Google released on Google chatgptthe company has greatly improved its pace.
In late March, Google launched the AI inference model. Gemini 2.5 ProThat Lead the industry On several benchmarks for measuring encoding and mathematical functions. The tech giant launched just three months after it debuted another model Gemini 2.0 flashthat was the most advanced at that time.
Tulsee Doshi, Google’s Gemini and Product Leader, told TechCrunch in an interview that the growing launch of the company’s model is part of a concerted effort to align with the rapidly growing AI industry.
“We are still working to figure out what is the right way to get these models out – the right way to get feedback,” Doshi said.
However, the release time frame seems to come at a price. Google No security report for its latest model has been releasedincluding Gemini 2.5 Pro and Gemini 2.0 Flash, raises concerns that companies will prioritize speed over transparency.
Today, the standards for Frontier AI labs, including OpenAI, Human, and Metabolic, are fairly standard, reporting safety testing, performance evaluation, and use cases whenever they launch a new model. Sometimes called system cards or model cards, these reports were proposed several years ago by researchers in the industry and academia. Google is actually one of the first people to suggest model cards 2019 research paper, Call them “a method of responsible, transparent and responsible practice in machine learning.”
Doshi told TechCrunch that the company has not released a model card for Gemini 2.5 Pro, because it believes the model is an “experimental” version. The purpose of these experimental versions is to release the AI model in a limited way, get feedback, and iterate the model before production releases.
According to Doshi, Google intends to release a model card for the Gemini 2.5 Pro when the model is usually available, adding that the company has conducted security testing and the confrontational red team.
In a follow-up message, a Google spokesperson told TechCrunch that security remains the company’s “top priority” and plans to move forward with more documentation around its AI model, including Gemini 2.0 Flash. Gemini 2.0 Flash (usually available) also lacks a model card. The last model card released by Google is for Gemini 1.5 Prothis appeared a year ago.
System cards and model cards provide useful information, sometimes not even widely publicizing their AI. For example, the system card released for the O1 inference model OpenAI shows that the company’s Models have a trend towards human “plan”And secretly pursue your goals.
Overall, the AI community sees these reports as a goodwill effort to support independent research and security assessments, but in recent years, these reports have become even more important. As transformer It was previously noted that Google told the U.S. government in 2023 that it would release security reports that all “important” public AI models are released “within scope”. company Make a similar commitment arrive Other governmentspromises to “provide public transparency.”
The United States has already conducted regulatory work at the federal and state levels to create security reporting standards for AI model developers. However, their adoption and success are limited. One of the most famous attempts is Vetoed California Act SB 1047the technology industry strongly opposes it. The lawmakers also proposed legislation that would authorize the US Institute of AI Security, the US AI standard setting agency, Establishing a model release guideline. However, the Institute of Security is now facing Possible cuts Under the leadership of the Trump administration.
From all appearances, Google lags behind some of its promises about model testing reports, while shipping models at the same time faster than ever. This is a bad precedent, many expert Dispute – especially as these models become more capable and refined.
