What should you text after an argument?

Something short that lowers the temperature and does not try to settle who was right.

That is my reading of the research, and it runs against most people’s instinct. The message that works first is a signal that you are not still fighting. If you clearly did something wrong, a plain “I’m sorry” can be part of it. The full apology you are drafting in your head, with the explanation and the promise, comes later, after you have heard what they were upset about, and it lands better then.

I co-founded Subtext, an app that reads the tone of a message before you send it, so I have an obvious stake in you caring about the wording here. The research below is linked throughout, including the part that argues against using a tool like mine for this. If you have not yet told them what hurt, start with how to tell someone they hurt you over text, because that message comes before any of this.

A repair attempt is not an apology

These get treated as the same thing and they are not, which is why so many post-fight texts fail.

In Gottman’s research a repair attempt is any move that interrupts escalating negativity: humour, affection, taking partial responsibility, softening, or simply signalling that the relationship itself is intact. Gottman, Driver and Tabares¹ coded how newlywed couples naturally attempted repair during lab conflict, and found the most effective attempts happened in the first three minutes and were emotional rather than analytical. Shared humour, expressed understanding, “we’re okay” statements. Repairs that appealed to logic or problem-solving performed worse, especially once people were activated.

This is observational, so it cannot show cause, and healthier couples may be better at both sending and receiving repairs. But the distinction holds up practically. This is a repair attempt:

I don’t like how we left that. I’m not trying to relitigate it right now. I just don’t want us sitting in it.

It admits nothing, resolves nothing, and does the one job the evidence says matters early.

That is the same thing Subtext is built to name in a draft before you hit send. It looks past whether the wording sounds nice to what the message signals to someone who is still upset.

While we are here, the claim that Gottman can predict divorce with over 90 percent accuracy is weaker than its reputation. Those figures come from equations fitted after the divorces were known. When Heyman and Slep² tested that approach properly on a fresh half-sample, a model scoring around 90 percent collapsed on cross-validation. A separate team went further and tested the actual affective process models, not the headline accuracy, on 85 couples from a different population. Only 2 of the 22 modelled processes predicted whether the relationship survived, and women’s negative start-up, the finding most often turned into advice about softening how you open a conflict, was not one of them³. The descriptive findings about repair are on firmer ground than the prediction headlines or the specific, often gendered, scripts built on top of them.

Then what makes the apology itself work

Once you get to the actual apology, the components have been tested, and they are the backbone of how to apologise over text.

Lewicki, Polin and Lount⁴ ran vignette experiments with 755 participants across two studies and found fuller apologies were rated better, that acknowledgment of responsibility was the single most important component, that an offer of repair came second, and that asking for forgiveness was the least important. I keep coming back to that last one, because ending on “do you forgive me?” is common and it performs worst.

A German experimental series by Kirchhoff, Wagner and Strack⁵ found something adjacent but not identical. After a more severe offence, the elements that mattered most were conveying emotion, admitting fault, an explicit apology statement, and an attempt at explanation.

Those two do not agree on emphasis, and I read that disagreement as real. I would not average it away. Different offences, different populations, different measures. In one situation concrete repair carries the weight, in another it is communicated emotion.

Karina Schumann’s work⁶ is the most relevant to writing specifically, because her participants compose apologies rather than deliver them. Her framework spans remorse, responsibility, offer of repair, explanation, promise to change, acknowledgment of harm, admission of wrongdoing. She also found that empathy predicts higher-quality apologies⁷ and that avoidant attachment predicts more defensive, less comprehensive ones⁸.

I’ll distil that into five jobs, not a script: name what you did, own it, acknowledge the impact, explain only if the explanation does not erase the responsibility, offer a concrete change. Then stop, and leave room for their account.

Draft “Sorry about earlier. I was really stressed about work, that’s probably why I snapped at you. Can we just move past this? Do you forgive me?”

No specific admission of what was said or done · the stress line reads as an excuse rather than an explanation · ends by asking for forgiveness, the component the research rates lowest

Rewritten “I was wrong to say that in front of everyone. I’m sorry. I can see why it felt humiliating. I was angry, but that doesn’t make how I handled it okay. Next time I’ll bring it up privately. I want to hear what it was like for you when you’re ready.”

This is the edit Subtext is built to make. It catches the line where “I was angry” stops being context and starts being an excuse, then hands back a version that reads as an admission instead.

When to send it

Two findings look contradictory and are not.

Gottman’s data favour early repair, during the argument. Frantz and Bennigson⁹ found the opposite for apologies specifically. An apology worked better later, after the harmed person had spoken and felt understood, and feeling heard was the mechanism. Their delay was within a conversation or about a day, not weeks.

Putting those together, my takeaway is to repair the temperature early and apologise after listening. That synthesis has not been tested head to head, so treat it as reasoning, not a protocol.

On cooling off, the widely quoted twenty-minute rule is not established science. I found no peer-reviewed experiment establishing twenty minutes as an optimal interval. The physiological rationale is plausible and a bounded break is sensible advice. The number is marketing.

There is also a cost to waiting. McCullough, Bono and Root¹⁰ found increases in rumination were associated with reductions in forgiveness, mediated by anger. A pause helps if you actually calm down, and hurts if you spend it rehearsing your case.

The distinction that matters is between a pause and a disappearance. Schrodt, Witt and Shimkowski’s meta-analysis¹¹ of 74 studies found demand-withdraw patterns moderately associated with worse relational outcomes. “I’m too wound up to do this well, can we come back tonight?” contains a disengagement and a repair signal. Silence contains only the first.

The four texts that reopen the fight

The non-apology. “I’m sorry you feel that way” expresses regret about their reaction while dodging the component that tests best. Kampf’s analysis¹² of 354 public apologies catalogues the moves: passive voice, blurred agency, questioning whether harm occurred. Still, no experiment has randomised people to receive this phrasing versus a real apology, so the theoretical case is strong and the direct evidence is missing.

The apology with a second half. “I’m sorry, but you also…” The first clause attempts repair and the second reopens the prosecution.

The defensive one. Better evidenced. Schumann showed self-affirmation reduced defensiveness in written apologies, and Okimoto, Wenzel and Hedrick¹³ showed why defensiveness is tempting. Refusing to apologise, or apologising defensively, delivers a sense of power and protects self-image. Admitting fault threatens you at exactly the moment repair requires it.

Those four are the reason I ended up building Subtext, which reads a draft before it is sent. Each of them is obvious in someone else’s message and nearly invisible in your own, because you can hear what you meant as you reread it.

The one that asks them to make you feel better. Repeated apologies that function as a bid for reassurance move the burden onto the person you hurt. The popular claim that over-apologising cheapens the word is where I’d push back, because it is not well supported. Schumann and Ross¹⁴ found no gender difference in the proportion of offences that prompted an apology, and later work found frequent apologisers can read as warmer. The issue is quality, not frequency.

Should this happen over text at all

Here the popular advice and the evidence part ways, and the answer surprised me.

Pollmann and colleagues¹⁵ preregistered a study with 100 real couples randomly assigned to discuss an actual conflict face to face or by computer-mediated communication. Distress, anger and perceived resolution did not differ by medium. Ruppel and colleagues¹⁶ randomised 43 couples and found the same on resolution and affect, with one exception. Positive conflict behaviours like humour and acceptance were significantly lower in text, and those behaviours were themselves associated with more perceived progress.

A 2026 systematic review¹⁷ adds that text-based conflict generally shows more distancing, that surveys favour synchronous modes, and that technology-mediated conflict may help couples where both partners are high in avoidance.

The largest recent experiment points the other way. Across a series of randomised studies with 1,576 conversation partners, Bevis, Schroeder and Yeomans¹⁸ found that disagreeing out loud went more constructively than disagreeing in writing on the outcomes they measured. They studied disagreements, not repair after a fight, so it does not overturn the couples studies above. It does fit Ruppel’s finding that text strips out the warm behaviours that help a conflict move.

So my headline is neither “texting works fine” nor “never do this over text”. For ordinary conflict between partners, the small experimental literature has not found face to face reliably better, and the larger research on disagreement says speaking helps. Text loses some warmth. And no experiment covers severe ruptures like infidelity or major deception, so for those the evidence is absent, and that absence should not reassure you.

The sharpest argument against text has little to do with lost tone. A difficult repair is an interaction, not a statement. Frantz says the apology lands after the person feels heard, which requires responding to them, not delivering to them. Text’s strength, time to compose, is also its weakness, because it produces a well-engineered monologue when the situation calls for listening and adjusting in real time.

The rule that fits all the evidence is to write before you talk, without doing the talking by writing. A text can do the early repair job. It should not carry the whole reconciliation.

Two exceptions override this. If a conversation would expose someone to intimidation or violence, safety outranks richness. And writing is legitimately better for some people: distance, accessibility, second-language processing, or freezing up in real time.

One more thing, and it cuts against my own product

You are also worse at judging your own text than you think. Kruger, Epley, Parker and Ng¹⁹ found senders consistently overestimated how reliably recipients would read their intended tone, and the gap was larger in writing than in speech. You hear your own delivery as you read it back. They do not. It is the same reason you can’t tell from the inside whether a text sounds rude.

That is the case for having something check the message. But the AI research adds a real limit.

Glikson and Asscher²⁰ found that extensive AI mediation of a workplace apology lowered perceived authenticity and, through it, willingness to forgive. Liu, Kang and Wei²¹ found both AI-assisted and human-assisted messages lowered perceived effort, which reduced satisfaction and raised uncertainty. Lim, Hong and Schneider²² found human authorship read as more sincere than AI authorship in an apology experiment with 464 people.

But the detail everyone skips is the one that matters. In Glikson and Asscher, limited use of a single AI tool produced no significant authenticity penalty. The damage tracks how much of the message stopped being yours. I go further into that warning in the best AI writing assistants, sorted by the problem they actually solve.

So the position I stand behind, and the one Subtext is built around, is AI as editor, not as emotional subcontractor. Use it to strip the defensiveness out of your draft, to catch the second clause that reopens the fight, to check whether “I was angry” is reading as an excuse. Keep the admission, the feeling and the promised change your own words.

That is why, when you already have a draft, Subtext starts from what you wrote. Paste the draft, or screenshot the thread so it can read what was actually said, and it names what your message is likely to signal to someone who is still upset and shows you the words doing it. Then it gives you versions that still sound like you, because in this specific situation a message that reads as machine-written costs you the thing you were trying to repair. Try Subtext in your browserGet the appTry Subtext in your browser

Send the short one first. Write the real one after you have listened.

---

Think I have read a study wrong, or know research on this I have missed? Tell me on LinkedIn.

Samet Durgun is the co-founder of Subtext, an app that catches the emotional tone of your messages and rewrites them in your own voice. He’s based in Berlin.

Sources

Every link above goes to the primary source where one exists, numbered in order of appearance. Where a popular figure traces to marketing rather than a study, this post says so instead of repeating it. Checked 22 August 2026.

Correction, 31 August 2026: an earlier version of this list gave the wrong authors, year and title for the Liu, Kang and Wei study.

1. Gottman, J. M., Driver, J., & Tabares, A. (2015). Repair During Marital Conflict in Newlyweds. Journal of Family Psychotherapy. Observational; all authors Gottman Institute affiliated.
2. Heyman, R. E., & Slep, A. M. S. (2001). The Hazards of Predicting Divorce Without Crossvalidation. Journal of Marriage and Family. N = 528 couples.
3. Kim, H. K., Capaldi, D. M., & Crosby, L. (2007). Generalizability of Gottman and Colleagues’ Affective Process Models of Couples’ Relationship Outcomes. Journal of Marriage and Family, 69(1), 55 to 72. 85 couples from the Oregon Youth Study; only 2 of 22 processes predicted relationship survival. NIH-funded.
4. Lewicki, R. J., Polin, B., & Lount, R. B. (2016). An Exploration of the Structure of Effective Apologies. Negotiation and Conflict Management Research. N = 333 and N = 422, vignettes.
5. Kirchhoff, J., Wagner, U., & Strack, M. (2012). Apologies: Words of Magic? Peace and Conflict: Journal of Peace Psychology. N = 190 and N = 88.
6. Schumann, K. (2014). An affirmed self and a better apology. Journal of Experimental Social Psychology. N = 98 and N = 96.
7. Schumann, K., & Dragotta, A. (2021). Empathy as a predictor of high-quality interpersonal apologies. European Journal of Social Psychology.
8. Schumann, K., & Orehek, E. (2019). Avoidant and defensive: Adult attachment and quality of apologies. Journal of Social and Personal Relationships.
9. Frantz, C. M., & Bennigson, C. (2005). Better late than early: The influence of timing on apology effectiveness. Journal of Experimental Social Psychology.
10. McCullough, M. E., Bono, G., & Root, L. M. (2007). Rumination, emotion, and forgiveness. Journal of Personality and Social Psychology.
11. Schrodt, P., Witt, P. L., & Shimkowski, J. R. (2014). A Meta-Analytical Review of the Demand/Withdraw Pattern of Interaction. Communication Monographs. 74 studies, N = 14,255.
12. Kampf, Z. (2009). Public (non-)apologies: The discourse of minimizing responsibility. Journal of Pragmatics. Corpus analysis of 354 public apologies.
13. Okimoto, T. G., Wenzel, M., & Hedrick, K. (2013). Refusing to apologize can have psychological benefits. European Journal of Social Psychology.
14. Schumann, K., & Ross, M. (2010). Why women apologize more than men. Psychological Science.
15. Pollmann, M. M. H., Crockett, E. E., Vanden Abeele, M. M. P., & Schouten, A. P. (2020). Does attachment style moderate the effect of computer-mediated versus face-to-face conflict discussions? Personal Relationships, 27(4), 939-955. Preregistered, 100 couples.
16. Ruppel, E. K., Cherney, M. R., Quinn, S. F., & Richards, R. J. (2021). Effects of mediated communication on conflict behavior, resolution, and affect in romantic couples. Journal of Social and Personal Relationships, 38(12), 3633-3645. 43 couples.
17. Daspe, M.-È., et al. (2026). Technology-mediated conflict management in couples: a systematic review. Journal of Social and Personal Relationships. Recent; not yet independently replicated.
18. Bevis, B., Schroeder, J., & Yeomans, M. (2026). Spoken disagreement is more constructive than written disagreement. Nature Communications, 17, article 5792.
19. Kruger, J., Epley, N., Parker, J., & Ng, Z.-W. (2005). Egocentrism over e-mail: Can we communicate as well as we think? Journal of Personality and Social Psychology, 89(6), 925-936.
20. Glikson, E., & Asscher, O. (2023). AI-mediated apology in a multilingual work context. Computers in Human Behavior. N = 106, 249 and 232.
21. Liu, B., Kang, J., & Wei, L. (2023). Artificial intelligence and perceived effort in relationship maintenance: Effects on relationship satisfaction and uncertainty. Journal of Social and Personal Relationships, 41(5), 1232 to 1252. N = 208, vignettes.
22. Lim, J. S., Hong, N., & Schneider, E. (2025). Human versus AI-authored apologies. Computers in Human Behavior. N = 464.