Model Cards 2.0: industry convergence on standardized AI safety reports
In one sentence Google, Anthropic, and Meta converge on structured second-generation model cards that include training data, safety evaluation results, red-team findings, limitations, and intended use. A first step toward auditable AI.
When you buy a food product, you find an ingredients label and nutritional values. Not perfect, but at least you know what you are consuming. For years, AI models had nothing similar.
Model cards — documents describing how a model was trained, what it can do, and what it cannot do — have existed since 2018 (proposed by Google). But the first generation was often superficial: a few lines of marketing with some benchmark numbers.
In 2025, a much more detailed new generation of model cards emerged. Google, Anthropic, and Meta began publishing structured documents that include: the datasets used for training and their limitations, the results of pre-release safety evaluations, findings from internal red-teaming teams, known limitations of the model, and the use cases the model was designed for and those it was not.
There is still no mandatory standardized format — each company uses its own structure — but this is the first concrete signal of convergence toward verifiable transparency. For those who must decide whether to use a model in a regulated context, having this data documented makes a difference.
Companies
Google, Anthropic, Meta
Tools
—
Tags
Sources