Three Generations, One Feed

Within four weeks of each other, three things appeared in Chinese feeds. In Hefei on August 17, a father ordering a ¥6 breakfast for his four-year-old — in hospital for leukemia chemotherapy — typed the child’s dietary restrictions into the delivery-note field; the shopkeeper called to confirm, cooked to the split specification, and slipped ¥100 and two cups of soy milk into the box. Across two weeks of August, on Douyin and Xiaohongshu, adults in their thirties uploaded selfies and let third-party AI tools render the children they used to be, because the real photographs were lost in moves, fused damp in rural albums, or never taken; some set the renderings as their profile pictures. And on a Monday at 7:48 a.m., a user — in a first-person post, not a news-verified account — reported a push notification that reached him while he was brushing his teeth: his ride was already en route, derived from three months of his 8:05 departures and, among other recurrences, his biweekly Saturday trips to his parents’ home.

Our Chinese-language column ran these as separate dispatches. Read together they are not three stories but one machine, and this piece takes it apart: what a Chinese family now has to write down about itself, where those writings live, and who holds the copies.

The gears

Sociologists have a name for the work of maintaining a family’s information — who remembers what the child can eat, where the albums are kept, which Saturday the visit falls on: kin-keeping. The three instances above are that work migrated, field by field, from memory and the shoebox into commercial systems, under three different custody arrangements.

In the first, family information is stated by a family member into a text box, read once by a human stranger, and answered in cash. In the second, it is synthesized: the missing record is generated from the adult’s present face by a model operated by parties the family cannot name, paid for with a biometric upload. In the third, it is inferred involuntarily — nobody states anything; the routine states itself, three months at a time, in an order log.

Run the three by birth year and a gradient appears. The four-year-old exists in the record before being able to read it — diet in an order field, condition in a note, before any consent is possible. The thirty-year-olds belong to the cohort whose childhoods predate cheap cameras and whose documentation has the holes to prove it; they are generating the missing records now. The oldest generation does not appear as a person at all: the ride log reads “biweekly Saturday, parents’ home” — grandparents present in the data as a destination. One family, three generations, one feed: each legible to a platform in a different form, none of them holding the copy.

Instance one: the chemo diet in a note field

The working side, in the sequence the reporting supports. August 17, around 6 a.m., Hefei. The order was a ¥6 soft egg pancake; the note read, in substance: child on chemotherapy, no seaweed or lettuce, half non-spicy, half mildly spicy. The shopkeeper telephoned the father back to confirm what the child could not eat, made the pancake to the split spec, put ¥100 and two cups of soy milk into the box unannounced, and replied to the father’s thanks with one line — “wish the child an early recovery; keep going.” The China News Service wire and SCMP carried the story abroad within days. SCMP’s arithmetic is the part worth keeping: ¥100 is roughly four days of profit for a breakfast stall of that size, about ¥25 a day on SCMP’s basis [unverified — profit figure as relayed by SCMP].

The failing side arrived hours later, in the comment section of the number-one trending topic. If “child on chemo” in a note field moves one merchant to give, is the next merchant who gives nothing thereby cold-blooded? And if the format works, what stops strangers from fabricating a note-field story for the same attention? Our column’s rebuttal was precise: the note was a dietary specification any father would send to any kitchen; the cash was the shopkeeper’s unprompted decision; reading pity-selling into the disclosure amounts to requiring every family with a sick child to establish its innocence before ordering breakfast. Both halves belong in the record: the money arrived on day one, and by the same evening the discourse had built a checkpoint — is this a script? — that the next genuine note, and the next shopkeeper, will have to pass.

Instance two: the childhood that had to be generated

The working side. Over roughly two weeks of August [unverified — trend window and platform heat as relayed by our column], Douyin and Xiaohongshu filled with adults uploading front-facing photos and receiving, seconds later, a child: same eyes, a missing tooth, a 1990s sweater. Styles to choose — retro ID photo, old film, hand-drawn cartoon; pairs of the adult and the six-year-old standing side by side; “family portraits” of three people who never sat for one together. The demand underneath is specific and, in the comment sections, explicit: family photos lost in a house move, an album that fused solid with damp, a rural childhood in which photographs were rare events. For users with no original, the artifact is the only version that exists — which is why, as our column put it, people who know exactly what it is still cry at it.

The failing side has two layers. The first is epistemic: the rendering is not the child who existed. It is inferred from the adult’s current features and smoothed toward an aesthetic average — bigger eyes, whiter skin, a tidier smile. In the column’s formulation: your six-year-old self may really have smiled with a missing tooth — just a different one. The second layer is the entry price: generating the image typically requires handing a front-view face to a third-party mini program, with no way to know where it is stored or what else it trains. This is a documented, recurring pattern rather than a hypothesis — the same warnings accompanied the Miaoya camera app’s viral run (Sixth Tone) — and the statute already treats the data as the most protected category: under the Personal Information Protection Law, in force since November 2021, biometric identification information is sensitive personal information, processable only for a specific, sufficient purpose with separate, un-bundled consent (China Briefing). The rule is on the books; the upload flow runs on the screen.

Instance three: the family routine, read from the order log

What platforms actually ship is documented and current. In an April 2026 hands-on, state radio CNR tested the new AI-hailing layer: Alibaba’s Qwen app takes one spoken or typed sentence and books a simple ride in around ten seconds; Didi’s “AI Xiaodi” accepts fuzzy constraints — fresh air, cheap, no motion sickness, nearest car — against more than 90 service tags the company says it maintains, proposes up to three plans, and requires manual confirmation, auto-cancelling on timeout. The adoption curve is steep: Didi’s figure, relayed by Chinese tech press on April 3, has AI-Xiaodi users up 37-fold since the start of the year (our column phrased the same multiple as one week’s surge — the relays differ on the basis; both are relayed), with users born after 2000 over 40% [unverified — CCN relay]. The working side is real: ten seconds against the unlock-locate-scroll-select routine, and the confirm button stays under the user’s thumb.

The failing side is the 7:48 a.m. post — and it keeps its label: a first-person account, not a news-verified event. While brushing his teeth, the user received a push: “You usually depart for the office at 8:05; a vehicle has been booked, arriving at your compound gate in about ten minutes.” His reconstruction of what the system knew: weekday 8:05 departures for three months; a Friday izakaya; biweekly Saturday trips to his parents; rain days five minutes earlier; Wednesdays 8 p.m. to a friend’s home; four visits to one hospital within a month. “I never told anyone I was seeing a doctor,” he wrote. “My trip records sold me out.” The legal expert CNR quoted — deputy director of an industry-and-information-technology-law key laboratory — put both halves in one sentence: the feature is a “benign interaction” and it carries “model hallucination” risk; platforms owe users anonymization and de-identification. The off-switch exists; using it returns every small task to manual. The user’s closing line — the more it knows you, the harder it is to leave — is a position, not a finding. What is verifiable is the substrate: the trips are logged, the logs persist, and the routine they encode — workdays, hospital, the parents’ Saturdays — is precisely the kin-keeping that used to live in someone’s head. We first covered this dispatch in watch Issue 3.

The ladder underneath

Set the three instances side by side and the gradient runs by birth year. A child born around 2022 enters the record before being able to consent to any of it. The cohort born around 1994, across much of rural China, has the opposite problem — the era’s photographs were scarce and fragile, and the survivors of that scarcity are now commissioning synthetic replacements from their own present faces. The pre-1970s generation, in these stories, is recorded mainly in other people’s data: an address in a ride log, a Saturday recurrence. The family information once held in memory and shoeboxes was lossy but family-held; the same information in platform fields is durable and third-party-held. The substrate is large — Didi’s China mobility averaged 39.4 million daily orders in the first quarter of 2026 [unverified — company figure relayed in Chinese press] — and adoption is youngest-first, post-00s over 40% of AI-hailing users: the most-documented generation is also the earliest adopting the systems that do the documenting.

What outsiders usually get wrong

Four corrections, all factual. First, the Hefei story was not an exposé of delivery platforms, and the argument it ignited was not about them: it trended as a private act of giving, and the national debate was over whether a father may write his child’s illness in a note field at all — with the note’s actual text a dietary specification. Second, the “AI childhood photo” is not photo restoration improving: no photograph is being recovered; the output is inferred from present features plus an aesthetic average, and the tools’ own users describe it that way while using it anyway. Third, “China has no privacy law for this” fails on the statute book — the PIPL has classed biometric data as sensitive personal information requiring separate consent since November 2021; the instances above document the upload flow, not the absence of the rule. Fourth, the 7:48 pre-emptive push was a first-person post, not a confirmed platform behavior; what is confirmed shipping is the opt-in AI-hailing layer with manual confirmation — the two are different things.

Sources

• ECNS / China News Service wire: breakfast shop owner’s gesture to a child with cancer goes viral (2026-08-21)
• SCMP: Chinese dad’s chemotherapy note prompts food vendor’s ¥100 (US$15) donation
• Sixth Tone: an AI photo app wows China, but privacy fears loom (the Miaoya precedent)
• China Briefing: how to legally handle sensitive personal information in China (PIPL)
• CNR (reporter Jiang Xiaochen): hands-on with AI ride features (April 2026)
• 10jqka/Tonghuashun: Didi’s one-sentence hailing era (2026-03-22)
• Baijiahao: the 7:48 first-person account (personal narrative, not news-verified)

Provenance & disclosure. This piece synthesizes three Chinese-language originals from our WeChat channel — “谁在消费苦难？外卖备注“孩子化疗”，店主塞进100元现金” (2026-08-19), “哭了。三十岁的人，在AI里看见六岁的自己” (2026-08-14) and “你还没下单，车已在楼下等你？” (2026-08-28, also covered in watch Issue 3) — drafted with AI assistance under human editorial direction and adapted to English 2026-09-06. Verification: the Hefei order, the note’s dietary content, the ¥100 + two soy milks and the one-line reply against ECNS and SCMP, including SCMP’s four-days-of-profit basis; the face-data risk pattern against Sixth Tone’s Miaoya coverage; the PIPL classification (biometrics as sensitive personal information, separate consent) against China Briefing; the platform layer (Qwen ~10 s, AI Xiaodi 90+ tags, three-plan manual confirm) against CNR’s April 2026 hands-on and 10jqka; the 7:48 account kept labeled as a personal narrative per its original post. Figures relayed only through our column or single-outlet press — the two-week trend window and platform heat, seconds-to-generate timing, style menus, the comment-section photo-loss anecdotes, the ¥6 order price, the 40% post-00s share (CCN), the ¥25/day stall-profit derivation (SCMP basis) and Didi’s 39.4-million Q1 daily orders — are marked [unverified] above. Basis note: the 37× user-growth multiple is Didi’s own figure as relayed April 3 by Chinese tech press, measured since the start of the year; our column’s “one week” phrasing is retained as an alternative relay, not a recomputation. Ratios recomputed: 100/6 ≈ 16.7× the order value; ¥100 ≈ 4 days of stall profit → ≈¥25/day; 4 hospital visits over 31 days ≈ one per 7.8 days. This is reported synthesis — not a SigPulse measurement, not medical or legal advice. Our first-party measurements live in the dispatches and the /data/ ledger.