Information date: 12 November 2026 — A foreign-owned auto-parts plant replaced its final-inspection bench with AI visual inspection. The model is discriminative: it classifies defects and returns a pass or fail label. The interpretive question is whether that counts as a public-facing generative service. Under the Interim Measures for Generative AI Services the duties attach to providers offering services to the public in China, and an internal quality-control tool normally falls outside that, as it does outside the deep-synthesis rules. But the line cameras capture faces and badge numbers, which is sensitive personal information: separate consent and a personal information protection impact assessment are required, and any offshore training use triggers cross-border rules. If the model trains on onshore data locally and only classifies defects, skip generative-AI filing but complete the impact assessment, separate consent and a retention limit before go-live. If any future release adds a natural-language interface used by staff or customers, re-screen it as a possible generative service. If any training data leaves the country, complete standard-contract filing or a security assessment before export, not after.
Verified facts and scope
A foreign-owned auto-parts plant replaced its final-inspection bench with AI visual inspection. The model is discriminative: it classifies defects and returns a pass or fail label. The interpretive question is whether that counts as a public-facing generative service. Under the Interim Measures for Generative AI Services the duties attach to providers offering services to the public in China, and an internal quality-control tool normally falls outside that, as it does outside the deep-synthesis rules. But the line cameras capture faces and badge numbers, which is sensitive personal information: separate consent and a personal information protection impact assessment are required, and any offshore training use triggers cross-border rules.
Scope has to be fixed before the numbers mean anything: Confirm before deployment: whether the model generates content at all; whether suppliers, auditors or customers can query it; whether cameras resolve faces clearly enough to identify an individual; where training images were captured and labelled; and whether the underlying model is hosted onshore or abroad. Anything outside that set is a separate question with its own evidence requirement, and mixing the two is the most common way a workable plan turns into an unusable one.
For a decision on “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore”, keep three layers apart: the published rule or statistic, the company’s own orders and process records, and the assumptions nobody has tested yet. The first layer defines the external conditions, the second proves what is true for this business, and the third is where the risk actually sits. Public material cannot substitute for a contract, a system log, a sample or a live test.
A practical way to test the scope is to ask which parts of “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore” would still be true if the company changed supplier, entity or product line tomorrow. Items that survive that change belong to the rule layer and can be taken from official publications; items that do not survive belong to the operating layer and have to be rebuilt from invoices, warehouse records, customs entries or platform statements. Where Confirm before deployment: whether the model generates content at all; whether suppliers, auditors or customers can query it; whether cameras resolve faces clearly enough to identify an individual; where training images were captured and labelled; and whether the underlying model is hosted onshore or abroad, the two layers are frequently mistaken for each other, and an action plan built on the wrong one either over-complies or misses a deadline it never saw coming.
How the effect reaches operations
The regulatory logic divides by service population, not by technology. A public, content-generating service draws security assessment and filing obligations; an internal classifier does not. Personal information law, however, does not care about model type — any processing of biometric data needs separate consent and an impact assessment. That is why the compliance work sits on the data pipeline, the camera placement and the labelling vendor, rather than on the model architecture.
The most common misjudgement is treating ‘internal tool’ as ‘no regulation’. Second is relying on the general employee onboarding consent, which is not separate consent for facial data. Third is the cross-border training loop: images shipped to an offshore engineering team, the model trained there, then deployed back into China — an outbound transfer nobody logged. Fourth is outsourcing the integration while retaining ownership of the dataset, so duties stay with the plant.
Translated into an operating sequence, The regulatory logic divides by service population, not by technology. A public, content-generating service draws security assessment and filing obligations; an internal classifier does not. Personal information law, however, does not care about model type — any processing of biometric data needs separate consent and an impact assessment. That is why the compliance work sits on the data pipeline, the camera placement and the labelling vendor, rather than on the model architecture. That ordering matters: check applicability first, then the transmission step, and only then the size of the gap in cost or timing. When an intermediate step cannot be confirmed, the honest response is to narrow the scope of the conclusion rather than to substitute an industry average for the company’s own figures.
The cost of getting the order wrong is asymmetric. A team that checks applicability too late usually discovers it at the point where the shipment, the filing or the invoice can no longer be changed without penalty, and by then the only remaining lever is price. A team that checks early spends a few hours on questions that turn out to be irrelevant, which is cheap by comparison. On “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore” the practical rule is therefore to spend the first hour on scope and the second on the numbers, never the other way round.
Decision
If the model trains on onshore data locally and only classifies defects, skip generative-AI filing but complete the impact assessment, separate consent and a retention limit before go-live. If any future release adds a natural-language interface used by staff or customers, re-screen it as a possible generative service. If any training data leaves the country, complete standard-contract filing or a security assessment before export, not after.
The conditions should be written down so the decision can be re-run by someone else: Confirm before deployment: whether the model generates content at all; whether suppliers, auditors or customers can query it; whether cameras resolve faces clearly enough to identify an individual; where training images were captured and labelled; and whether the underlying model is hosted onshore or abroad. When those hold, proceed as described; when one is missing, close the gap first and decide afterwards. This is an analytical case note, not AI, privacy or cybersecurity legal advice. Filing thresholds depend on facts and local CAC practice; obtain a written assessment before deployment.
Two workstreams usually have to move in parallel. The compliance workstream closes the document side of “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore” — registrations, confirmations, filings — while the commercial workstream protects the delivery date, the pricing already quoted and the customer commitments that were made on the earlier reading of the rules. Running them as one queue is how small documentation gaps become missed shipment dates: the commercial side waits for a certificate that could have been requested a week earlier. If the model trains on onshore data locally and only classifies defects, skip generative-AI filing but complete the impact assessment, separate consent and a retention limit before go-live. If any future release adds a natural-language interface used by staff or customers, re-screen it as a possible generative service. If any training data leaves the country, complete standard-contract filing or a security assessment before export, not after. Set one owner per workstream and one shared status sheet, and treat a red item on either side as a reason to freeze only the affected order.
Implementation checklist
- Classify the model in writing before the cameras go live.
- Run the employee-image impact assessment as a pre-condition, not a follow-up.
- Keep the training dataset and labelling work onshore by default.
- Name one decision owner, one execution owner and one review date.
- Keep the source page, access date, entity and internal evidence.
- On any change, reopen only the affected judgement.
Evidence and review
Start with the smallest test that can change the answer: Classify the model in writing before the cameras go live. Pick one real entity or one representative case, write down the continue, watch and stop conditions in advance, and record the input version, the owner, the start time and the actual result. The first pass should only look at facts that move the decision; formatting preferences are not risk.
The second pass is completed by “Run the employee-image impact assessment as a pre-condition, not a follow-up”, and someone who did not prepare the file should spot-check one piece of critical evidence. The most common misjudgement is treating ‘internal tool’ as ‘no regulation’. Second is relying on the general employee onboarding consent, which is not separate consent for facial data. Third is the cross-border training loop: images shipped to an offshore engineering team, the model trained there, then deployed back into China — an outbound transfer nobody logged. Fourth is outsourcing the integration while retaining ownership of the dataset, so duties stay with the plant. A single small deviation can be corrected in place; a repeated deviation, or one that touches a statutory duty or changes the economics, is what justifies pausing the affected part.
After “Keep the training dataset and labelling work onshore by default”, compare the actual result with the baseline rather than with the impression of the team. If the model trains on onshore data locally and only classifies defects, skip generative-AI filing but complete the impact assessment, separate consent and a retention limit before go-live. If any future release adds a natural-language interface used by staff or customers, re-screen it as a possible generative service. If any training data leaves the country, complete standard-contract filing or a security assessment before export, not after. When the arrangement is extended to more products, customers or processes, keep the same success test and one complete record per batch. If later information changes, update only the affected fields.
One last discipline keeps the file usable next quarter: write the closing note as if a colleague who has never spoken to the supplier has to act on it. That note states what was verified, on which date, against which source, and what remains open — on “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore” it also states which of the three actions was actually finished and which was only scheduled. Teams that skip the closing note re-litigate the same question every month, and each round costs more than the documentation would have. This is an analytical case note, not AI, privacy or cybersecurity legal advice. Filing thresholds depend on facts and local CAC practice; obtain a written assessment before deployment.
Limits of the conclusion
This is an analytical case note, not AI, privacy or cybersecurity legal advice. Filing thresholds depend on facts and local CAC practice; obtain a written assessment before deployment.
The limits are structural rather than cautious wording. Public sources describe rules and market conditions; they do not observe “Case: A Foreign-Owned Auto-Parts Plant Switches to AI Visual Inspection — Whether the Model Counts as a Public-Facing Generative Service, the Employee-Image PIPIA and the Training Data It Had to Keep Onshore” inside a specific company. Where the analysis quotes a second-hand interpretation, the original release has to remain reachable, and a figure that cannot be traced back does not enter the decision table. Where the point depends on Confirm before deployment: whether the model generates content at all; whether suppliers, auditors or customers can query it; whether cameras resolve faces clearly enough to identify an individual; where training images were captured and labelled; and whether the underlying model is hosted onshore or abroad, the conclusion is only as current as the last check of those conditions.
