Can an "AI wife" reduce the mental load?

On 5 October 2026 (US Pacific time) Sonia Baschez, whom Business Insider describes as a marketing entrepreneur and podcaster, posted on X that Silicon Valley keeps building her a husband when what she wants is a wife¹²³. Ben Tossell’s newsletter Ben’s Bites picked it up on 8 October². Her point, in my paraphrase, is that agents which do the tasks you assign them leave her life largely unchanged, and she wants one that knows what is going on in it and anticipates what is needed¹⁴. On 9 October Business Insider ran a column by Katie Notopoulos saying she needs her agent to be a wife³⁴.

Can software take the mental load off a household, or only some of the errands? As of 12 October 2026, assistants can take on some of the errands, and I found no controlled study of whether they reduce the thinking around them. I’m not a sociologist, so treat this as one founder’s reading of the research.

I should also say where I stand. I co-founded Subtext, an AI writing assistant, with Anton Licht⁵. It helps with one hard message at a time and runs no part of a household. Near the end I say where it fits and when you shouldn’t use it, and I have a stake in that section.

Why do some people want an “AI wife”?

As I read it, Baschez was describing two kinds of work: the kind that starts when someone tells you what to do, and the kind that starts before anyone asks, like noticing the school form. Notopoulos tested agents on family logistics. Meta’s Muse could fill in and send a school form, yet for many other tasks she still had to give it step-by-step instructions, which she compared to asking her husband to do something³⁴. Baschez told her she knows the metaphor plays on gender stereotypes that don’t fit every couple, and finds it useful shorthand anyway⁴.

The debate is small: the post (about 791,000 views according to the newsletter’s embed, captured at an unknown time), the newsletter item, the column and an AI-generated aggregator page²³⁶. I found no published response from researchers, single parents, disabled people or LGBTQ+ writers, though X, Threads and TikTok couldn’t be searched directly.

The phrases describe a wish. I found no household assistant sold as an “AI wife” or an “AI husband”. “AI wife” already meant an AI romantic companion, reported since at least January 2023⁷, which is why Ben’s Bites added “not like that” to its subtitle².

The image is older than the post. To me it echoes Judy Brady’s satirical 1970 essay “I Want a Wife”⁸, and in 2020 Yolande Strengers and Jenny Kennedy’s book The Smart Wife criticised assistants made feminine by design⁹. This time a user asks for a wife as a job description.

What researchers mean by mental load

Sociologists often call it cognitive labour, and the standard model comes from Allison Daminger. Her 2019 study draws on 70 interviews with members of 35 couples, college-educated, middle- to upper-middle-class married parents of young children, mostly in different-sex couples¹⁰¹¹. It splits the thinking side of running a home into four stages:

• Anticipating: noticing that something will be needed before anyone asks.
• Identifying options: working out the ways that need could be met.
• Deciding: choosing one.
• Monitoring: checking that it happened and worked.

Her 2025 book renames the last stage “following up”¹², which I prefer, because it implies closing the loop.

In 26 of 32 different-sex couples the wife did more cognitive labour overall, and women did more of the anticipating and monitoring in particular. Deciding, the stage most tied to power, was shared roughly equally, and identifying options was more often split¹⁰¹¹. Heavier loads may also go with strain. In an online convenience sample of 368 women, higher mental-load scores went with more anxiety and depressive symptoms, and a second sample of 1,871 mothers linked mental load to neuroticism. Both are cross-sectional, so neither can show which causes which¹³.

One more finding matters for AI. A focus-group study of 25 mothers counted “managerial thinking”, including delegating and instructing, as mental labour¹⁴. To my mind, writing instructions for an agent belongs in that category.

What the research says about who carries the load

Daminger’s findings, and most that followed, describe a division of labour in some different-sex couples. None of it shows that anticipating is a female capacity.

A 2023 systematic review of 31 studies, all of heterosexual individuals or couples, attributes the gender gap mainly to motivation shaped by stereotypes and social roles, and argues that it has no basis in ability¹⁵.

Men carry part of the load, and the split is specific. In a German probability sample of 2,675 people in different-sex couples, surveyed in early 2023, about half of the women saw themselves as mainly responsible for daily-life mental work, and about a quarter of the men said their partner was. Men claimed main responsibility for only 3 of 21 tasks: repairs, tradespeople and bills¹⁶.

In a US survey of 2,133 partnered heterosexual parents, mothers reported 13.72 of 21 cognitive tasks as “mostly me” and fathers 8.18. Mothers who were employed or had high incomes did less physical housework, but their cognitive labour was no lower¹⁷. The survey is cross-sectional, so it shows a pattern and no cause. Still, if higher income doesn’t go with less thinking work, an app that takes over execution has no obvious way to reduce it either.

So the metaphor imports three problems. It turns a distribution of work into a personality. It casts men as competent and oblivious, hiding the homes where men do the anticipating. And it assumes a two-adult, different-sex home, which leaves out single parents (who report doing nearly all of it, according to a co-author’s summary of one US survey¹⁸), same-sex couples, flatmates and families organised around a disabled member’s care. Noting the female name of a family-assistant app called Anna, the journalist Rose Hirst wrote that “even when outsourcing our mental load to AI, we must pass it to a woman”¹⁹.

From here on I’ll use plainer terms: an instruction-following assistant acts once told; an outcome-owning one notices a need, acts within agreed limits and follows up until the matter is closed.

Can you outsource your mental load?

Taking responsibility for an outcome, as I read two papers on delegating to AI agents, needs five things: the assistant notices a need on its own, decides within agreed limits, follows up, handles failure before you notice, and someone answers for it when it goes wrong. An OpenAI white paper holds that at least one human should be accountable for every uncompensated direct harm an agent causes²⁰. A February 2026 Google DeepMind preprint defines delegation as handing over authority, responsibility and accountability, and requires irreversible, high-stakes failures to be escalated to a human or stopped²¹. Software cannot be held to account in that sense, so read together they suggest to me that the work can move to software while a person still answers for it, as the agent’s supervisor.

An instruction-following assistant takes over only the execution. When Hirst tested two family-assistant apps for The Spinoff in September, both needed her to notice tasks and feed them in, a school-donation reminder never arrived, and a 2:18 am alert concerned a bill already on direct debit¹⁹. The sociologist Leah Ruppanner told her: “You can hold the mental load, even if you’ve delegated the labour”¹⁹.

Delegating creates work of its own. Lisanne Bainbridge argued in 1983 that designers automate what they can and leave the operator whatever remains, and that it is humanly impossible to keep monitoring for unlikely abnormalities²². She wrote about industrial process control, so applying this to a family calendar is my inference. A psychotherapist quoted by Business Insider in September warned that couples sharing an agent called Instinct could end up with a new argument about who is responsible for checking it²³.

“Proactive” covers four different things

Assistants come in four kinds, and marketing calls the last three proactive. Only one of them reaches the anticipating stage.

• Reactive: answers when asked.
• Persistent-context: remembers what you said, still waits to be asked.
• Scheduled or triggered: runs something you set up when a time or condition arrives, such as an email from the school.
• Inference-driven: decides on its own that something may need attention.

Most proactive features that vendors document as of 12 October 2026 are the third kind, and Meta says Muse makes unprompted suggestions²⁴. A trigger still depends on someone having anticipated the need when they set it up. OpenAI shows the drift. Its release notes of 17 June 2026 say Pulse, its proactive daily briefing, is being sunset as proactive updates move into scheduled tasks²⁵.

What the main assistants can do, and where

Several can send email, buy things or place calls, mostly with your approval and mostly in the US. The table shows what the main products say they do, as of 12 October 2026. Every capability below is a vendor claim, untested independently; Muse’s availability comes from press reports.

Disclosure: Subtext runs on models from Anthropic, which makes Claude⁴³.

We build Subtext in Berlin⁵, so I checked what a household in Germany can use. It can use ChatGPT’s and Claude’s scheduled tasks and ChatGPT’s cloud browser on their paid plans (Plus and above for ChatGPT), whose help pages state no EU restriction²⁶²⁷⁴¹, plus Alexa+ and Gemini for Home in early access³⁷⁴⁴. It can’t use Dots on Pro, Spark, Muse, Siri AI on an iPhone or Google’s AI calling, and Alexa’s auto-buy isn’t documented for Germany^253234373940. The most autonomous features in the debate are the ones a Berlin household can least try.

Can AI help with the mental load?

Possibly, though nobody has shown it. I found no peer-reviewed or preregistered study, whether a trial, diary study or long-term survey, measuring whether an AI assistant reduces household mental load or life admin.

The two closest studies are small and point the same way. In ParPal, a January 2026 preprint, 11 families (23 people) used an AI system for planning family learning for one to two weeks. Over half stopped carrying out the subtasks it assigned them while still using its help with individual tasks⁴⁵. In Croatia, five highly educated employed mothers used ChatGPT to organise their households for two weeks in early 2025. They reported that planning felt easier, and responsibility stayed with them⁴⁶. Neither study had a control group.

Part of the reason is that the research that measures household cognitive labour hasn’t been paired with an AI tool, and the teams building AI family tools haven’t used its measures. The measures are also thin. A 2023 review found no validated mental-labour scale¹⁵, and a validated “invisible family load” scale published the same year dropped its deciding and anticipating items⁴⁷. Two of the most autonomous products, Muse and Dots, only launched in September 2026²⁴²⁹. Settling it would take random assignment, several months, separate measures of who anticipates and who follows up, reports from both partners, and a log of the work of feeding and checking the assistant.

Three numbers you may have seen. Skylight’s marketing for its Sidekick feature promises “10 hours back (or more) every month” and states no method⁴⁸. A Motherly headline says AI cut one Zurich mother’s mental load by 97%, with no baseline⁴⁹. A Delphi-style exercise with 65 AI experts in the UK and Japan estimated that 39% of domestic task time could technically be automated within ten years, and its authors call the sample too small to generalise⁵⁰. That is marketing, an anecdote and a forecast. None of them measured anyone’s mental load.

Documented failures fall into a few patterns

The incidents on record with consumer AI agents fall into a few patterns. OpenAI has already withdrawn one such product. Its help page says ChatGPT agent is no longer available and names ChatGPT Work and the cloud browser as replacements⁵¹.

• Acting without the confirmation it was meant to ask for. In February 2025 a Washington Post journalist asked OpenAI’s Operator only to find the cheapest dozen eggs for delivery. It bought eggs through Instacart for $31.43 in total and reported $19.68⁵².
• Obeying hidden instructions. Security researchers showed in 2025 and early 2026 that text planted in a Google Calendar invite could make Gemini delete events, leak email or private meeting details, or operate smart-home devices⁵³⁵⁴. OpenAI says prompt injection is unlikely ever to be fully solved⁵⁵.
• Losing the instruction. A user told her OpenClaw agent to suggest inbox deletions and wait for her go-ahead. On her real inbox it began deleting, and she couldn’t stop it from her phone⁵⁶. That is one person’s account.
• Giving one person control over the others. In a 2019 study of 18 people in shared smart homes, whoever installed the devices held outsized control⁵⁷. A New York study of 39 survivors of intimate-partner violence and 50 professionals described abusers as authenticated users of their victims’ accounts⁵⁸. Both studies predate AI agents.

The first three are individual cases, and none of them shows how common such failures are. The fourth comes from studies of smart homes and shared accounts, not of AI agents. How to tell if an AI assistant is proactive goes through each pattern in more detail. If someone at home uses shared accounts or an assistant’s history to watch you, treat it as a safety problem and talk to a specialist service from a device that person cannot open. Helplines are listed under “When raising it isn’t safe” in how to ask your partner to share the mental load.

Every confirmation step hands a decision back

Most vendors’ main safeguard is a confirmation step. OpenAI says Dots need approval for purchases and that asking for a draft does not authorise sending it³⁰³¹; Google says Spark confirms before sends, purchases and form submissions³².

Confirmations protect you, and each one also returns a decision, and some checking, to you. So the household question is which decisions you want to keep. A family might let a system reorder the same cat food below a set price and still want to approve a partner’s birthday present. Amazon’s US auto-buy sits at the far end. By Amazon’s description, you pick an item and a target price, and Alexa later buys it with your default address and payment method, overnight if need be, with no confirmation step described³⁶.

More initiative has other costs. In a vignette experiment with 92 US participants with programming experience, unrequested AI help led to lower satisfaction with the system than help on request, through a dent in people’s sense of competence⁵⁹. It concerned programming help, so I treat it only as a hint that an assistant getting ahead of you can feel like being managed. OWASP’s security guidance and a paper from the UK’s competition authority both favour human approval for high-impact or high-risk actions⁶⁰⁶¹.

A short test for “proactive” claims

When a product says it takes things off your plate, I’d ask who notices that something needs doing, what it may do without asking, whether it asks only when your judgement is needed, what happens when it fails, and who can reconstruct what it did. Those questions draw on published governance and security guidance²⁰⁶⁰⁶¹. I go through them, plus two more (what the assistant knows and where it learned it, and whether anyone has measured the benefit), with dated product examples, in a seven-question checklist for proactive AI claims.

The simplest version is one question. After the first week of set-up, do you think about the responsibility less? An agent can save time and still fail that test by becoming one more thing to supervise.

Do AI family assistants create more work?

Nobody has measured that for AI assistants, and household technology has a long, disputed record. Joann Vanek concluded in 1974 that full-time housewives spent as much time on housework as their counterparts 50 years earlier⁶². Ruth Schwartz Cowan’s 1983 history More Work for Mother argued that appliances mostly replaced work once done by men, children and servants while standards rose⁶³. Using Australian time-use data from 1997, Michael Bittman, James Mahmud Rice and Judy Wajcman found that domestic technology rarely reduced women’s unpaid working time⁶⁴.

Jonathan Gershuny replied that a snapshot of appliance ownership can’t show historical change⁶⁵. With Teresa Harms he linked 1920s diaries of US rural homemakers to later surveys. Daily cooking, clearing and cleaning fell from 271 minutes to 108 by the 2000s, while care for children and adults rose from 35 minutes to 88; they credit technology for only part of the fall⁶⁶.

Both camps support a qualified reading. Technology cut specific task time, and new tasks and higher standards absorbed part of the gain. This is physical housework, so the analogy to mental load is indirect. The sociologist Marianne Cooper put the cautious version in a TIME opinion piece: “Because technology is laid over existing cultural beliefs and social relations, it rarely fixes social problems”⁶⁷.

Strengers and Kennedy argue that smart-home devices add work of their own, because someone has to choose, configure and repair them⁶⁸. In 22 Australian households studied in 2013 and 2014, 17 of 29 male and 6 of 28 female household members were counted as digital experts⁶⁹, though Jennifer Rode and Erika Shehan Poole dispute the account in which men are simply the household’s technical experts⁷⁰. AI agents need little technical set-up and a steady supply of context and checking, which to me looks like the monitoring stage. My untested guess is that this work drifts to whoever already holds the mental load.

The part of the load made of words

The part of the load I think AI fits best is the wording: the reply to a relative, the note to the school, the message after bad news, the no. Micaela di Leonardo described this kin work (visits, letters, calls, presents, cards, holiday gatherings) in 1987 as mostly done by women⁷¹, and a 1993 study found its breadth associated with depression in both women and men⁷². I cover these messages in why family messages feel like a second job, and reply avoidance in why you can’t reply to messages you want to answer.

Drafting tools work on the text. The noticing, remembering and deciding to reach out stay with you, and the research treats those as the core burden. I found no study showing that AI drafting reduces the anxiety, delay or emotional cost of personal messages. The measured gains are small and come from one study. In an online chat experiment with 219 pairs of crowdworkers, having smart replies available raised messages per minute by 10.2%, and people who used them tended to be rated as more cooperative by their partners⁷³.

There’s a social cost too. In experiments, mostly with US crowdworkers and hypothetical messages, people who know or suspect AI helped tend to trust the sender less; in one, trust in an email’s writer averaged 4.02 out of 5 when attributed to the sender and 3.51 when attributed to AI⁷⁴. In scenario studies of a work conflict, disclosing the help didn’t remove the penalty for an AI-assisted apology⁷⁵, and in two US experiments with 647 and 654 adults, recipients who weren’t told did not suspect AI⁷⁶. The research doesn’t settle the ethics; my position is in does AI make your texts sound like a robot.

Where Subtext fits, and when you shouldn’t use it

As of 12 October 2026, Subtext helps you draft, rewrite, proofread and understand messages⁴³. You paste or upload what you want help with, as text, a screenshot of the thread or a voice note⁷⁷. It suggests up to three versions, each marked safe to send or with a warning, based on Subtext’s own estimate⁷⁷⁷⁸. It can also sum up a long message you received as what the sender wants from you and the points that need a response⁷⁹. You copy the version you want and send it yourself. Subtext sends nothing for you, isn’t a keyboard and connects to none of your email, calendar or messaging accounts; it only reads what you hand it⁴³⁷⁸. More in what Subtext is and who it’s for.

Measured against the mental load, that is help with one narrow task, composing a reply to a relative or a tense school email, and nobody has tested whether it makes even that lighter. Subtext carries none of the load itself. It doesn’t notice that a reply is owed, decide what you want to say or check what happens next. Our own research page is plain about the evidence: “No study on this page tested Subtext”⁸⁰. It adds that the tags, feeling word and safe-to-send estimate haven’t been measured against how real readers took real messages⁸⁰. I found no independent review of Subtext either.

When I’d skip it:

• When the hard part is remembering or planning. Subtext does neither. A shared list or the reminder tools in the best apps for ADHD adults who take forever to reply will serve you better.
• When your problem is volume. Email triage and auto-reply tools are a different category; I sorted them in the best AI reply app depends on whether your problem is volume or stakes.
• When the words belong to someone who wouldn’t want them shared. What you submit goes to processors in the US, including Anthropic and, for voice recordings, Deepgram. It isn’t used for training, we have no zero-retention arrangement with any provider, and conversations are deleted five days after last use (90 if pinned). Our Terms make you responsible for confirming you may share someone else’s message⁴³. I compared policies in which AI writing assistants train on your messages.
• When a tool on your phone already does enough. The drafting tools built into phones and mail apps may be all you need; I set out where they beat us in every phone rewrites your messages now.
• When the person would be hurt to learn AI helped. Given the research above, write that one yourself.

For the one message you keep rewriting, Subtext reads your draft, marks the words that could land badly and offers up to three versions, and you send the one you choose. Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

Where I landed

Can AI reduce the mental load? As of 12 October 2026, assistants can take over execution and some option-finding, and watch for conditions you define in advance. Anticipating and following up still mostly begin and end with a person, who now also supervises the agent. I found no study that measured whether the total goes down, and neither of the two closest small studies showed that it does.

The gendered metaphor names a real distinction, and I’d retire it. “Outcome-owning” and “instruction-following” describe the same thing without turning a division of labour into a personality.

If you try an assistant anyway, start with one bounded job where you’d spot a failure quickly, give it the narrowest permissions that work, and ask who will check it. If the answer is the person who already carries the load, talk about that first; how to explain the mental load to a partner may help.

Questions people ask

What does an AI family assistant app do? By the vendors’ own accounts, mostly capture and reminders: Cozi Max turns invitations in your inbox into calendar events⁸¹ and Ollie for Meals suggests meal plans⁸². In Rose Hirst’s test of two such apps, both needed her to notice tasks and feed them in¹⁹.

Why are AI assistants female by default? Brookings documented in 2020 that the four major voice assistants all launched with female-sounding default voices⁸³. Heather Suzanne Woods argued that feminine personas draw on roles like caretaker, mother and wife, and UNESCO made a similar case⁸⁴⁸⁵. Evidence on effects is thin. One experiment with 234 people found implicit stereotyping and no difference in trust between gendered and gender-ambiguous voices⁸⁶.

Can Subtext take the mental load off me? No. It helps with the wording of one message at a time, and you still notice, decide and send. Our own research page says that none of the studies it lists tested Subtext⁸⁰.

Sources

Numbered in order of first appearance. Product facts were checked on 12 and 13 October 2026, and help pages change often. Where I couldn’t read a primary source, the note says what I read instead.

1. Silicon Valley keeps building me a husband when what I really want is a wife lol (post on X). Sonia Baschez, X. 6 October 2026, 00:30 UTC (5 October, US Pacific time). (Social post; text and date confirmed on 13 October 2026. The view count is from the Ben’s Bites embed.)
2. Text is so 2023. Ben Tossell, Ben’s Bites newsletter. 8 October 2026. (Checked 12 October 2026. Newsletter commentary; the view count comes from its embed of the post, captured at an unknown time.)
3. I don’t need an AI agent to be an executive assistant. I need it to be a wife. Katie Notopoulos, Business Insider. 9 October 2026. (Checked 13 October 2026. First-person column; cites no studies.)
4. I don’t need an AI agent to be an executive assistant. I need it to be a wife. Katie Notopoulos, Business Insider, syndicated by Yahoo Style Canada. 9 October 2026, 09:05 UTC. (Checked 12 October 2026. Syndicated copy of the column.)
5. About Subtext: who builds the app and why. Subtext (Samet Durgun and Anton Licht). Undated. (Checked 12 October 2026. Our own page.)
6. ‘AI Wife’ Concept Reveals Demand for Proactive Personal Assistants. HyperAI (AI-generated aggregation, no author). About 10 October 2026. (Cited only to show the page exists; it says it was assembled by AI and quotes no one, so it is no evidence of reach or reactions.)
7. A DIY Coder Created a Virtual AI ‘Wife’ Using ChatGPT. Vice. January 2023. (Press. Cited only for the older companion meaning of the phrase.)
8. Judy Brady Syfers. Wikipedia contributors. Accessed 12 October 2026. (Encyclopedia summary. The link between the essay and the column is my reading; the column does not cite Brady.)
9. The Smart Wife: Why Siri, Alexa, and Other Smart Home Devices Need a Feminist Reboot. Yolande Strengers and Jenny Kennedy, MIT Press (Penguin Random House listing). 2020. (Book; publisher description read, full text not read.)
10. 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. Crossref lists Harvard University as funder and showed no correction on 12 October 2026.)
11. How Couples Share “Cognitive Labor” and Why it Matters. Allison Daminger, Behavioral Scientist. 19 September 2019. (Essay by the study’s author; not peer reviewed. Source of the sample description.)
12. What’s on Her Mind: The Mental Workload of Family Life. Allison Daminger, Princeton University Press. 9 September 2025. (Book; publisher page read, which gives no sample counts.)
13. Mental load in women and mothers: Causes of distress, personality traits, and psychopathology. Stefanie M. Jungmann, Stella F. Garbe Huedo, Karlotta A. S. Jacobs and Nina K. Pothorn, PLOS One 21(9): e0356255. 2 September 2026. (Peer reviewed; two online convenience samples (368 women; 1,871 mothers), cross-sectional; anxiety and depression were measured in the first sample only. Authors report no specific funding and no competing interests.)
14. Mothers and Mental Labor: A Phenomenological Focus Group Study of Family-Related Thinking Work. Lindsey G. Robertson, Tamara L. Anderson, M. Elizabeth Lewis Hall and Christina Lee Kim, Psychology of Women Quarterly 43(2): 184 to 200. Online 11 March 2019. (Peer reviewed; 25 mothers in seven focus groups. Funding not shown on the publisher page.)
15. 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 individuals or couples. Partly funded by FernUniversität in Hagen; no conflicts declared.)
16. Who shoulders the mental work? Evidence on the distribution of mental work in different-sex couples in Germany. Lena Hipp, Olga Leshchenko and Armin Sauermann, Social Problems (advance article). 10 February 2026. (Peer reviewed; national probability sample, early 2023, N = 2,675 people aged 23 to 65 in cisgender different-sex couples; cross-sectional, self-reported, one partner per couple.)
17. Take a Load Off? Not for Mothers: Gender, Cognitive Labor, and the Limits of Time and Money. Ana Catalano Weeks, Helen Kowalewska and Leah Ruppanner, Socius 11. 2025. (Peer reviewed; online quota survey (Dynata), February to March 2023, 2,133 partnered heterosexual US parents analysed of 3,000 surveyed. Funded by the Australian Research Council.)
18. Do Mothers Do More of the “Mental Load”? Gender Divisions in Daily and Episodic Cognitive Household Work. Ana Catalano Weeks, University of Bath Institute for Policy Research blog. 16 December 2024. (Co-author’s blog summary of a survey of about 3,000 US parents; the paper gives the single-parent subsample as 460.)
19. Can you actually outsource your mental load? I desperately tried to. Rose Hirst, The Spinoff. 13 September 2026. (First-person journalism; cites no studies. Quotes confirmed on the live page on 13 October 2026.)
20. Practices for Governing Agentic AI Systems. Yonadav Shavit and colleagues, OpenAI. Undated (references to late 2023). (Vendor white paper; not peer reviewed.)
21. Intelligent AI Delegation. Nenad Tomašev, Matija Franklin and Simon Osindero, Google DeepMind. 12 February 2026. (Preprint.)
22. Ironies of automation. Lisanne Bainbridge, Automatica 19(6): 775 to 779. November 1983. (Peer-reviewed argument paper about industrial process control; applying it to households is my inference.)
23. Couples are using a viral AI agent to do their emotional labor. Julia Pugachevsky, Business Insider, syndicated by Yahoo Style Canada. 27 September 2026. (Checked 12 October 2026. Press; couples’ accounts are anecdotes.)
24. Introducing Muse: The World’s First Personal AI Agent Built for Everyone. Meta Newsroom. 8 September 2026, modified 30 September 2026. (Checked 12 October 2026. Vendor announcement; “World’s First” in the title is a marketing claim.)
25. ChatGPT release notes. OpenAI Help Center. Updated about 9 October 2026. (Checked 12 October 2026. Vendor page.)
26. Scheduled tasks in ChatGPT. OpenAI Help Center. Updated about 2 October 2026. (Checked 12 October 2026. Vendor page.)
27. Using cloud browser in ChatGPT. OpenAI Help Center. Updated 28 August 2026. (Checked 12 October 2026. Vendor page.)
28. ChatGPT Work and Codex. OpenAI Help Center. Undated. (Checked 12 October 2026. Vendor page.)
29. Introducing dots. OpenAI. 29 September 2026. (Checked 12 October 2026. Vendor announcement; capabilities are company claims.)
30. How we build safety, security, and privacy into dots. OpenAI. 29 September 2026. (Checked 12 October 2026. Vendor page.)
31. Getting started with your dot. OpenAI Help Center. Updated about 11 to 12 October 2026. (Checked 12 October 2026. Vendor page.)
32. Use Gemini Spark to manage your tasks & workflows in Gemini Apps. Google, Gemini Apps Help. Undated. (Checked 12 October 2026. Vendor page; Google calls Spark an experimental feature in early development.)
33. Meta’s Muse AI Agent Is Now Available in Canada. Austin Blake, iPhone in Canada. 18 September 2026. (Press.)
34. Meta Muse kommt aufs iPad, nach Deutschland aber weiterhin nicht. Jan Gruber, Notebookcheck. 8 October 2026. (Press (German). Meta names no European date.)
35. Alexa+ now available to everyone in the US. Amazon (About Amazon). 4 February 2026, updated 1 September 2026. (Checked 12 October 2026. Vendor page.)
36. Getting started with Alexa+: Simplify home organization with recipes, calendars, and reminders. Amazon (About Amazon). 4 February 2026. (Checked 12 October 2026. Vendor page; describes target-price auto-buy and no confirmation step or cancellation window.)
37. Alexa+ startet in Deutschland. Amazon (About Amazon Deutschland). 7 May 2026. (Checked 12 October 2026. Vendor page. Amazon’s German help pages could not be read.)
38. Alexa+ arrives in India as Amazon’s next-generation AI assistant expands internationally. Amazon (About Amazon). 19 March 2026, modified 16 September 2026. (Checked 12 October 2026. Vendor page; still lists Germany as Early Access.)
39. How to get Siri AI. Apple Support. 22 September 2026. (Checked 12 October 2026. Vendor page.)
40. Ask Gemini to handle your everyday phone calls. Google, Gemini Apps Help. Undated. (Checked 12 October 2026. Vendor page; US only, Pixel 11 with a US SIM, English, Phone by Google public beta.)
41. Schedule recurring tasks in Claude Cowork. Anthropic, Claude Help Center. Undated. (Checked 12 October 2026. Vendor page. Anthropic is one of Subtext’s processors.)
42. Use Google Workspace connectors. Anthropic, Claude Help Center. Undated. (Checked 13 October 2026. Vendor page.)
43. Terms, Privacy and your account. Subtext (Anton Licht, sole proprietor). Terms last updated 26 July 2026; Privacy Policy last updated 10 October 2026. (Checked 12 October 2026. Our own legal page.)
44. Get started with Gemini for Home features. Google, Google Home and Nest Help. Undated. (Checked 12 October 2026. Vendor page; early access in 20 countries including Germany.)
45. ParPal: A Human-Centred AI System for Multi-Actor Planning and Collaboration in Family Learning. Chen, Xie, Li, Lin, Zhang, Wang, Huang, Yu, Metoyer, Hua and Chawla. 28 January 2026. (Preprint; 11 families (23 people), one to two weeks, no validated load measure, no control group.)
46. Invisible household labour and digital solutions: Gender divisions in mental labour and the role of artificial intelligence. Ksenija Klasnić, Anita Dremel, Krešimir Žažar and Đurđica Degač, conference proceedings, pp. 137 to 154. 2025. (Conference proceedings, review process not stated; the two-week test had five mothers and no control group. Funded by the University of Zagreb Faculty of Humanities and Social Sciences.)
47. Who’s Remembering to Buy the Eggs? The Meaning, Measurement, and Implications of Invisible Family Load. Julie Holliday Wayne, Maura J. Mills, Yi-Ren Wang, Russell A. Matthews and Marilyn V. Whitman, Journal of Business and Psychology 38(6): 1159 to 1184. Online 30 May 2023. (Peer reviewed; scale developed on Prolific and MTurk panels. Funding not shown in the pages read.)
48. Meet Sidekick, powerful planning for busy families. Skylight. Undated. (Checked 12 October 2026. Marketing page; no method stated.)
49. Mom uses AI to cut 97% of her mental load and spend more time with her kids. Himanshi Bahuguna, Motherly. 11 August 2025. (Press; one person, no baseline or method (anecdote).)
50. The future(s) of unpaid work: How susceptible do experts from different backgrounds think the domestic sphere is to automation? Vili Lehdonvirta, Lulu P. Shi, Ekaterina Hertog, Nobuko Nagase and Yuji Ohta, PLOS ONE 18(2): e0281282. 22 February 2023. (Peer reviewed; expert forecast from 65 AI experts in the UK and Japan, 2020 to 2021. Funded by UKRI, JST and ERC.)
51. ChatGPT agent. OpenAI Help Center. Updated about 15 September 2026. (Checked 12 October 2026. Vendor page; gives no removal date.)
52. I let ChatGPT’s new ‘agent’ manage my life. It spent $31 on a dozen eggs. Geoffrey A. Fowler, The Washington Post, syndicated by Anchorage Daily News. 7 February 2025. (Press; one journalist’s test.)
53. Invitation Is All You Need: Invoking Gemini for Workspace Agents with a Simple Google Calendar Invite. Or Yair, Ben Nassi and Stav Cohen, SafeBreach Labs. 6 August 2025. (Security researchers’ write-up; Google described mitigations.)
54. Weaponizing Calendar Invites: A Semantic Attack on Google Gemini. Liad Eliyahu, Miggo Security. 19 January 2026. (Security researchers’ write-up.)
55. Continuously hardening ChatGPT Atlas against prompt injection attacks. OpenAI. 22 December 2025. (Vendor post.)
56. A quote from Summer Yue. Simon Willison (weblog). 23 February 2026. (User anecdote reported on a weblog.)
57. Who’s In Control? Interactions In Multi-User Smart Homes. Christine Geeng and Franziska Roesner, CHI 2019. 2019. (Peer-reviewed conference paper; interviews and experience sampling with 18 people, before AI agents.)
58. A Stalker’s Paradise: How Intimate Partner Abusers Exploit Technology. Diana Freed, Jackeline Palmer, Diana Minchala, Karen Levy, Thomas Ristenpart and Nicola Dell, CHI 2018. 2018. (Peer-reviewed conference paper; 39 survivors and 50 professionals in New York, before AI agents.)
59. When AI-Based Agents Are Proactive: Implications for Competence and System Satisfaction in Human-AI Collaboration. Diebel, Goutier, Adam and Benlian, Business & Information Systems Engineering 68(2): 259 to 278. Online 3 January 2025. (Peer reviewed; vignette experiment, 92 US participants with programming experience recruited through Prolific. The article acknowledges German Research Foundation funding; the Crossref record lists TU Darmstadt.)
60. LLM06:2025 Excessive Agency. OWASP GenAI Security Project. 2025. (Security guidance.)
61. Agentic AI and consumers. UK Competition and Markets Authority. 9 March 2026. (Regulator paper.)
62. Time Spent in Housework. Joann Vanek, Scientific American 231(5): 116. November 1974. (Magazine analysis of time-use studies.)
63. More Work for Mother: The Ironies of Household Technology from the Open Hearth to the Microwave. Ruth Schwartz Cowan, Basic Books (Hachette listing). 1983. (Book; publisher description read.)
64. Appliances and their impact: the ownership of domestic technology and time spent on household work. Michael Bittman, James Mahmud Rice and Judy Wajcman, British Journal of Sociology 55(3): 401 to 423. September 2004. (Peer reviewed; Australian Bureau of Statistics 1997 Time Use Survey. Sample size not checked.)
65. Domestic equipment does not increase domestic work: a response to Bittman, Rice and Wajcman. Jonathan Gershuny, British Journal of Sociology 55(3): 425 to 431. 2004. (Peer-reviewed reply; publication record read.)
66. Housework Now Takes Much Less Time: 85 Years of US Rural Women’s Time Use. Jonathan Gershuny and Teresa Attracta Harms, Social Forces 95(2): 503 to 524. 2016. (Peer reviewed; links 1920s rural homemaker diaries to later surveys; 1920s data lack equipment records.)
67. Can AI Free Women from the Mental Load of Caregiving? Marianne Cooper, TIME (Ideas, opinion). 5 August 2026. (Opinion; cites no study testing AI’s effect on mental load.)
68. Yolande Strengers and Jenny Kennedy on The Smart Wife. Rotman Institute for Gender and the Economy (GATE), University of Toronto. Undated. (Event summary.)
69. Digital housekeepers and domestic expertise in the networked home. Jenny Kennedy, Bjorn Nansen, Michael Arnold, Rowan Wilken and Martin Gibbs, Convergence 21(4): 408 to 422. 24 April 2015. (Peer reviewed (author-accepted manuscript); 22 Australian households, 2013 and 2014.)
70. Putting the Gender Back in Digital Housekeeping. Jennifer A. Rode and Erika Shehan Poole, Proceedings of GenderIT ’18, pp. 79 to 90. May 2018. (Conference paper; abstract only read.)
71. The Female World of Cards and Holidays: Women, Families, and the Work of Kinship. Micaela di Leonardo, Signs 12(3): 440 to 453. 1 April 1987. (Peer reviewed; full text paywalled, bibliographic record checked.)
72. 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 (AGRIS record). 1993. (Peer reviewed; abstract read.)
73. 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 (publisher correction 3 October 2023). (Peer reviewed; randomised experiments with online workers. The correction added the omitted funding statement (NSF).)
74. 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 conference paper; Wizard-of-Oz survey of 147 MTurk workers plus 10 interviews. Funding not stated in the text read.)
75. 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 and funding not available on the record page.)
76. 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.)
77. How Subtext works: a real message, start to finish. Subtext. Undated. (Checked 12 October 2026. Our own page.)
78. Subtext app: AI writing assistant for texts and replies. Subtext. Undated. (Checked 12 October 2026. Our own page.)
79. Subtext: AI Writing Assistant. Subtext (seller Anton Licht), Apple App Store listing. Undated. (Checked 12 October 2026. Our own store listing.)
80. The research behind Subtext. Subtext. Undated. (Checked 12 October 2026. Our own page.)
81. Cozi Max: Cozi Family Organizer. Cozi. Last modified 14 January 2026. (Checked 12 October 2026. Vendor page.)
82. Ollie for Meals: Meal Planning. Ollie, Apple App Store listing. Undated. (Checked 12 October 2026. Vendor store listing.)
83. How AI bots and voice assistants reinforce gender bias. Caitlin Chin-Rothmann and Mishaela Robison, Brookings Institution. 23 November 2020. (Think-tank report.)
84. Asking more of Siri and Alexa: feminine persona in service of surveillance capitalism. Heather Suzanne Woods, Critical Studies in Media Communication 35(4): 334 to 349. 2 July 2018. (Peer-reviewed critical analysis; read through a summary by the National Communication Association.)
85. I’d blush if I could: closing gender divides in digital skills through education. Mark West, Rebecca Kraut and Han Ei Chew, UNESCO and EQUALS Skills Coalition. 2019. (Policy report.)
86. Female by Default? Exploring the Effect of Voice Assistant Gender and Pitch on Trait and Trust Attribution. Suzanne Tolmeijer, Naim Zierau, Andreas Janson, Jalil Sebastian Wahdatehagh, Jan Marco Leimeister and Abraham Bernstein, Extended Abstracts of CHI 2021. 8 May 2021. (Extended abstract; online experiment with 234 participants. Funded by the Swiss National Science Foundation.)