in ordinary interaction, and brief samples can carry them. Yet where a relationship will go remains hard to predict, false alarms erode reliance, and people misjudge automated judgements in both directions. Mars insight cards, their evidence chains to the original quotes and source conversation, and the default view of active insights are designed on this evidence: they surface signals with their grounds and leave the judgement to the person reading them.

References

Listed in APA 7th edition style, each with its original DOI link.

1. Ambady, N., & Rosenthal, R. (1992). Thin slices of expressive behavior as predictors of interpersonal consequences: A meta-analysis. Psychological Bulletin, 111(2), 256–274. https://doi.org/10.1037/0033-2909.111.2.256
2. Ambady, N., LaPlante, D., Nguyen, T., Rosenthal, R., Chaumeton, N., & Levinson, W. (2002). Surgeons’ tone of voice: A clue to malpractice history. Surgery, 132(1), 5–9. https://doi.org/10.1067/msy.2002.124733
3. Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 1–16. https://doi.org/10.1145/3411764.3445717
4. Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 1–21. https://doi.org/10.1145/3449287
5. Burton, J. W., Stein, M.-K., & Jensen, T. B. (2020). A systematic review of algorithm aversion in augmented decision making. Journal of Behavioral Decision Making, 33(2), 220–239. https://doi.org/10.1002/bdm.2155
6. Carrère, S., & Gottman, J. M. (1999). Predicting divorce among newlyweds from the first three minutes of a marital conflict discussion. Family Process, 38(3), 293–301. https://doi.org/10.1111/j.1545-5300.1999.00293.x
7. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
8. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155–1170. https://doi.org/10.1287/mnsc.2016.2643
9. Dixon, S. R., Wickens, C. D., & McCarley, J. S. (2007). On the independence of compliance and reliance: Are automation false alarms worse than misses?. Human Factors, 49(4), 564–572. https://doi.org/10.1518/001872007X215656
10. Gottman, J. M., & Levenson, R. W. (1992). Marital processes predictive of later dissolution: Behavior, physiology, and health. Journal of Personality and Social Psychology, 63(2), 221–233. https://doi.org/10.1037/0022-3514.63.2.221
11. Hoff, K. A., & Bashir, M. (2015). Trust in automation: Integrating empirical evidence on factors that influence trust. Human Factors, 57(3), 407–434. https://doi.org/10.1177/0018720814547570
12. Joel, S., Eastwick, P. W., Allison, C. J., Arriaga, X. B., Baker, Z. G., Bar-Kalifa, E., Bergeron, S., Birnbaum, G. E., Brock, R. L., Brumbaugh, C. C., Carmichael, C. L., Chen, S., Clarke, J., Cobb, R. J., Coolsen, M. K., Davis, J. L., de Jong, D. C., Debrot, A., DeHaas, E. C., . . . Wolf, S. T. (2020). Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies. Proceedings of the National Academy of Sciences, 117(32), 19061–19071. https://doi.org/10.1073/pnas.1917036117
13. Le, B., Dove, N. L., Agnew, C. R., Korn, M. S., & Mutso, A. A. (2010). Predicting nonmarital romantic relationship dissolution: A meta-analytic synthesis. Personal Relationships, 17(3), 377–390. https://doi.org/10.1111/j.1475-6811.2010.01285.x
14. Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
15. Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
16. Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886

Cite this article

Mars Research. (2026). Can opportunities and risks in a relationship be seen early?. MarsDataAI. https://marsdata.ai/research/signals-in-relationships