Google has released an essential report that details how its latest AI model, Gemini 2.5 Pro, was constructed and evaluated, three weeks after making it publicly available as a “preview” version.
AI governance experts had blasted the company for allegedly breaking pledges it made to the U.S. government and at several international AI safety conferences by releasing the model without disclosing documentation of safety assessments it had conducted and potential risks the model might pose.
A Google spokeswoman claimed in an emailed statement that any claims that the business had broken its promises were “inaccurate.” Additionally, the corporation stated that a more thorough “technical report” will be released after the Gemini 2.5 Pro “model family” is completely accessible to the general public.
However, at least one AI governance specialist has also criticized the recently released six-page model card for offering “meager” details regarding the model’s safety assessments.
In a lengthy post on social media site X, Kevin Bankston, senior advisor on AI Governance at the Center for Democracy and Technology, a think tank based in Washington, D.C., expressed concern over the model card’s lack of clarity and delayed distribution.
“As companies rush their models to market, this scant documentation for Google’s top AI model tells a troubling story of a race to the bottom on AI safety and transparency,” he added.
He claimed that Google hadn’t finished its safety testing before releasing its most potent model and that it still hasn’t done so as of right now because of the model card’s delayed release and the absence of important safety evaluation results, such as information about “red-teaming” tests to trick the AI model into serving up dangerous outputs like bioweapon instructions.
According to Bankston, it’s also possible that Google has completed its safety assessment but has implemented a new policy that prevents it from disclosing the evaluation’s findings until the model is made available to all Google users. At the moment, Google refers to Gemini 2.5 Pro as a “preview.” It is available through Google Labs and Google AI Studio, but there are some restrictions on what users may do with it. According to Google, the model will be broadly accessible to college students in the United States.
According to the Google representative, the business will publish a more thorough AI safety report “once per model family.” According to Bankson on X, this could imply that Google will no longer provide independent evaluation findings for more refined models, like those designed for cybersecurity or coding. He pointed out that this could be risky since refined AI models can display behaviors that differ significantly from the “base model” from which they were derived.
There are other AI companies that appear to be backing off on AI safety besides Google. Experts in AI safety also questioned Meta’s model card for its recently released Llama 4 AI model, which is comparable in length and detail to the one Google recently provided for Gemini 2.5 Pro. Since the company’s “chain of thought” reasoning models, such o3 and o4-mini, outperformed the recently published GPT-4.1 model on numerous benchmarks, OpenAI stated that it was not issuing a technical safety report for the model, claiming that it was “not a frontier model.”
OpenAI said that GPT-4.1 was more powerful than its GPT-4o model, whose safety assessment had revealed that the model might present certain dangers, but that these were far below the point at which the model would be deemed hazardous for public release. It’s unclear if GPT-4.1 will now surpass those limits because OpenAI stated it had no plans to release a technical report.
With the launching of its new o3 and o4-mini models on Wednesday, OpenAI did provide a technical safety report. At the same time, however, it revised its “Preparedness Framework” earlier this week, outlining how the business will assess its AI models for critical risks—from assisting someone in creating a biological weapon to the potential for a model to start improving itself and eluding human control—and work to reduce those risks.
The update removed “Persuasion,” which refers to a model’s capacity to manipulate someone into destructive behavior or to believe false information, as a risk category that the business would evaluate prior to release. Additionally, it altered the way the business would decide whether to release more risky models. For example, it stated that if a competitor had already released a model that was similar, the business would think about shipping an AI model that posed a “critical risk.”
The update removed “Persuasion,” which refers to a model’s capacity to manipulate someone into destructive behavior or to believe false information, as a risk category that the business would evaluate prior to release. Additionally, it altered the way the business would decide whether to release more risky models. For example, it stated that if a competitor had already released a model that was similar, the business would think about shipping an AI model that posed a “critical risk.”
AI governance specialists had differing opinions on such adjustments. While some praised OpenAI for being open about its methodology and clearly outlining its release standards, others expressed alarm.
