Why does replying to a simple message feel like so much work?

“Dinner Friday at seven?” arrived on Tuesday. It is now Thursday. Typing yes would take four seconds, and you have opened the message three times. I co-founded Subtext, and people describe this to me, usually with some version of “I don’t know why this is so hard.”

The term people reach for is mental load. I understand why, and I think it is the wrong name for most of what is going on, because the right name changes what helps. This piece takes the reply apart: what mental load actually refers to, the separate jobs hidden inside one short reply, why texting can wear you out even when each message is easy, and where a writing tool fits and where it does not.

It sits between two pieces I have already written. Why can’t I reply to messages I actually want to answer? is about not being able to start. Why does a two-line reply take me an hour? is about not being able to stop editing. This one is about what the reply is made of.

Is replying to messages part of the mental load?

Partly, and mostly in the part people do not mean.

The best-defined version of it is what the sociologist Allison Daminger calls cognitive labour. Her 2019 study in the American Sociological Review, based on 70 interviews with 35 couples, described the cognitive side of household labour as four activities: anticipating needs, identifying options, deciding, and monitoring whether things got done¹. Women in her sample did more of the anticipating and monitoring, while deciding was shared more evenly. The couples were in the Boston area and mostly college educated, so this is one well-off American group rather than households in general.

That describes running a household over weeks: noticing the birthday is coming, remembering the form is due, checking the plumber came. A message can carry a piece of that. “Can you sort something for Mum’s birthday?” lands in the middle of it, and most of the work it creates comes after you reply, in the planning and tracking. Wording the reply is a smaller, separate job, done in minutes, where the household work runs across weeks. The messages that do carry a household are in why family messages feel like a second job, and whether software can take over any of the noticing is in can AI reduce the mental load?

Three other terms get used for the same feeling, and each comes from a different field.

Cognitive load comes from educational psychology. John Sweller’s work in the late 1980s was about how much a task demands of working memory while you learn or solve a problem². It is a useful lens for a reply, because you are holding the message, your own constraints, possible readings and a draft in mind at once. It was built for teaching, though, so applying it to texting is an analogy that nobody has tested on messages.

Emotional labour comes from sociology. In The Managed Heart, Arlie Hochschild separated emotional labour, the managing of feeling that is sold for a wage in jobs like cabin crew, from emotion work, the same kind of effort in private life³. Softening a reply to a friend who is upset is emotion work in her terms. I keep “emotional labour” for paid work, which is what the word was made for.

Communication overload comes from information systems research. A 2026 scale-development study by Graf and colleagues argues it should be measured separately from information overload, because the two are driven by different processes and need different fixes⁴.

The names matter because each points at a different remedy. Household mental load gets lighter when someone else takes over the anticipating and monitoring. Cognitive load gets lighter when you hold fewer things in mind at once. Emotion work gets easier with time to cool down and a clear sense of what you want. A writing tool touches only some of these, which I come back to below.

The four jobs inside a short reply

This is my breakdown, built from the research below. It is a way of seeing where you are stuck, and nobody has tested it as a stage model.

Reading. Work out what the message is asking and how it was meant. Text strips out voice and face. In Kruger and colleagues’ five experiments, people who wrote messages overestimated how well their tone would come across⁵. I went through those studies, and the figure that circulates from one of them, in does this text sound rude. As the reader you are guessing too, and are they mad at me covers how good those guesses are.

Deciding. Work out what you want, what you can do and what you are willing to promise. “Dinner Friday?” is a calendar question, a money question, an energy question and sometimes a question for your partner before it is a writing question.

Wording. Turn the decision into sentences.

Calibrating. Choose how warm, how long, how firm and how apologetic those sentences should be, for this person, today. Wording and calibrating are the two jobs Subtext was mainly built to help with, and I come back to its limits below.

A fifth job, following through, sits outside the message. Booking the table, calling tomorrow as promised, coming back once you have checked. That is where household mental load comes back in, because it is ongoing and it lives in your head until it is done.

My read is that the job that stalls most replies is deciding. Once the decision is made, the wording usually follows quickly. Until it is made, rewording does not settle anything, and a lot of overthinking a text is a decision dressed up as a wording problem.

What the email research shows about putting replies off

The best behavioural evidence on deferred replies comes from workplace email.

Sarrafzadeh and colleagues interviewed 15 Microsoft employees and analysed two weeks of logs from about 40,000 enterprise users of Outlook on the web, in May 2018⁶. Five of the six authors were at Microsoft, and the logs recorded actions, not message text. The emails people deferred were more likely to ask for a reply or careful reading, and deferral rose with workload. In the interviews, people described the questions they ran through before replying: whether they knew the answer, whether something had to be done first, whether replying meant switching to a different task. Some deferred on purpose, to think about how to respond.

That maps closely onto reading and deciding. It also separates two things the procrastination literature can blur. Putting off a reply until you know the answer is triage, and it is often the sensible thing to do. Putting it off because it feels bad is the mood-repair pattern I covered in why can’t I reply to messages. From the outside, both look like silence.

There are two limits. This is enterprise email, so it says nothing about how often people put off a text to a friend. And I could not find a comparable study of WhatsApp or SMS that combines logs with people’s own reasons. Everything this section says about personal messages is an inference from email.

Switching costs and decision fatigue

Switching costs are real and hard to turn into minutes. In Rubinstein, Meyer and Evans’s lab experiments with University of Michigan students, the cost of switching between tasks grew as the rules got more complex, and cues reduced it⁷. A reply that needs a calendar, a document and an earlier thread involves several switches. A figure for minutes lost per text has no study behind it.

A caution about decision fatigue. The idea that each decision drains a limited daily supply comes from ego-depletion research, and that effect has not held up well. A 23-lab preregistered replication with 2,141 participants found an effect close to zero⁸, and a 36-site test led by one of the theory’s own developers, with 3,531 participants, also found no reliable effect⁹. “I can’t face another decision tonight” is a fair description of how an evening feels. Citing decision fatigue as the science behind it is a stretch.

Why texting can be exhausting even when every message is easy

If one reply is a small job, a phone full of them is a different problem. The research here is about availability, guilt and expectations more than about any single message.

Jeffrey Hall and Nancy Baym studied what they called mobile maintenance expectations in close friendships¹⁰. Expecting to keep in touch by phone was linked to dependence on the friend, which went with higher satisfaction, and to overdependence, which went with lower satisfaction. Entrapment, the guilt and pressure to respond, was linked to lower satisfaction on its own. I could only read the abstract, so I cannot confirm the sample from the paper itself.

Hall’s later work followed this into daily life¹¹. A survey of 300 people came first, then an experience-sampling study in which 112 people reported on 1,879 interactions over five days. People who started the week feeling more entrapped reported lower well-being and a weaker sense of connection during their text exchanges. In the survey, larger networks and more texting were linked to lower well-being through entrapment. It is correlational, and what tracked with lower well-being was the sense of being trapped by the phone.

Being reachable can also be good for you. Taylor and Bazarova followed 114 cohabiting couples for a week in 2019, with five short surveys a day, in a preregistered study¹². Being connected and available to each other, by phone and their other media, was linked to better mood and life satisfaction, for the person and for their partner. Their prediction that one partner’s availability would stress the other was not supported. The effects were small, the design was correlational, and the sample was mostly white, married and in opposite-sex couples. It is still a useful counterweight. The same reachability can feel like a leash in one relationship and like closeness in another.

Expectations also shift with the rest of your life. In a study of WhatsApp in Argentina, Matassi, Boczkowski and Mitchelstein combined 158 interviews with a survey of 700 people¹³. Younger adults described being always on. For people in midlife, WhatsApp was bound up with work and care responsibilities. Older adults were less continuously available. It is observational, from fieldwork in 2016 and 2017. A quick question often lands on a phone already carrying several roles, and the study cannot say whether that makes replies harder to write.

Timing carries meaning when someone is waiting. Huang and Yao analysed answers from 162 undergraduates at a public university in Austin, Texas, who imagined a late reply during a conflict with a partner¹⁴. They expected replies during conflict much faster than usual, about ten minutes against nineteen. Students who expected faster replies, students higher in attachment anxiety and students more committed to the relationship reported more negative feeling about the imagined delay. It was a hypothetical scenario with students, and it measured only negative feeling. Across 171,303 responses in corporate email, course forums and an online question service, Kalman and colleagues found that at least 80 per cent of replies in each dataset arrived within that dataset’s average response time. Among the 89 heavy users they could check one by one, nearly all sent at least 70 per cent of their replies within their own average¹⁵. A late reply stands out against your usual rhythm, which is part of why the person waiting notices.

I looked for research on reply burden among carers, or people in emotionally demanding relationships, and found none. The link is plausible and untested.

Does anxiety make you misread texts?

Less clearly than the popular version says. Earlier work, which I covered in are they mad at me, linked higher social anxiety to more negative readings of ambiguous texts. A 2025 study by Goldman and colleagues tested this with a WhatsApp-style conversation and did not find it¹⁶. They screened 337 first-year students at Tel Aviv University and picked 30 with high and 30 with low social anxiety. Each read a made-up chat from the day after a first date while an eye tracker followed where they looked. Both groups rated the ambiguous parts more negatively than the irrelevant parts. The two groups did not differ in how they interpreted them, and the statistics favoured no difference. The low-anxiety group, if anything, lingered longer on the ambiguous lines.

Sixty students and an invented chat is a small, artificial test, so it does not show that anxiety never colours a reading. It does mean the claim that anxious people systematically misread texts is less settled than it sounds. If you are anxious and a message reads as cold, the message may simply be ambiguous.

ADHD, autism and slow replies

There is real research here, and none of it lets you diagnose anyone from how they text.

A meta-analysis of 13 studies of adults with ADHD found medium-sized differences from comparison groups on executive tasks such as inhibition and switching, and differences of a similar size on tasks that are not executive at all, which led the authors to conclude the difficulties may not be confined to executive function¹⁷. None of the studies measured messaging. I went through the ADHD evidence on procrastination and starting tasks in the best apps for ADHD adults who take forever to reply, including the fact that no study I found treats replying as an outcome.

For autism, the evidence runs in a direction people often do not expect. In a survey of 245 autistic adults, Howard and Sedgewick found phone calls were among the least liked ways to communicate, about level with letters, and “anything but the phone” was the strongest theme in people’s comments¹⁸. That is self-report from an online sample, and preferences vary from person to person. It gives no support to the idea that autistic people find texting harder. The work many autistic people describe in writing, adjusting a message so a non-autistic reader takes it the right way, is in why a two-line reply takes an hour.

Five ordinary replies, taken apart

These are worked examples, not data. I have not put a time on any step, because no study I found measures them.

Confirming dinner. “Dinner Friday at seven?”

1. Is this a firm plan or an idea?
2. Check the calendar, which means leaving whatever you were doing.
3. Check the rest: money, getting there, energy, anyone else’s plans.
4. Decide yes, no, or yes with a condition.
5. Write it with a bit of warmth.

The heavy step is usually the third. Once it is done, the reply nearly writes itself.

Answering a worried friend. “I don’t know what I’m going to do.”

1. Work out whether they want to vent, get advice or get practical help.
2. Notice your own reaction.
3. Decide how much time and attention you can actually give tonight.
4. Check whether anything sounds serious enough to need more than a message.
5. Write something that does not shrink how they feel and does not promise what you cannot give.

The heavy steps are the first and third. If anything suggests someone may be in danger, that calls for a phone call or for help, and the wording of a message comes a long way second.

Postponing a reply. “Can we talk about what happened yesterday?”

1. Recognise that a rushed answer could make it worse.
2. Decide whether silence will be read as an answer.
3. Pick a real time to come back, instead of “later”.
4. Send a short holding message.
5. Actually come back.

The fifth step is follow-through, and it is the one that slips. I could not find anyone who has tested holding messages directly, as I wrote in what makes a reply land well, though the attribution research there suggests that naming a reason for a delay changes how the other person explains it.

Saying no. “Could you take this on for me?”

1. Decide whether you are willing.
2. Separate “I don’t want to say no” from “I’m not sure”.
3. Decide whether you owe a reason.
4. Decide whether to offer something smaller, or nothing.
5. Write it without stacking apologies.
6. Accept that they might be disappointed.

The first four are decisions, and only the fifth is writing. How to say no to extra work and whether setting boundaries is good for you cover the research on the deciding part.

Clearing up a misunderstanding. “That’s not what I meant.”

1. Reread and find the sentence where your readings split.
2. Separate what you meant from what you wrote.
3. Decide whether to apologise, explain, or both.
4. Write a specific correction, without a verdict on who was wrong.
5. Ask whether it lands, or suggest a call if text keeps making it worse.

The heavy step is the first. My guess is that a reply that argues about intent before finding that sentence tends to make things worse.

Which kind of reply is this?

This is the decision tree I use. It is a practical summary of the research above, and nobody has tested it as an instrument.

1. Is something urgent or unsafe happening in real life? Deal with that first, by phone or in person if needed. The wording comes second.
2. Do you know what you want to say or do?
   • Yes, and the words come easily. Reply now.
   • Yes, but you cannot tell how it will land. Draft with help, from a template, a friend or a tool such as Subtext, then edit it back into your own words before you send.
   • Not yet, because you are missing information, someone else’s input or time to cool down. Acknowledge and defer: “I’ve seen this. I need to check something and I’ll reply tomorrow evening.” Then put the reminder somewhere outside your head.
   • Not yet, because you do not want what is being asked. Set a boundary. Decide first, then work on the wording.

Eight replies, before and after

I wrote these to illustrate the idea, and none of them has been tested. The column that matters is the last one. Each “after” was easier to write because something got decided first.

What helps without any app

Saying how you reply. “I’m slow on messages during the working day, ring me if it’s urgent.” The entrapment research puts the strain in guilt and expectations, and this sentence changes the expectations. Nobody has tested the sentence itself. In a related finding I covered in what makes a reply land well, one line from an email sender saying there was no rush reduced the recipient’s sense of urgency.

A short acknowledgement. It separates “I’ve seen this” from “here is my full answer”. It also creates a promise, so it only helps if the reminder lives somewhere you will see it.

Batching. Answering at set times cuts down on switching. It suits most messages and suits a conflict badly, where Huang and Yao’s students expected answers within minutes.

Boundaries. These work on the deciding job directly. If you keep rewording a yes you do not mean, the thing to change is the yes.

Templates. Pre-decided phrases for situations that repeat: “I’ve seen this and need to check before I answer.” “Thanks for asking. I can’t commit to that.” They save the wording job for familiar cases and do little for a new or delicate one.

What an AI writing tool can and cannot do here

A writing tool works on the wording and calibrating jobs. It does nothing for the decision or the follow-through, and it cannot know what the other person meant.

AI help has a cost when the other person suspects it. In Hohenstein and colleagues’ preregistered experiment with 219 pairs of online participants, AI smart replies made conversations faster and more positive¹⁹. When people believed their partner was using AI, they rated that partner as less cooperative and less affiliative, roughly less warm, even though actual use of the suggestions improved those ratings. A second experiment with 291 pairs varied how positive or negative the suggestions were. Negative suggestions made conversations more negative than positive or standard suggestions did, and the shift came from the suggestions people actually sent. The publisher later corrected the paper to add its funding statement, which does not change the results. What rewriting does to voice is a separate question, covered in does AI make your texts sound like a robot.

I would hold any reply tool to these criteria:

• When a message needs a decision, it asks what you want before it writes.
• It tells you how your draft reads, instead of only replacing it.
• Its versions stay close to your own words.
• It leaves sending to you.
• It tells you where your text goes and how long it stays there.

I co-founded Subtext, so read this part knowing that. It was built for the wording and calibrating jobs, and it helps with part of the reading. You bring it a draft, a screenshot of the thread or a voice note. For a message you received, it tells you in plain words what the other person wants, listing up to five things that need a response²⁰. For an important message, it stops and asks what you want to get across before it writes anything²¹²². That hands the deciding job back to you instead of guessing. It marks the words in your draft that could land badly, offers up to three versions with a verdict on each, and lets you make one warmer, shorter, more casual, funnier or more confident with a tap²¹. Nothing is sent for you. Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

What it does not do matters as much. Subtext cannot tell you what the other person privately meant, and its own FAQ says so²¹. It does not decide whether you should say yes, apologise or end something. It is not a reminder app either. Conversations delete themselves five days after you last use them, or 90 days if you pin one²³, so it is the wrong place to keep track of who you owe a reply. A calendar or a reminders app does that job, and the ADHD apps piece compares the options. If you want to compare Subtext with the rewrite buttons built into your phone, I did that in every phone rewrites your messages now.

Before you paste a private conversation into any AI

A screenshot of a thread contains someone else’s words, often their name, and sometimes their health, money or address. They did not agree to any of it going to an AI provider.

For Subtext, under the privacy policy updated on 10 October 2026, your text, images and voice transcripts go to Anthropic for analysis and drafting, voice recordings go to Deepgram for transcription, and the AI processing happens in the United States²³. Nothing you send is used to train a model. Anthropic deletes inputs and outputs within 30 days but can keep requests flagged for a safety review for up to two years, and the policy says plainly that Subtext has no zero-retention arrangement with any provider. The terms also ask you to confirm you are allowed to submit other people’s words. “Not used for training” and “never processed or kept” are different promises, for every tool, mine included. I compared thirteen tools on this in which AI writing assistants train on your messages, and what goes wrong when AI reads a screenshot is in AI can read a screenshot of your texts.

The habit is the same whatever you use. Crop to the part that matters. Remove names, health details, money and addresses that have nothing to do with the wording. Leave out anything you were told in confidence. And when a suggestion comes back, change it until it sounds like you, because the person reading it knows how you text.

What I’d take from this

My read is that a reply that feels heavy is usually carrying a decision. Work out which job is stuck. If it is the decision, a tool will only make the indecision sound nicer. If it is the words, help is useful, as long as you keep your voice. If it is the follow-through, write the reminder down somewhere you will see it, because the chat will not hold it for you.

And if you are the one waiting, the person on the other side may be doing all four jobs while juggling several other roles. Something may still be owed, and silence is a poor guide to how much they care.

Sources

Numbered in the order they appear above. Studies were checked against the publisher page or the authors’ own copy on 12 and 13 October 2026, and Subtext’s pages on 12 October 2026. Where a figure could not be verified against the primary source, the text or the entry says so.

1. Daminger, A. (2019). The Cognitive Dimension of Household Labor. American Sociological Review, 84(4), 609 to 633. 70 in-depth interviews with 35 couples. The Boston area and the mostly college-educated, middle and upper-middle class sample come from the author’s dissertation and later writing, not the article page I could open. Crossref lists Harvard University as funder.
2. Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12(2), 257 to 285. Research on learning and problem solving, not messaging.
3. Hochschild, A. R. (1983). The Managed Heart: Commercialization of Human Feeling. University of California Press. The distinction between emotional labour, sold for a wage, and emotion work in private life is in the note on page 7, which I checked through academic quotations of it rather than the book itself. Her earlier article, Emotion Work, Feeling Rules, and Social Structure (American Journal of Sociology, 85(3), 1979), uses only “emotion work”.
4. Graf, B., Beinig, R., Schwickert, S., and Schischke, D. (2026). Consolidation and integration of information overload and communication overload: A theory-based derivation of distinct measurement models and scale development. International Journal of Information Management, 88, 103045. Preregistered scale development: a development sample of 46, then two professional samples of 105 and 288. Open access. Funder not listed.
5. Kruger, J., Epley, N., Parker, J., and 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 to 936. Five studies with Cornell and University of Illinois students. Funded by a University of Illinois grant and NSF SES-0241544.
6. Sarrafzadeh, B., Hassan Awadallah, A., Lin, C. H., Lee, C.-J., Shokouhi, M., and Dumais, S. T. (2019). Characterizing and Predicting Email Deferral Behavior. Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining (WSDM 2019). Interviews with 15 Microsoft employees and logs of about 40,000 enterprise Outlook web users, 6 to 19 May 2018, with no message text. Five of the six authors were at Microsoft, and the first did the work there. Read in the authors’ preprint.
7. Rubinstein, J. S., Meyer, D. E., and Evans, J. E. (2001). Executive control of cognitive processes in task switching. Journal of Experimental Psychology: Human Perception and Performance, 27(4), 763 to 797. Four lab experiments with University of Michigan undergraduates, N = 12, 36, 36 and 24. Funded by NIMH grant MH16892.
8. Hagger, M. S., Chatzisarantis, N. L. D., and colleagues (2016). A Multilab Preregistered Replication of the Ego-Depletion Effect. Perspectives on Psychological Science, 11(4), 546 to 573. 23 labs, N = 2,141, d = 0.04, 95 per cent confidence interval −0.07 to 0.15.
9. Vohs, K. D., Schmeichel, B. J., and colleagues (2021). A Multisite Preregistered Paradigmatic Test of the Ego-Depletion Effect. Psychological Science, 32(10), 1566 to 1581. 36 sites, N = 3,531, d = 0.06, not significant.
10. Hall, J. A., and Baym, N. K. (2012). Calling and texting (too much): Mobile maintenance expectations, (over)dependence, entrapment, and friendship satisfaction. New Media and Society, 14(2), 316 to 331. Published online in 2011. Findings as stated in the abstract; the full text is paywalled, so the sample size and design are not verified here. Funder not found.
11. Hall, J. A. (2017). The experience of mobile entrapment in daily life. Journal of Media Psychology, 29(3), 148 to 158. Study 1, survey, N = 300. Study 2, experience sampling, N = 112, 1,879 interaction reports, five a day for five days. Correlational. Funder not found.
12. Taylor, S. H., and Bazarova, N. N. (2021). Always Available, Always Attached: A Relational Perspective on the Effects of Mobile Phones and Social Media on Subjective Well-Being. Journal of Computer-Mediated Communication, 26(4), 187 to 206. Preregistered, 114 cohabiting couples (228 people), recruited January to March 2019, seven days of five surveys a day, 78 per cent answered. 87 per cent white, 82.5 per cent married, 91 per cent opposite-sex couples. Funded by the Cornell Center for Social Sciences.
13. Matassi, M., Boczkowski, P. J., and Mitchelstein, E. (2019). Domesticating WhatsApp: Family, friends, work, and study in everyday communication. New Media and Society, 21(10), 2183 to 2200. 158 semi-structured interviews and a survey of 700 in Argentina. The fieldwork dates and the survey’s Buenos Aires area come from a later account of the same project, not the article. Funder not found.
14. Huang, Q., and Yao, S. (2024). No response? Chronemic expectancy violation and relational turbulence in technologically-mediated romantic relational conflict. Journal of Social and Personal Relationships, 41(5), 1183 to 1206. A pilot survey of 50, then 200 students aged 18 to 23 at a public university in Austin, 79.5 per cent women, 162 in the final analysis. Hypothetical scenario; negative feeling was the only outcome measured. Not preregistered, no funding.
15. Kalman, Y. M., Ravid, G., Raban, D. R., and Rafaeli, S. (2006). Pauses and Response Latencies: A Chronemic Analysis of Asynchronous CMC. Journal of Computer-Mediated Communication, 12(1), 1 to 23. 171,303 responses from Enron email, university course forums and Google Answers.
16. Goldman, L., Basel, D., Hallel, H., Hertz-Palmor, N., and Lazarov, A. (2025). Instant messaging communication in social anxiety: no support for the combined cognitive bias hypothesis (CCBH). Motivation and Emotion, 49, 439 to 454. 337 first-year Tel Aviv University students screened, 30 with high and 30 with low social anxiety tested with eye tracking. Not preregistered; data shared on OSF. Open access funded by Tel Aviv University; no conflicts of interest declared.
17. Boonstra, A. M., Oosterlaan, J., Sergeant, J. A., and Buitelaar, J. K. (2005). Executive functioning in adult ADHD: a meta-analytic review. Psychological Medicine, 35(8), 1097 to 1108. 13 studies of adults with ADHD and comparison groups. The authors report no financial support.
18. Howard, P. L., and Sedgewick, F. (2021). “Anything but the phone!”: Communication mode preferences in the autism community. Autism, 25(8), 2265 to 2278. 245 autistic adults recruited online, self-report. No funding.
19. Hohenstein, J., Kizilcec, R. F., DiFranzo, D., and colleagues (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. Experiment 1 preregistered, 219 pairs (438 online workers). Experiment 2, 291 pairs. A Publisher Correction on 3 October 2023 added the omitted funding statement, NSF grant CHS 1901151. No competing interests declared.
20. Subtext. Read the message you received before you reply. Checked 12 October 2026.
21. Subtext. Subtext homepage and FAQ. Checked 12 October 2026.
22. Subtext. What Subtext is built on. Checked 12 October 2026.
23. Subtext. Terms, privacy and your account. Privacy policy last updated 10 October 2026, terms last updated 26 July 2026. Checked 12 October 2026.