# China AI Compliance Checklist and Algorithm Filing Tracker The compliance set for offering AI services in China: an applicability matrix across the CAC instruments, a security assessment checklist, a content labeling specification, and an algorithm filing tracker. **Who this is for:** The compliance owner for AI products available to users in mainland China. Source playbook: https://aigovernance.com/playbook/china-ai-regulation-compliance --- ## China AI regulatory applicability matrix _Which CAC instrument applies to which product or feature._ ### Template | Product / feature | Generative AI Measures | Deep Synthesis Provisions | Algorithm Recommendation Provisions | Data / PIPL | Security assessment required? | |---|---|---|---|---|---| | | Y / N | Y / N | Y / N | Y / N | Y / N | ### Worked example | Product / feature | Generative AI Measures | Deep Synthesis | Algorithm Recommendation | Data / PIPL | Security assessment? | |---|---|---|---|---|---| | Text assistant (public) | Y | N | N | Y | Y (public-facing generative service) | | Image generation | Y | Y | N | Y | Y | | Product recommendation feed | N | N | Y | Y | filing, not full assessment | | Internal analytics (no public output) | N | N | N | Y | N | ### Acceptance criteria - Every China-available feature is assessed against each instrument. - The trigger for a security assessment versus a filing is recorded per feature. - PIPL and data-export obligations are tracked alongside the AI-specific instruments. --- ## CAC security assessment checklist _The areas a self-assessment (and any filed assessment) must cover for a public generative service._ ### Template | Area | Requirement | Status | Evidence | |---|---|---|---| | Training data | lawful sources; IP respected; no unlawful content; personal data handled per PIPL | | | | Data annotation | annotation rules; staff training; quality checks | | | | Content safety | filtering for prohibited content categories; refusal behaviour | | | | Output testing | pre-release testing against a content-safety test set; documented pass rate | | | | Model transparency | disclosure of service provider; complaint channel | | | | Real-name and minors | user identity verification; minors protection measures | | | | Incident handling | takedown, model tuning, and reporting process for unlawful content | | | ### Worked example | Area | Status | Evidence | |---|---|---| | Training data | Complete | provenance records; content filter on ingestion | | Data annotation | Complete | annotation SOP; annotator training log | | Content safety | Complete | multi-category filter; refusal tests | | Output testing | Complete | 2,000-prompt safety set; pass rate documented | | Model transparency | Complete | provider disclosure + complaint form in-product | | Real-name / minors | In progress | identity check via partner; minors mode pending | | Incident handling | Complete | takedown + retrain + report runbook | ### Acceptance criteria - Every area has a status and attached evidence. - Output testing has a documented test set and a pass rate, not a claim. - A content-incident handling process exists and has been exercised. --- ## Content labeling implementation specification _What must be labeled as AI-generated, how, and in which formats, per the labeling rules._ ### Template | Content type | Explicit label (visible/audible) | Implicit label (metadata) | Placement / format | Implemented? | |---|---|---|---|---| | Generated text | | | | | | Generated images | | | | | | Generated audio | | | | | | Generated video | | | | | ### Worked example | Content type | Explicit label | Implicit label | Placement / format | Implemented? | |---|---|---|---|---| | Generated text | "AI-generated" notice near the output | provider + generated flag in response metadata | prepended line; not removable in the UI | Yes | | Generated images | corner watermark + caption | C2PA-style metadata; provider ID | bottom-left, min 5% width | Yes | | Generated audio | spoken disclosure at start | metadata tag | first 2 seconds | Partial | | Generated video | on-screen label first + persistent corner mark | metadata tag | first 3 seconds + corner throughout | Partial | ### Acceptance criteria - Both explicit (visible/audible) and implicit (metadata) labels are specified per content type. - Label placement and format meet the size and duration expectations in the rules. - Partial items have an owner and a date. --- ## Algorithm filing tracker _Filing status, registration numbers, and renewal dates for each algorithm subject to filing._ ### Template | Algorithm / service | Filing type | Submitted | Registration number | Approved | Renewal / update due | Owner | |---|---|---|---|---|---|---| | | initial / change | YYYY-MM-DD | | YYYY-MM-DD | YYYY-MM-DD | | ### Worked example | Algorithm / service | Filing type | Submitted | Registration number | Approved | Renewal / update due | Owner | |---|---|---|---|---|---|---| | Text assistant (generative) | initial | 2026-05-10 | (redacted) | 2026-07-02 | on material change | China Compliance | | Recommendation feed | initial | 2026-04-01 | (redacted) | 2026-05-20 | annual review 2027-05 | China Compliance | ### Acceptance criteria - Every algorithm subject to filing has an entry with its current status. - Material changes to a filed algorithm trigger a change filing, tracked here. - Renewal and review dates are on the China compliance calendar. --- ## Governance controls this kit produces evidence for - **CMP-001**: The applicability matrix is multi-jurisdiction mapping for the China instruments. - **CMP-006**: The labeling specification is the AI content watermarking and labeling compliance record. - **DGC-001**: The training-data area of the CAC checklist maps to training data provenance. - **SAF-004**: Output safety testing against a content-safety set is AI reliability testing evidence. - **CMP-002**: The compliance calendar for CAC guidance updates is part of standards and regulatory monitoring.