Google’s latest AI model report lacks key safety details, experts say
On Thursday, weeks after launching its most powerful AI model so far, Gemini 2.5 Pro, Google has published a technical report Show the results of the internal safety evaluations. However, the report is light on the details, experts say, making it difficult to determine which risks the model can be.
Technical reports offer useful – and sometimes not – flattering – information that companies do not always advertise on a large scale about their AI. In general, the AI community sees these reports as good wedding efforts to support independent research and safety evaluations.
Google uses a different safety report -approach than some of its AI rivals and only publishes technical reports when it regards a model as a graduate in the ‘experimental’ phase. The company also does not include any findings from all its evaluations of “dangerous capacities” in these descriptions; It reserves it for a separate audit.
Several experts who spoke WAN were still disappointed by the sparsity of the Gemini 2.5 Pro report, of which they noticed Frontier Safety Framework (FSF). Google introduced the FSF last year into what it described as an attempt to identify future AI possibilities that can cause ‘serious damage’.
“This [report] Is very scarce, contains minimal information and weeks after the model was already made available to the public, “Peter Wildeford, co-founder of the Institute for AI policy and strategy, WAN said.” It is impossible to verify whether Google realizes its public obligations and therefore impossible to assess the safety and security of their models. “
Thomas Woodside, co-founder of the Secure AI project, said that although Google has issued a report for Gemini 2.5 Pro, he is not convinced of the company’s dedication to provide timely safety evaluations. Woodside pointed out that the last time that Google published the results of dangerous capacity tests in June 2024 – for a model announced in February that same year.
Google has not inspired much self -confidence and has not made a report available for Gemini 2.5 Flash, a smaller, more efficient model that the company announced last week. A spokesperson told WAN that a report for Flash will be ‘soon’.
“I hope this is a promise from Google to start publishing more frequent updates,” Woodside told WAN. “Those updates must contain the results of evaluations for models that have not yet been publicly used, because those models can also form serious risks.”
Google was perhaps one of the first AI laboratories that proposes standardized reports for models, but it is not the only one accused of subdivision of transparency lately. Meta has one Likewise poor safety evaluation Of his new Lama 4 open models, and OpenAi chose not to publish a report for his GPT-4.1 series.
Hanging above the head of Google are the guarantees that the technology giant has made for regulators to maintain a high standard of AI safety tests and reporting. Two years ago, Google told the US government It would publish safety reports for all “important” public AI models “within Scope”. The company followed that promise with similar obligations Unpleasant Other countriesPromises to “offer public transparency” around AI products.
Kevin Bankston, a senior adviser for AI Governance at the Center for Democracy and Technology, the trend of Sporadic and vague reports a “race to the bottom” on AI safety.
“Combined with reports that competing laboratories such as OpenAI have shaved their safety test before release of months to days, this lean documentation for the Top AI model of Google tells a disturbing story about a race to the soil on AI safety and transparency while companies bring their models to the market,” he said Te Tech.
Google has said in statements that, although not detailed in its technical reports, the safety tests and “opponents Red Teaming” performs for models prior to release.
