The Safety Paperwork Missed Launch Day

The model has left the station. The safety paperwork is apparently still looking for its platform.

In an October 8 report, SemiAnalysis reviewed 857 identifiable releases from nine major Chinese AI developers, covering 2021 through September 15, 2026. It found public, model-specific safety results available at or before launch for nine releases: 1.1% of the sample.^[1][2]

Read the small print before the big number

• 31 releases, or 3.6%, had qualifying public safety results in the census.
• Nine had those results by launch; 16 received them later, with a median delay of 42 days and a maximum of 349.
• The remaining six had results whose timing or exact-model match could not be established.^[1]

Viddy the distinction, droogs. A report arriving six weeks later may improve the record. It cannot travel backwards and help the buyer who had to decide on launch day.

The researchers counted actual results, including substantive findings, rather than generic assurances of evaluation. They did not automatically apply a flagship model’s results to its variants. Their search covered public materials; an absent disclosure does not establish that internal testing was absent, or that a model was unsafe. The census also does not supply a comparable US disclosure percentage.^[1][2]

Ask for the receipt

The sensible buyer’s question is wonderfully unglamorous: which results describe this exact model, and were they available when we committed to using it? A safety slogan has excellent uptime. It is the supporting evidence that needs a service-level agreement.