Google delivers twin models faster than its AI safety reports

More than two years after Google was caught with a flat foot since Openai’s exit ChatgptThe company dramatically picked up the pace.

At the end of March, Google launches AI’s reasoning model, Twins 2.5 Prothis leads the industry On several indicators, measuring encoding and mathematics capabilities. This launch came just three months after the technological giant debuted another model, Twins 2.0 flashIt was the most modern of the time.

Google Director and Gemini Product Director, Tulsee Doshi, told TechCrunch in an interview that the growing cadance of launching the company’s model is part of the agreed efforts to keep up with the fast -growing AI industry.

“We are still trying to understand what is the right way to expose these models – what is the right way to get feedback,” Dosha said.

But it seems that the timing frame for release has come at a price. Google Is yet to post safety reports on its most recent modelsIncluding Gemini 2.5 Pro and Gemini 2.0 Flash, raising fears that the company gives priority to transparency.

Today, it is quite standard for border AI laboratories – including Openai, Anthropic and Meta – to account for safety testing, performance evaluation and use of cases where a new model is launched. These reports, sometimes referred to as “system cards” or “model cards”, were offered years ago by researchers in the industry and academics. Google was actually one of the first to offer model cards in a 2019 Research Document, Calling them “approach for responsible, transparent and responsible practices in machine learning.”

Doshi told TechCrunch that the company did not publish a model card for the Gemini 2.5 Pro because it considers the model a “experimental” edition. The purpose of these experimental versions is to place an AI model in a limited way, to obtain feedback and to repeat the model before the start of production, she said.

Google intends to publish the model card of the Gemini 2.5 Pro when it makes the model as a whole, according to Doshi, adding that the company has already tested the safety and a racing red team.

In a subsequent message, a Google spokesman told TechCrunch that safety continues to be a “top priority” for the company and that it plans to release more documentation around its AI models, including the Gemini 2.0 Flash, moving forward. Gemini 2.0 Flash, which is usually available, also lacks a model card. The last card Google release was for Gemini 1.5 Prowhich came out more than a year ago.

System cards and model cards provide useful – and non -volatile, at times – information that companies do not always advertise for their AI. For example, the OPENAI system card launched for its O1 reasoning model revealed that the company of the company The model tends to “scheme” against people And he secretly pursues his own goals.

Overall, the AI ​​community perceives these reports as conscientious efforts to support independent assessments of research and safety, but reports have become more important in recent years. AS Transformer He noted earlier that Google told the US government in 2023 that he would publish safety reports for all “significant”, public AI models in the range. The company made a similar commitment yes other governmentspromising to “provide public transparency”.

The federal and state level has regulatory efforts to create safety reporting standards for AI model developers. However, they were greeted with limited adoption and success. One of the more remarkable attempts was Veto California Bill SB 1047which the technology industry strongly opposed. Legislators also set out legislation that would authorize the American AI Safety Institute, AI Standard Authority, Standard Determination, In order to establish guidelines for edition modelsS However, the Safety Institute is now upright Possibilities at the Trump administration.

Of all the appearances, Google is lagging behind in some of its promises to report the model testing, while delivering models faster than ever. It’s a bad precedent, very much Experts Argue – especially since these models become more capable and complex.