Google is shipping Gemini models faster than its AI safety reports

More than two years after being caught on flat foot by Google OpenAI ChatziptThe company has dramatically accelerated.

In late March, Google launched an AI argument model, Gemini 2.5 ProThat Leads the industry Coding and mathematical capabilities measure a number of criteria. This launch came just three months after the Tech Giant launch a model debut, Gemini 2.0 flashIt was sophisticated for time.

Google’s director and head of Jemini’s product, Tulsi Doshi told TechCranch in an interview that the growing cadens of the company’s model launches was part of a joint effort to retain the rapidly developed AI industry.

“We are still trying to determine what the right way to keep these models – what is the right way to get the response,” Doshi said.

However, the frame seems to have spent the rampd-up release time. Google For its latest models have not yet been able to publish security reportsGemi 2.5 Pro and Gemi 2.0, with flash, raises anxiety that the company is giving priority to the speed rather than transparency.

Today, for Frontier AI labs, including OpenAI, anthropologists and meta, it uses cases to be fairly standard, security tests, performance evaluation and whenever a new model launches. These reports, sometimes known as “System Card” or “Model Card”, proposed many years ago art and academia researchers. Google was actually the first one to suggest a model card in one 2019 research paper, They are called “an approach to being responsible, transparent and accountable in machine learning”.

Doshi told TechCrunch that the company did not publish a model card for Jemi 2.5 Pro because it considered the model as a “experimental” manifestation. He said that the goal of this experimental manifestation is to keep an AI model in limited ways, to get feedback and repeat the model before the production of a production, he said.

Google Gemi wants to release the model card of 2.5 Pro when it usually makes the model available, according to Doshi, adds that the company has already tested security tests and adverse red taming.

In a follow-up message, a Google spokesperson told TechCrunch that it continued as a “top priority” for the security agency, and it was planned to publish more documentation around its AI models, including Jemie 2.0 flash. Gemi 2.0 flash, which is usually available, also lack a model card. The latest model card was published by Google Gemini 1.5 ProWhich came out more than a year ago.

System cards and model cards are useful – and unforgettable, from time to time – provide information that companies do not always advertise their AI. For example, the system card published for the OpenAIA and 1 reasonable model has revealed that the company There is a tendency to “scheme” against people And secretly follow its own goals.

Greaterly, the AI ​​community realizes these reports as a good faith attempt to support independent research and protection evaluation, but in recent years reports have taken extra importance. As Transformer As mentioned earlier, Google told the US government in 2023 that it would publish a safe report for publication of all “notable,” public AI Model “. “Agency Gave similar promises From Other governmentPromise to “supply public transparency”.

There has been regulatory efforts at the Federal and State level in the United States to create a security report value for AI model developers. However, they have been limited and successful. Was one of the more significant efforts Vetode California Bill SB 1047Which was strongly opposed to the technology industry. Lawmakers have also published laws that will approve the Standard-Setting Body of the United States AI Safety Institute, the United States, Place the guideline for model releaseThe However the security Institute is now facing Possible Under the administration of the Trump.

From all presence, some promises to report on Google Model Exam are falling behind while the models are shipping faster than ever. This is a bad example, a lot Expert Argument – especially these models become more capable and sophisticated.