An AI Update Broke an Autistic Shopper’s 23-Item Routine

The grocery notebook stayed open on the kitchen table while she packed her bags. On the right-hand page was a list of 23 items, copied from her phone and grouped in the order she expected to encounter them: produce first, refrigerated food later, household supplies near the end.

She had made lists before using an AI assistant. They were usually incomplete, crowded with revisions or arranged according to when she remembered each item. That meant crossing the store several times, making fresh decisions under bright lights while carts stopped around her and announcements interrupted what she was trying to hold in mind.

She is autistic, and crowded grocery stores can demand more processing than she has available. A missing ingredient is manageable at home. Discovering it in the store, while deciding whether another size or unfamiliar substitute will work, can stall the whole trip.

For 14 months, the assistant reduced those decisions. She gave it a loose account of the coming week, including the meals she could imagine eating, what remained in the pantry and how much she wanted to spend. The system turned that conversation into quantities and categories, then arranged the items according to the path she normally took through the store.

The exchange mattered because she did not begin with structured data. She might remember halfway through that lunch needed to travel well, or that a dinner should leave enough food for the next day. The assistant could incorporate the new condition into the existing plan without requiring her to rebuild a spreadsheet or explain everything to another person.

She still checked its work. Once it suggested a quantity that would have left her with far too many tortillas. Another time it treated an ingredient she disliked as interchangeable with one she used often. Those were contained errors.

She corrected them, and the conversation continued.

The grocery notebook became the final authority. The assistant could propose, infer and reorganize, but nothing counted until she copied it onto the page.

The answer that would not stay short

The change arrived after a software update in August 2024. The company described an improved system, but the assistant began responding differently to the same kind of request.

Instead of producing the compact list she expected, it added meal descriptions, storage suggestions and encouraging commentary. Some responses grouped food by recipe. Others used broad departments that did not match her route. When she asked for a correction, the assistant occasionally generated a fresh list with different quantities rather than editing the version already on screen.

The problem was not that the prose sounded cold. It sounded more attentive than before, which made the disruption harder to name. The assistant acknowledged her preferences and explained its reasoning, but she needed it to stop explaining and preserve the list.

During the first trip after the update, she brought a page with 27 items. Several appeared under new headings, and one ingredient was duplicated in separate parts of the list. She noticed the duplicate in the store, then began checking whether other lines had changed. The notebook no longer narrowed her attention.

It created another comparison task.

She left without two items and spent $18.40 more than planned, partly because she chose familiar prepared food when the list stopped feeling reliable. That purchase was useful, not a failure. It also meant the assistant had ceased doing the specific job around which she had built the trip.

At home, she placed the receipt inside the grocery notebook. It remained there for weeks, marking the page where the routine changed.

What the update changed

A conversational AI assistant does not usually fetch a single fixed answer attached to a request. It generates a response piece by piece, using the current conversation, any available saved preferences and instructions supplied by the company, while choosing among multiple plausible continuations.

That design lets it turn an unstructured account of a week into a useful shopping plan. It also means identical prompts can produce different wording or organization, especially when the underlying model, system instructions or product settings change. A preference such as “keep it short” is interpreted alongside other instructions rather than enforced like a locked spreadsheet rule.

She had assumed the assistant remembered the routine as a person might remember a repeated favor. In practice, its apparent continuity came from a mixture of chat history, saved memory features and the model’s ability to infer what she meant. The relationship felt stable because those parts had produced stable behavior for more than a year.

After the update, the assistant still knew several of her preferences. It could mention her usual budget and avoid foods she had rejected before. Yet it no longer gave those facts the same weight when formatting the answer, and there was no setting labeled with the outcome she needed: preserve the exact grocery-list structure that already works.

This was the machine’s part in the story. Its conversational flexibility had made the accommodation possible, since it could translate incomplete thoughts into a coherent plan. The same flexibility meant that a revised model could reinterpret familiar instructions without deleting them or displaying an obvious error.

Rebuilding without starting over

For six weeks, she tested the assistant at home rather than trusting it immediately before a shopping trip. She used one ordinary meal plan more than once and kept the outputs in the grocery notebook, where differences were easier to see.

The results varied. One version contained 23 items and fit on a page. Another expanded the same plan to 31 by listing optional toppings and pantry staples she already had. A later response stayed short but sorted items alphabetically, which was tidy and useless for her route.

She found that a saved instruction helped when it described the output’s boundaries rather than her general preference. She specified that the assistant should return only the shopping list, retain quantities already approved during the conversation and use the store sequence recorded in her account settings. She also began asking it to revise the existing list instead of generating a replacement.

Those changes made the responses more consistent. They did not make them fixed.

Her sister had helped with groceries during the testing period, mostly by joining her on more crowded trips. Before the assistant, the two of them had exchanged long messages about meals and substitutions, a practical conversation that could become tiring for both. The AI had taken over part of that work without requiring another person to be available, and her sister understood why restoring it mattered.

The rebuilt routine also changed what she told the system. Earlier chats included detailed explanations of sensory preferences and days when eating had been difficult. That context sometimes improved suggestions, but it was personal information stored in an account she did not control.

She replaced some of it with narrower statements about food itself. The assistant needed to know that certain textures should not be suggested. It did not need an account of the day that taught her this. She removed old conversations from the visible history, while recognizing that deleting a chat from view did not answer every question about retention or past system use.

This introduced a trade-off she could feel in the answers. Less personal context meant the assistant sometimes needed another correction. It also made the grocery task feel more contained, closer to planning food than recording her life.

The list she uses now

Three months after the update, the grocery notebook held a new 23-item list. The assistant had built it from four planned dinners, two leftover lunches and a spending limit of $85. It grouped the food according to her route and left out commentary.

She checked every line against the meal plan. One quantity needed changing. Nothing had been duplicated.

The trip went as expected, though she no longer treated an expected trip as proof that the system had settled permanently. Before leaving home, she compared the generated list with the previous page and copied only the version she had reviewed. The notebook added work, but it also placed the final structure somewhere an update could not silently rewrite.

She kept using the assistant. Its ability to gather fragments from a conversation remained useful, and she did not want to return the task to her sister or force every meal into a rigid planning template. Dependence was not the whole description. Neither was independence.

The receipt from the disrupted trip eventually moved out of the notebook. The page stayed.

Questions people ask

Why did the

AI give different grocery lists for the same request?

The assistant generated each response from several possible word and formatting choices rather than retrieving a fixed saved document. After the underlying model or its instructions changed, it interpreted familiar preferences differently, adding optional items or reorganizing categories even when the shopper’s request appeared unchanged.

Can saved

AI preferences prevent an update from changing a routine?

Saved preferences improved consistency in her case, especially when they described a narrow output format. They did not lock the behavior. Company-level model changes and hidden system instructions could still affect how the assistant interpreted those preferences, and she had no setting that preserved an earlier version of the tool.

How did she limit the personal information in her grocery chats?

She replaced broad accounts of difficult days with narrower food preferences and removed older conversations from the visible history. That choice sometimes required more corrections, and it did not resolve every question about data retention, but the current grocery exchange contained less of her personal life.

Does she still trust the AI-generated grocery list?

She relies on it as a draft rather than a final record. Before each trip, she checks quantities and order, then copies the approved version into the grocery notebook. The most recent page contains 23 items, with one quantity corrected in pen.