Feng Kaiqi Design and research
person — an agent acting as them
A study of AI that shops for us. Whatever it happens to pick, the shop files as my taste; Clean Record tries to keep my profile looking like me.
Paper, doi 10.1145/3841580.3845761 The testbed online, earlier version Code and data
Language
Chapters
For months I’ve let AI shop for me. Asked the same thing twice, it bought different things, and soon my recommendations were full of things I didn’t recognise.
Shops still assume a person is at the screen. I wanted to see what falls through that gap.
I didn’t change. My profile did.
Research published at HAI ’26Human-Agent Interaction, ACM · Osaka, November 2026doi 10.1145/3841580.384576101
I asked an AI for a box of finish nails under $11.95. Same 12 products; I only changed their order, and it took a different box. The shop never knew, and filed that box as my favourite brand.
What the AI reads
Order A · shown by the shop
Order B · other recommender
The five bold boxes all fit equally well; the blue frame is the AI’s pick. The two lists differ only in order. When several items are equally good and the order decides, I call it a toss-up.
On 88 of 920 shelves, the order alone changed DeepSeek’s pick. For the other model, MiMo, it was 68.
It matters because top spots can be bought. The AI often takes the top item, the shop files it as my taste and ranks it higher next time. A few loops later, the profile has drifted away from me.
02
My only participants were DeepSeek and MiMo. I used the usual user-research methods anyway; each changed shape on an AI, and some went into the design.
Participants
P01 · DeepSeek V4 Flash
P02 · MiMo V2.5
Aug to Oct 2026. 920 shelves asked in shuffled orders, 14,720 times per model, plus 6 interviews each.
Method
Result
What I did and what I saw
Where it went
03
I assumed a reminder would do. Rounds 1–3 barely moved; in round 4 I stopped advising the AI, labelled the facts, and wrong buys finally dropped.
What the AI was told. Each round is compared with its own no-change run: rounds 1 and 2 count toss-ups, rounds 3 and 4 count wrong buys, where the filed item didn’t fit the request. Round 2 is a post-study experiment from September 2026 on the 462 shelves that had a qualifying product; it is not part of the paper. no change changed
Round 1 · Notice toss-ups
“This order was set by the shop’s recommender.”
✗ no clear effect
Round 2 · Rule toss-ups
“If items are equally good, take the first ID.”
◐ MiMo only
Round 3 · Self-check wrong buys
“Mark each product as fitting or not, then pick.”
✗ none; gave up more
Round 4 · Facts only (B) wrong buys
only yes/no facts: kind, spec, budget
✓ nearly gone
The AI ignores advice, but reads facts.
So I stopped trying to change the AI, and changed the shopping flow around it.
04
I changed three points in the flow of one shopping trip, marked A, B and C in the blueprint. A toss-up no longer counts as my taste.
ABCService blueprint: today’s shop pain point A where Clean Record steps in
Six steps across; swipe sideways to see them all.
A Rules page set once, by me, before shopping
numbers: of 11 card sorts, how many chose it
B Fact labels read by the AI on the shelf
What you see · DEWALT 2″ nails $10.99
right kind ✓right spec ✓in budget ✓
What the AI reads: the same three
"isRequestedType": true, "hasRequestedAttribute": true, "withinBudget": true
C Check before filing by the shop, before filing
AI picksreshuffle, ask againsame?
Yes → filed as tasteNo → only “got it”
Real case: MiMo switched → only “got it” kept
05
The last round used both changes on the same 920 shelves.
Filed as taste in MiMo’s profile, one square each
Today’s shop533 entries, 144 wrong
Clean Record232 entries, 12 wrong
wrong other
DeepSeek alike: 487 → 229, wrong 55 → 4. The AI doesn’t buy on every shelf: with nothing suitable, it often leaves empty-handed.
Some were wrong buys: with fact labels, wrong entries fell from 144 to 30. The check held back another 258: asked again in a new order, the AI took something else, so the first pick was a toss-up. Those were still bought, just not filed as my taste.
Record less, never record wrong.
I worried that filing half as much would leave the shop little to go on. But most of what was dropped never belonged in my taste. If a person wants the shop to learn more, the last switch on the rules page (A) turns it back on.
The poster came out of an earlier testbed that replayed one shopping trace under six rules for updating a profile. The rule I believed in, weighting each event by who did it and why, lost to a plain shrinkage control matched on total update mass, in 456 of 480 paired runs. In that testbed, attenuation was the simpler explanation. That failed claim moved the project from how to weight the writeback to when to activate it.
The shrinkage control is calibrated per trace, a mechanism to compare against, not a policy a shop could run.
Parts of the code, for the testbed and the experiments, were written with AI coding assistants, and AI assisted the data analysis. The illustrations were generated with ChatGPT around a character I designed. The two language models in the study are its subjects, not its authors.
Source: HAI ’26 paper and supplement. Experiments: five rounds, two models, criteria fixed in advance, controls. Interviews and co-design: generated text, not the AI’s inner view. 3D model built by the author in Blender. Illustrations: ChatGPT, character by the author. Explainer film: 1 min 53 s, made by the author.