Google announced Friday that the company is expanding Gemini's latest in-depth investigation mode to 40 additional languages.
The company launched deep exploration mode earlier this month, allowing users of its Google One AI premium plan to unlock a type of AI-powered research assistant. Advanced features work in multiple steps, from creating a research plan to finding relevant information. Then, based on that information, the tool performs the search again to extract knowledge. After repeating this process several times, a report will be created.
Languages supported by Gemini include Arabic, Bengali, Chinese, Danish, French, German, Gujarati, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malayalam, Malay Tea, Polish, Portuguese, Swahili, Spanish, Tamil, Telugu, Thani, Ukrainian, and Urdu.
The challenge for Google is to find reliable sources of information in a particular language and summarize it in their native language without making grammatical mistakes.
HyunJeong Choe, director of engineering for the Gemini app, said in a conversation with TechCrunch in early December that while the company uses clean data and trusted sources to train its models, He said Google's AI summaries written in Japanese tend to be inaccurate summaries.
“We typically rely on native sources of data and also use Google search on the back end to ground that information. We also evaluate native language data and facts before deploying models. We run checks,” Choe said.
“Obtaining factuality or correct information is a well-known research problem in generative AI in general. Models already contain a lot of information in pre-training mode, but we need to use the information appropriately. We're focused on training the model to do that,” Choe said.
Jules Walter, product lead for international markets for Gemini apps, said the company has a testing program to ensure quality checks from a native perspective. He said the company is generating data to train models. In addition, local teams also review those datasets.
Earlier this week, TechCrunch reported that contractors working to improve Gemini by rating responses were told that Google's guidelines no longer allow contractors to skip quick answers, regardless of their expertise. Reported.
After the report was released, a Google spokesperson said that contractors not only evaluate responses for content, but also consider style, format, and other factors.