Where Is the AI Safety Scorecard?

Every time a new AI model is launched, it often comes with an impressive set of numbers (benchmarks!). Often accompanied by charts showing how the new model compares with the competition across various categories.

But how often do we see equally detailed safety evaluation results?

This was the question that a recent study published by SemiAnalysis attempted to answer.

The researchers examined 857 AI model releases from nine leading Chinese AI developers between 2021 and September 2026.

Some of the findings:

• Only 31 releases (3.6%) had publicly available, model-specific safety evaluation results.
• Just 9 releases had those results available at or before launch.
• For example, Alibaba published safety results for 7 of its 238 releases, Tencent for 1 of 133, and ByteDance for 2 of 120.

One thing to note: the absence of published safety results doesn’t mean the models weren’t tested. The study looked at public disclosure, not internal testing.

And while this research focused on Chinese developers, I think the question it raises is relevant to the wider AI industry.

Companies such as OpenAI, Anthropic and Google do publish safety reports for some of their major models. But the level of detail and consistency of reporting varies.

We probably expect (maybe demand is a better word) a safety scorecard too?

Reference: SemiAnalysis — Beijing Will Not Pace the Frontier