Why Can't a ¥3,299 AI Homework Machine Read a Child's Handwritten 7? The Take Inside China

A machine that grades homework got a simple thing wrong in front of a family: it read a child’s handwritten “7” as a 9, and on another evening read a multiple-choice option meaning “cannot be determined” aloud with a garbled final syllable. The dispute that followed — over a ¥3,299 Xiaoyuan S2 “study tablet” bought in Chengdu — became our WeChat column’s vehicle for taking AI grading apart “one gear at a time.” The piece ran 2026-08-28, building on Sichuan consumer reporting from July 2025 and August 2026 trending parent complaints. This entry translates it.

The numbers

• Device: Xiaoyuan S2 study tablet, ¥3,299, purchased May 9 (2025); first misgrade May 12; voice error June 9; further errors to June 24 (Consumer Quality Daily timeline)
• Market: 1.265M learning tablets sold Q1 2025, +29.4% YoY (RUNTO); 1.77M+ social notes on the category; flagship price point ¥5,999 pushed hardest in stores
• All case details are domestic-sourced [unverified — the primary consumer report is outlet-root-linked only; no deep article URL located]

The take inside China

The teardown. “AI grading” is a pipeline with three stations, and each can fail differently. Recognition: a child’s handwriting isn’t typeface — a hurried 7 with a curved tail reads as 9 (a user test on SMZDM documented the same). Question bank: recognition isn’t explanation — if the item isn’t in the library, the machine grades but won’t teach; the store itself admitted it couldn’t promise per-question coverage. Generation: the LLM layer is the confident classmate — fast, fluent, and wrong with full assurance; August 2026 trending complaints center exactly there (children copying logically flawed answers). The take’s verdict on the three: recognition failure is immaturity; bank gaps are undelivered service; confident wrong answers are the most hidden failure — a child’s first instinct when misgraded is that they erred, not the machine.

The responsibility vacuum. The store’s defense: “no brand achieves 100% grading accuracy” and returns require third-party lab reports. The lawyer’s counter (Wang Bo): defects defeating core function may violate product-quality law; activation-no-return may violate consumer protection. The take’s framing of the industry’s posture: sold as tutor (“AI辅导, frees the parents”), defended as tool (“for reference only, bears no teaching responsibility”) — charging tutor prices while accepting tool liability. Its fairness note: the 100%-accuracy disclaimer is honest as far as it goes; the problem is that “will err” never appears at the same volume in the sales pitch.

The anxiety economics. Parents are not buying a tablet; they are buying “not supervising homework after 8 p.m.” The counter-trend the take savors: parents using free general-purpose assistants (Doubao-class chat apps) with zero expectations are calmer than buyers of professional AI hardware — because nobody promised them professionalism. Promise saturation determines collapse size. The closing question, left open: when a machine is both teacher and referee and blows the whistle wrong, who owns the whistle?

What the Chinese take left out

Effect sizes — no error-rate measurements exist in any of the cited reporting (the category’s central spec, absent). Also missing: MOE curriculum-compliance rules for such devices and what the sales contracts actually bind. Our own dispatches measure models on hardware; nothing about this category has been measured first-party — which is exactly why it stays in Watch and out of /data/.

Why it matters outside

The “sold as tutor, warranted as tool” gap is not China-specific — it is the coming consumer-protection argument for every AI homework/grading product anywhere, and the three-layer failure model (recognition / retrieval / generation) is a serviceable checklist for any parent evaluating one. The calm-users-of-free-tools observation is also a general one: expectation setting is a product’s hidden spec.

Sources

• Sichuan Online · Consumer Quality Daily (2025-07): the Chengdu case (outlet root; no stable article URL located)
• RUNTO 洛图科技: Q1 2025 learning-tablet shipments (outlet root)
• Liberation Daily: Can AI learning machines free the parents? (outlet root)

Provenance & disclosure. Originally published in Chinese on our WeChat channel on 2026-08-28 (“大几千的AI学练机，连孩子写的’7’都认不准”); drafted with AI assistance under human editorial direction. Translated to English on 2026-08-29 (AI-assisted, human-reviewed). The consumer-case timeline and RUNTO figures could only be checked against outlet roots, not deep article URLs — the case details carry [unverified] status at the figure level. This is translated commentary — not a SigPulse measurement. Our first-party measurements live in the dispatches and the /data/ ledger.