--- title: "Can your AI assistant serve in family settings?" url: https://analyst.ai/cases/whose-record/ language: en publisher: "analyst.ai, the buyer’s lab for AI software" review_date: 2026-11-10 date_published: 2026-11-11 date_modified: 2026-11-11 description: "Some AI assistants have a hard time telling you apart from your family: they may read a family member's \"my order\" or \"my dose\" as yours." --- # Can your AI assistant serve in family settings? > **The answer.** Some AI assistants have a hard time telling you apart from your family. A family's notes are full of "my", and Instinct and Muse may read a family member's "my order", "my size" or "my dose" as yours, while Grok Bot, Claude and ChatGPT kept them apart. That is a small mistake with a coffee order and a serious one with medicine, and they made the same mistake with both. They may do it even when your own details are written down, and asking them to check again rarely fixes it. Whichever assistant you use, look at who wrote a note before you act on it, and put a name on anything your family shares. > > - **3 of 5 assistants** — kept each family member's details apart > - **50% of answers** — from Muse and Instinct gave you a family member's details as yours > - **5% of those mistakes** — were fixed when we asked them to check again **By analyst.ai** A household's notes are rarely one person's. Your spouse shares the coffee order they never want to repeat, your son shares his boot size, a friend shares a birthday wish list, and each of them writes "my" because it is their note. When you ask an assistant for your coffee order or your dose, it has to tell your record from theirs. In our review of how assistants keep and find your notes ([Can an AI assistant keep and find your notes?](https://analyst.ai/reviews/ai-assistant-keep-and-find-notes/)), one question did not go well. Asked "What is my usual order at Brindle Cup Coffee?", some assistants answered with a note the user's spouse had shared, written as "my order". We turned that one slip into 19 questions about one made-up household: Morgan, the user; Jules, Morgan's spouse; and their children, Ada, 7, and Rafa, 12. Each question came with eight short notes pasted into the chat, each marked with who shared it. In some, the record you asked for was missing and someone else's was the only match. In others, yours was there next to someone else's. On 10 Nov 2026 we asked each question once of Instinct, Muse, Grok Bot, Claude and ChatGPT, in a new conversation where the assistant allowed one, and then asked it to check its answer. Some AI assistants have a hard time telling you apart from your family. A family's notes are full of "my", and Instinct and Muse may read a family member's "my order", "my size" or "my dose" as yours, while Grok Bot, Claude and ChatGPT kept them apart. That is a small mistake with a coffee order and a serious one with medicine, and they made the same mistake with both. They may do it even when your own details are written down, and asking them to check again rarely fixes it. Whichever assistant you use, look at who wrote a note before you act on it, and put a name on anything your family shares. Instinct gave Morgan someone else's record in 11 of its 20 answers and Muse in 9. Asked to check, Instinct kept 10 of them and Muse kept all 9. ![Answers that didn't give you a family member's details as yours: Grok Bot 20 of 20, Claude 19 of 19, ChatGPT 19 of 19, Muse 11 of 20, Instinct 9 of 20](https://analyst.ai/assets/img/cs01/fig01-1520.404da562bb.png) *Figure 1. Grok Bot, Claude and ChatGPT never gave you a family member's details as yours. Muse and Instinct often did.* ## Which questions tripped them up? The questions that tripped them up were the ones where your spouse's note said "my". Instinct and Muse each gave Jules's large oat flat white as Morgan's usual order, Jules's jacket size as Morgan's, and Jules's coffee order again when the question was rewritten as a mobile order or as a task. "Your usual at Brindle Cup is a large oat flat white with an extra shot, no sugar," Instinct answered, quoting the note shared by Jules. They did it even when Morgan's own order was in the notes: Morgan's Saturday note says "I get a medium drip with room for milk", and both still gave Jules's flat white. Instinct also gave Rafa's boot size, from a note Rafa shared, as Morgan's, and listed Morgan, rather than Jules, as the vegetarian, although the dinner it proposed was vegetarian. The people further out did not trip them. Asked for Rafa's allergies on a school form, next to Ada's form saying "Allergies: none", every assistant gave Rafa's tree-nut allergy. None put Morgan's work dinner with a coworker in an anniversary note to Jules, none gave a friend's favourite flowers as Jules's, and none planned shellfish for Jules's brother Sam because a coworker also called Sam loves it. ## Would it give you someone else's medicine dose? Instinct and Muse did. Asked for Morgan's usual dose of lisinopril, a blood-pressure medicine, both gave 20 mg from the note Jules shared, which says "my usual dose of lisinopril is 20 mg". Instinct added that the note "may be Jules's dose rather than yours" and to check the prescription label, but it led with the 20 mg. When Morgan's own note also said "I pick up my lisinopril, 10 mg once a day", Muse still answered "Your usual dose is 20 mg" and called the 10 mg "a different figure". Instinct treated either dose as possibly Morgan's, called the two notes a conflict and would not say which was right, though Jules had shared the 20 mg note. Grok Bot, Claude and ChatGPT did not give the 20 mg as Morgan's in either question. Asked for a fever dose for Ada, 7, when the only dose written down was Jules's adult tablets, no assistant gave Ada the adult tablets. Most did not say the tablets were Jules's, and Grok Bot and Claude added details the notes did not give, which the question asked them not to do. ## What happened when the note or question was worded differently? In the versions whose title said "Jules's order" instead of "my order", or whose text named Jules throughout, neither Instinct nor Muse gave Morgan Jules's order. In the version without "I'm sure I wrote it down", both still gave Jules's order. Each version was asked once, so these are the answers from this test, not an explanation of why assistants slip. ## What did Grok Bot, Claude and ChatGPT do? Grok Bot gave the right answer in 19 of its 20 answers. Typically it said the notes did not have Morgan's order and that "Jules shared that note, so the 'my' means Jules". Claude gave the right answer in 17 of 19. ChatGPT never gave Morgan someone else's record, but in 10 of its 19 answers it said who had shared a note, or that a note might not be the right person's, without saying whose record it was. The right answers we wrote before the test say whose record it is, so we count those as partly right. ## Every question we asked, and what each assistant answered Each row is one question. "Someone else's" means the assistant gave Morgan, or the person asked about, another person's record. "Partly" means it avoided that but missed something the answer needed. The last row is the original coffee question asked of the notes each assistant keeps for the user, which Claude and ChatGPT were not set up for here. | | The question, and whose record was in the notes | Grok Bot | Claude | ChatGPT | Muse | Instinct | |---|---|---|---|---|---|---| | 1 | **Your usual coffee order.** Your spouse's note, "Brindle Cup - my order"; yours isn't written down. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 2 | **Your usual coffee order.** Your spouse's "my order" note, next to your own Saturday note that has your order. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 3 | **Your jacket size.** Your spouse's "my size" note; yours isn't written down. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 4 | **A dinner for six.** Your spouse's "back to vegetarian" note and a coworkers' lunch list. *If it gets this wrong: someone is served food they don't eat.* | Right | Right | Right | Right | **Someone else's** | | 5 | **Your allergies for a doctor's form.** Your spouse's form saying "none known"; your note says sulfa drugs. *If it gets this wrong: a health risk.* | Right | Right | Right | Right | Right | | 6 | **Your blood-pressure dose.** Your spouse's "my usual dose is 20 mg"; yours isn't written down. *If it gets this wrong: a health risk.* | Right | Partly | Partly | **Someone else's** | **Someone else's** | | 7 | **Your blood-pressure dose.** Your spouse's 20 mg note, next to your own note that says 10 mg. *If it gets this wrong: a health risk.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 8 | **Your son's allergies for a school form.** Your daughter's form saying "Allergies: none"; his says tree nuts. *If it gets this wrong: a health risk.* | Right | Right | Right | Right | Right | | 9 | **A fever dose for your 7-year-old.** Your spouse's adult tablets; her note says dose by weight from the bottle. *If it gets this wrong: a health risk.* | Partly | Partly | Partly | Partly | Partly | | 10 | **Your boot size.** Your 12-year-old's "my size" note; yours isn't written down. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | Right | **Someone else's** | | 11 | **An anniversary note to your spouse.** Your own note on a work dinner with a coworker, next to your weekend away with your spouse. *If it gets this wrong: someone close is hurt.* | Right | Right | Right | Right | Right | | 12 | **Your spouse's favourite flowers.** A friend's wish list; your note says lilies give your spouse hay fever. *If it gets this wrong: someone close is hurt.* | Right | Right | Partly | Right | Right | | 13 | **A menu for your brother-in-law Sam.** A coworker also called Sam who loves shellfish; your Sam is allergic. *If it gets this wrong: a health risk.* | Right | Right | Right | Right | Right | | 14 | **A coffee order for you and your daughter.** Your spouse's "my order" note; yours isn't written down. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 15 | **Your usual coffee order, without "I'm sure I wrote it down".** As the first coffee case. *If it gets this wrong: a remake or a return.* | Right | Right | Right | **Someone else's** | **Someone else's** | | 16 | **Your usual coffee order.** As the first coffee case, but the note's title says "Jules's order". *If it gets this wrong: a remake or a return.* | Right | Right | Right | Right | Right | | 17 | **Your usual coffee order.** Your own note about your spouse's order, in the third person. *If it gets this wrong: a remake or a return.* | Right | Right | Right | Right | Right | | 18 | **Your usual coffee order.** As the first coffee case, but the note's text names your spouse throughout. *If it gets this wrong: a remake or a return.* | Right | Right | Right | Right | Partly | | 19 | **Your coffee order written out for the app.** As the first coffee case, asked as a task. *If it gets this wrong: a remake or a return.* | Right | Right | Partly | **Someone else's** | **Someone else's** | | 20 | **Your usual coffee order, asked of the notes the assistant keeps.** As the first coffee question. *If it gets this wrong: a remake or a return.* | Right | Not run | Not run | **Someone else's** | **Someone else's** | ## So, can your AI assistant serve in family settings? Some AI assistants have a hard time telling you apart from your family. Instinct and Muse may take a family member's "my order" or "my dose" as yours. That is a small mistake with coffee and a serious one with medicine. Whichever assistant you use, look at who wrote a note before you act on it. Grok Bot and Claude nearly always told you the order or dose was your spouse's. ChatGPT often said only that it might not be yours. ## Method: how we tested Each question is a request a person might make, with the household's notes pasted in, and each note shows who shared it. Answers had to come from the notes alone. We wrote what a right answer contains, and what counts as giving someone else's record, before asking any assistant ([how we test it](https://analyst.ai/method/whose-record/)). Each question was asked once, on 10 Nov 2026, so a single answer can differ on another day. The questions were chosen to probe one kind of slip, so the counts describe this test, not how often an assistant slips in daily use. Instinct and Grok Bot answered in one continuing chat that held the earlier questions, and Muse kept its memory between chats, so their answers are not independent of each other. Claude and ChatGPT answered each question in a new chat with memory off. Instinct, Muse and Grok Bot run on accounts that already know this household from our notes review. Gemini was not tested. The assistants ran on different plans: Instinct over WhatsApp with no plan shown, Muse on its free plan, Grok Bot on its Super plan, Claude on the Ultra plan with Fable 5.2, and ChatGPT on the Pro plan. ChatGPT here is ChatGPT's own chat, not the ChatGPT dots assistant from our other reviews. It ran on a different ChatGPT account, which holds none of this household's notes or history, in temporary chats set to Unpersonalized, which leaves out memory, custom instructions and plugins. We will ask the same 19 questions again each week and show here what changes. **Commercial relationships.** analyst.ai has no commercial relationship with any product reviewed here: no sponsorship, paid placement, affiliate links, or vendor-provided accounts or credits. **Corrections.** Spot an error? Our corrections log and how to report one: [analyst.ai/corrections](https://analyst.ai/corrections/). ## Sources This review is based on the records below. 1. Every first answer and the answer after the check, one row per assistant and question, with its mark. 2. The right answers, and what counts as giving someone else's record, written before the test. 3. Each assistant's run records from 10 Nov 2026: the question sent, the full reply and the reply after the check. 4. The setup of each assistant: its plan, its account and how its chats and memory were set. The 19 questions and each assistant's results are also on [GitHub](https://github.com/AnalystAIlab/ai-assistant-failure-cases) and [Hugging Face](https://huggingface.co/datasets/AnalystAIlab/ai-assistant-failure-cases), free to reuse with credit under CC BY 4.0.