Why do family messages feel like a second job?

Family messages feel like a second job because many of them are part of one. Remembering the birthday, noticing that an aunt has gone quiet, deciding who answers the group chat about Christmas, finding words for a parent’s disappointed text. Researchers have called this work kinkeeping since the 1980s, and they have found it falls mostly on women¹²³. A family message is usually the visible end of a longer job. Someone noticed it was needed, decided what to say and will check how it landed.

That makes it part of what people now call the mental load, which got fresh attention this October after a viral post asked, in effect, whether AI could be the partner who notices things. I go through that debate and the evidence in the main piece, can AI reduce the mental load? This piece is narrower. It covers the messages themselves, what research says about getting AI help with them, and the privacy question raised by pasting other people’s words into an app.

I co-founded Subtext, an AI writing assistant for personal messages, so I have a stake in this topic. I say where Subtext fits, and when not to use it, near the end. I have written separately about reply procrastination, condolences, apologies, arguments and boundaries, and I link to those where they fit.

What is kin keeping?

It is the work of keeping relatives connected. Carolyn Rosenthal defined “kinkeeping” in a 1985 paper that reported women were more likely than men to be kinkeepers; I’ve only seen that paper summarised, so treat the detail as second-hand³. In 1987 Micaela di Leonardo published “The Female World of Cards and Holidays: Women, Families, and the Work of Kinship”¹. The title names the territory. She described kin work, such as visits, letters, calls, presents, cards and the organising of holiday gatherings, as work done mostly by women¹.

The effort shows up in the research. In 1993 the sociologists Gerstel and Gallagher found that kinkeeping was done mostly by women, though men did some, and that its breadth was associated with depression for both². That is a correlation, and the abstract does not give the sample. The link was with breadth, the number of relatives a person helps. Depth, the hours spent helping them, showed no such link.

The thinking stages of household work apply here too. Allison Daminger’s model, built from interviews with members of 35 couples, splits cognitive labour into anticipating a need, identifying options, deciding and monitoring⁴. A birthday message runs through all four. You notice the date is coming, think about what to send, choose, and check later that the card arrived or the call happened. If one partner in your home does all of that for both families, how to explain the mental load to your partner covers that conversation.

Effortful messages wait longer, at least in email

In logs from about 40,000 users of a commercial email client, deferral was more likely when handling a message needed effort, such as replying or careful reading. In the same study’s interviews, people said that when a message has several recipients they may hold back, expecting someone else to reply⁵. That was workplace email and has not been tested on families. I’d expect the same pattern in a family group chat.

Family messages usually carry more history than work email. A reply to a sibling is also a reply to years of being siblings. The reasons a short reply can take an hour are covered in why a two-line reply takes an hour, and the reasons we avoid replies we want to send are in why I can’t reply to messages. The separate jobs inside one reply are in why replying to a simple message feels like so much work. This piece is about which messages carry the load.

Family messages that carry the mental load

Most of the heavy ones involve a decision before any typing starts.

How do you respond to guilt trips from your mum?

Answer the feeling, then state what you will do. This goes for a guilt trip from any parent or relative. I haven’t found research on guilt-trip texts specifically, so what follows is my own approach, and you should adapt it.

A guilt trip usually bundles two things: a feeling (“I miss you”, “I feel left out”) and a request you have already decided about. Replying to the request alone sounds cold. Replying to the feeling alone leaves the request hanging, and it will come back. So do both, briefly, and offer something you can give. An example you could adapt:

Received: “I suppose you’re too busy for us again this Sunday.” Reply: “It sounds like you’d like to see more of us, and I miss you too. I can’t do Sunday. Could we call on Tuesday evening?”

I’d leave out long justifications, because each reason you give can become something to argue with. If the guilt trips are part of a wider pattern of control, better wording won’t fix that, and my post on whether setting boundaries is good for you goes into it, including when to get support.

How do you say no to family without feeling guilty?

You may not manage the “without feeling guilty” part, and I’d treat that as normal. What you can control is a message that is clear and kind, and I think it helps to say the no once, plainly, without circling back to it.

I’d keep a refusal short, for the reason above. An example for a request to host relatives:

“Thank you for thinking of us. We can’t host this time, the flat is too small for a week of guests. Could we meet for lunch while you’re in town?”

That gives one reason, no apology spiral, and a real offer. If money is the request, I’d keep the reason even shorter, or leave it out. The feature page say no without over-explaining covers the wording. Saying no at work is a different post: how do you say no to extra work?

Who ends up making the list in the group chat

In many families, I suspect, it is the same person every time. The “someone else will reply” effect that interviewees in the email study described for messages with several recipients⁵ is a plausible description of a family chat where eight people read “who’s bringing what on Saturday?” and one person ends up making the list.

If that person is you, two small habits may help. Name an owner for each decision in the chat itself, for example “Mia, can you book the restaurant by Wednesday?”, so the job is visible. And when a long thread settles something, post a two-line summary of what was decided.

Can AI help me respond to a text?

It can help you write one. Whether it helps the relationship may depend on whether the other person finds out. The research on AI-mediated communication splits into two strands, and both apply to Subtext as much as to any other tool.

The first strand suggests that help nobody notices can go with better conversations. In a 2023 study in Scientific Reports, 219 pairs of crowdworkers chatted on an app with or without Google’s suggested replies. Having suggestions available increased messages per minute by 10.2%. Where people used the suggestions, their partners tended to rate them as more cooperative, and the language was more positive⁶. In a second experiment with 291 pairs, negative suggested replies made conversations more negative⁶.

The second strand is about what happens when people know or suspect. In that same study, people who thought their partner was using suggested replies rated them as less cooperative, whether or not it was true⁶. In a CHI 2022 survey, 147 crowdworkers rated trust in an email’s writer at 4.02 out of 5 when it was attributed to the sender, 3.68 with AI autocomplete, and 3.51 when attributed to AI⁷. In scenario studies of a workplace conflict, using AI tools to write an apology lowered its perceived authenticity and people’s willingness to forgive, and disclosing the help did not remove the penalty⁸.

In experiments, recipients rarely noticed on their own. Two US experiments with 647 and 654 adults found that, without being told, people did not suspect AI in realistic personal messages such as an apology. Told that AI wrote it, they judged the sender more harshly⁹. I’ve read the abstract, a preprint of the first experiment and the co-author’s own summary, not the full paper. A study with representative UK (477) and US (765) samples found that secret use of AI communication tools was judged less acceptable than open use, though only when people compared the two directly, as the UK sample did¹⁰.

These findings sit uncomfortably together. Help that goes unnoticed seems to do no harm and may do some good, disclosed help is penalised, and people disapprove of hidden help when they compare it with open help. One recipient quoted by Business Insider asked, in effect, whether their message was so hard to answer that the sender needed ChatGPT¹¹. Most of these studies are vignettes or experiments with online crowdworkers. None looked at real family messages.

There is also a style risk. In an experiment with 118 participants from India and the US, AI suggestions pulled Indian participants’ writing towards Western styles¹². If your family writes in another language or with its own conventions, read the draft back and make sure it still sounds like your family.

Should you tell someone you used AI?

The research does not settle it, and my position is the one I set out in does AI make your texts sound like a robot? If the words are yours and AI only tidied them, I think saying so is fine. The studies above suggest it can still cost some trust, and the one study I found on light edits, described in that post, points the other way. If AI wrote the message, rewrite it in your own words before you send it, and then the question mostly goes away. The apology studies offer one hint. Limited use, one tool out of three, did not harm how authentic the apology seemed⁸.

For family messages I’d add one test. Would you be comfortable if the recipient saw the draft and your edits side by side? If yes, you used a writing aid. If no, the message probably needs more of you in it.

Using ChatGPT to analyse text messages: whose words are you sharing?

When you paste a family chat or upload a screenshot into any AI tool, you share other people’s words with a company they never chose. Your mother’s message about her diagnosis or your brother’s money worries go wherever your tool sends them. I found no study on how people feel about their messages being pasted into an AI tool.

The rules depend on the tool. I can speak for Subtext’s own legal page. It acknowledges that screenshots often contain someone else’s words, and by submitting them you confirm you may lawfully do so. Content goes to named processors, including Anthropic and Deepgram in the US, and is not used for training. The policy states that Subtext has no zero-retention arrangement with any provider. Conversations are deleted five days after last use, or 90 if you pin them¹³. Neither the Terms nor the Privacy Policy says anything more about the people in your screenshot¹³.

Other tools make different choices. Google says Gmail’s Help me write, which personal accounts without a paid Google plan get only in the US, draws on your other emails and Drive files to personalise drafts¹⁴. WhatsApp says its optional Writing Help, launched in English in the US and some other countries in 2025, runs on Private Processing, so that Meta cannot read the message or the rewrites, and cites outside audits¹⁵. Apple says its Private Cloud Compute does not store data¹⁶. All of these are vendor claims. ChatGPT uses conversations on personal accounts for training unless you turn that setting off. I quote its policy in which AI writing assistants train on your messages?, and what happens to screenshots is in AI can read a screenshot of your texts.

In the EU, an individual using a tool for purely personal or household purposes may fall outside the GDPR, while the company providing the tool is still covered¹⁷. If the exemption applies, it covers you, and your relatives’ words are still personal data in the provider’s hands. I’d crop the screenshot to the messages that matter, cover names and numbers, leave out other people’s health details, and where possible describe the situation in a sentence of my own.

Should you use AI to write a condolence message?

I wouldn’t let it write one. Edit with it if you like. The research is mixed. In the CHI 2022 survey, consolation emails about a lost pet were trusted more than product enquiries, even when labelled as AI-written⁷. In the same project, all 10 interviewees said fully autonomous AI writing would be inappropriate for highly interpersonal messages⁷. A condolence is about as interpersonal as messages get.

What to say, and what to avoid, is covered in what do you say when you don’t know what to say? Write the first sentence yourself, however clumsy, and keep it.

Where Subtext fits, and when not to use it

Subtext is an AI writing assistant that helps you draft, rewrite, proofread and understand messages. You paste or upload what you want help with, as text, a screenshot or a voice note. It suggests up to three versions, each marked safe to send or with a warning based on Subtext’s own estimate¹⁸. It sends nothing for you. You copy the version you want and send it yourself, and it connects to none of your accounts¹⁸¹³. Its research page states that “No study on this page tested Subtext”¹⁹.

For family messages, Subtext suits two jobs. One is summing up a long message as what the sender wants from you and the points to answer, before you reply. The other is marking the words in your own reply, to a guilt trip or to a request, that could land badly. Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

It does none of the kinkeeping. It cannot remember the birthday, notice the silence or decide who books the restaurant. If you want an assistant that acts on your behalf, the checks in how to tell if an AI assistant is proactive apply, and Subtext is the wrong tool by design.

Don’t let it, or any AI tool, write the message when the message exists to show effort, such as a condolence or a long-overdue apology. Don’t paste a relative’s private crisis when a one-line description would do. And if family contact itself feels unsafe, no draft fixes that. A specialist helpline where you live is the place to start, and how to ask your partner to share the mental load lists some under “When raising it isn’t safe”. For the wider question of what software can take off a household, the main piece is can AI reduce the mental load?

Questions people ask

Why do women do most of the kin keeping? Studies since the 1980s have found women more likely to be kinkeepers, and men do some too²³. For household mental labour generally, a 2023 review of studies of heterosexual couples attributes the gap mainly to motivation shaped by stereotypes and social roles, and argues that it has no basis in ability²⁰.

How do you respond to a long text message? List the questions it asks and answer those first. Messages that take effort are more likely to be put off, at least in workplace email logs from about 40,000 users⁵.

How do you reply to a text you ignored? Briefly acknowledge the gap and answer the message. In the same email study, 16% of logged users deferred at least one email a day⁵. More in why you can’t reply to messages you want to answer.

How can you tell if someone used AI to text you? Often you won’t. In two US experiments with 647 and 654 adults, recipients did not suspect AI in realistic personal messages unless told⁹.

Is it OK to use AI to polish my writing? The evidence is kinder to light editing. In apology scenarios, using AI tools lowered perceived authenticity, and limited use (one tool of three) did not⁸.

Can I use ChatGPT for texting? For drafting, yes. Read it back in your own voice, and check what the tool keeps. The measured benefits of suggested replies so far come from lab chats between crowdworkers⁶.

Why are people using AI to text their friends? Business Insider reported people using it to rehearse hard conversations and to de-escalate co-parenting exchanges¹¹.

How do you text a friend or relative you haven’t talked to in years? Keep it short and name the gap without a long apology. I wrote a reconnection script in why do people ghost?

Sources

Numbered in order of first appearance. Each entry says what kind of source it is and how much of it I could read. Pages were checked on 12 and 13 October 2026.

1. The Female World of Cards and Holidays: Women, Families, and the Work of Kinship. Micaela di Leonardo, Signs 12(3): 440 to 453. April 1987. . Peer reviewed; paywalled, so her list of kin work is as later authors quote it from p. 442. Checked 12 October 2026.
2. Kinkeeping and distress: gender, recipients of care, and work-family conflict. N. Gerstel and S. K. Gallagher, Journal of Marriage and the Family 55(3): 598 to 607. 1993. . Peer reviewed; read via AGRIS abstract record; sample not stated in the record; correlational. Checked 12 October 2026.
3. Kinkeeping. Wikipedia contributors. Accessed 12 October 2026. . Encyclopedia summary, cited here for Rosenthal, C. J. (1985), “Kinkeeping in the familial division of labor”, Journal of Marriage and the Family 47(4): 965 to 974; original not read. Checked 12 October 2026.
4. The Cognitive Dimension of Household Labor. Allison Daminger, American Sociological Review 84(4): 609 to 633. Online 9 July 2019. . Peer reviewed; qualitative, 70 interviews with members of 35 couples. Checked 12 October 2026.
5. Why people don’t reply to your emails. University of Waterloo, Cheriton School of Computer Science. 3 April 2019. . University news summary of Sarrafzadeh et al., WSDM 2019; 15 interviews and logs from about 40,000 users of a commercial email client; workplace email. Checked 12 October 2026.
6. Artificial intelligence in communication impacts language and social relationships. Hohenstein, Kizilcec, DiFranzo, Aghajari, Mieczkowski, Levy, Naaman, Hancock and Jung, Scientific Reports 13: 5487. 4 April 2023. . Peer reviewed; two randomised experiments with Mechanical Turk crowdworkers (219 and 291 pairs); publisher correction of 3 October 2023 added the omitted funding statement (NSF). Checked 12 October 2026.
7. Will AI Console Me when I Lose my Pet? Understanding Perceptions of AI-Mediated Email Writing. Liu, Mittal, Yang and Bruckman, CHI 2022. 2022. . Peer reviewed; Wizard-of-Oz survey of 147 MTurk workers plus 10 student interviews. Funding not stated in the text read. Checked 12 October 2026.
8. AI-mediated apology in a multilingual work context: Implications for perceived authenticity and willingness to forgive. Glikson and Asscher, Computers in Human Behavior 140: 107592. 2023. . Peer reviewed; three scenario studies; sample sizes not available on the record page. Funding not disclosed on the record. Checked 12 October 2026.
9. Blissful (A)Ignorance: Despite the widespread adoption of AI in communication, people do not suspect AI use in realistic contexts. Jiaqi Zhu and Andras Molnar, Computers in Human Behavior 180: 108929. 2026. . Peer reviewed; two US online experiments, 647 and 654 adults. Full text not read: I read the abstract, the preprint of the first experiment and the co-author’s summary for The Conversation, reposted by the University of Michigan at https://prod.lsa.umich.edu/psych/news-events/all-news/faculty-news/most-people-do-not-realize-when-a-personal-message-they-receive-.html. Checked 13 October 2026.
10. People have different expectations for their own versus others’ use of AI-mediated communication tools. Zoe A. Purcell, Mengchen Dong, Anne-Marie Nussberger, Nils Köbis and Maurice Jakesch, British Journal of Psychology 117(2): 548 to 566. Online 4 September 2024. . Peer reviewed; representative samples, UK n = 477 and US n = 765; read in the preprint at https://arxiv.org/abs/2305.01670. Checked 13 October 2026.
11. ChatGPT has entered the group text. Amanda Hoover, Business Insider (syndicated by Yahoo). Updated 2 July 2025. . Journalism; read via syndicated copy. Checked 12 October 2026.
12. AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Agarwal, Naaman and Vashistha. 2025. . Controlled experiment, 118 participants from India and the US; ACM DOI consistent with CHI 2025, read via arXiv. Checked 12 October 2026.
13. Terms, Privacy and your account. Subtext (Anton Licht, sole proprietor). Terms updated 26 July 2026; Privacy Policy updated 10 October 2026. . Our own page. Checked 12 October 2026.
14. Draft emails with Gemini in Gmail. Google, Gmail Help. Undated. . Vendor page; Help me write for personal accounts without a paid plan is US only. Checked 12 October 2026.
15. Get the Tone of Your Message Right with Private Writing Help. WhatsApp Blog. 27 August 2025. . Vendor page; audit claim not independently checked; EU availability not verified. Checked 12 October 2026.
16. Apple Intelligence comes to iPhone, iPad, and Mac starting next month. Apple Newsroom. 9 September 2024. . Vendor page. Checked 12 October 2026.
17. Recital 18: Not Applicable to Personal or Household Activities. gdpr-info.eu (unofficial reproduction of Regulation (EU) 2016/679). Accessed 12 October 2026. . Checked 12 October 2026.
18. Subtext app: AI writing assistant for texts and replies. Subtext. Undated. . Our own page; first-party description. Checked 12 October 2026.
19. The research behind Subtext. Subtext. Undated. . Our own page. Checked 12 October 2026.
20. Gendered Mental Labor: A Systematic Literature Review on the Cognitive Dimension of Unpaid Work Within the Household and Childcare. Natalia Reich-Stiebert, Laura Froehlich and Jan-Bennet Voltmer, Sex Roles 88: 475 to 494. 29 April 2023. . Peer-reviewed systematic review of 31 studies, all of heterosexual people or couples. Checked 12 October 2026.