Experts say that the latest report of the artificial intelligence model lacks the main safety details
On Thursday, weeks after the launch of the strongest artificial intelligence model so far, Gemini 2.5 ProGoogle A published Technical report Show the results of internal safety assessments. However, the report shines on the details, as experts say, which makes it difficult to determine the risks that the model may pose.
Provides technical reports useful – and unequalSometimes – information that companies do not always announce on a large scale about artificial intelligence. In general, the artificial intelligence community believes that these reports are making good efforts to support independent research and safety assessments.
Google follows a different approach to reporting safety from some of its competitors of artificial intelligence, as technical reports are published only once you are considered a model that comes out of the “experimental” stage. The company also does not include results from all its “dangerous ability” reviews in this writing; It maintains separate scripts.
Many experts who spoke with Techcrunch are still disappointed due to the variation Border safety framework (FSF). Google FSF presented last year in what he described as an attempt to determine the capabilities of future artificial intelligence that could cause “severe damage”.
“this [report] “It is very little, and it contains minimal information, and came out weeks after the form of the form available to the public,” said Peter Wildfiorb, co -founder of the Institute of Politics and Strategy of IQ.
Thomas Woodside, co -founder of the AI Secure project, said that although happy, Google has released a report on Gemini 2.5 Pro, but he is not convinced of the company’s commitment to providing complementary safety assessments in a timely manner. Woodside noted that the last time Google published the results of dangerous ability tests was in June 2024 – a model announced in February in the same year.
Do not inspire much confidence, Google did not provide a report on it Gemini 2.5 flashSmaller and more efficient model the company announced last week. A TECHRUNCH spokesman said a Flash report that is “soon.”
“I hope this is a promise of Google to start publishing more frequent updates,” Woodside told Techcrunch. “These updates should include the results of the models’ evaluation evaluations that have not been publicly published, because these models may also pose serious risks.”
Google may have been one of the first artificial intelligence laboratories that suggested unified reports of models, but it is not the only one that existed Accused recently. She was released dead Likewise, high safety assessment Again Lama 4 Open models, Openai chose not to publish any report for them GPT-4.1 series.
The Google head is the assurances made by the technology giant for organizers to maintain a high level of artificial intelligence testing and reports. Two years ago, Google told the US government Safety reports will be published for all important artificial intelligence models “within the scope. Company Follow this promise with similar obligations to Other countriesPledge to “provide public transparency” about artificial intelligence products.
Kevin Bankston, the chief adviser to artificial intelligence governance at the Center for Democracy and Technology, described the intermittent and mysterious trend to “race to the bottom” on the integrity of artificial intelligence.
“By assisting reports that competing laboratories such as Openai have shaved the time of safety test before release from months to days, these fierce documents of the best AI from Google tells a worrying story about a race to the bottom on the integrity of artificial intelligence and transparency with companies rushing their models to the market.”
Google said in the phrases that, although it is not detailed in its technical reports, it is conducting a safety test and the “Red Teaming Team” of the characteristics of the models that were released.
