Google said Friday that the company is expanding Gemini’s latest in-depth research mode to 40 more languages.
The company launched the in-depth research mode earlier this month, allowing Google One AI premium plan users to unlock an AI-powered research assistant of sorts. The in-depth function works in a multi-step method, from creating a research plan to finding relevant information. Then, based on that information, the tool performs a search again to extract knowledge. After repeating that process a few times, it creates a report.
Gemini’s supported languages include Arabic, Bengali, Chinese, Danish, French, German, Gujarati, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malayalam, Marathi, Polish, Portuguese, Swahili, Spanish, Tamil, Telugu, Thani, Ukrainian, and Urdu.
The challenge for Google is to find reliable sources in a particular language, and then summarize it in the native language without messing up the grammar.
In conversation with TechCrunch in early December, HyunJeong Choe, director of engineering for the Gemini app, said that while the company trains the model using clean data and trustworthy sources, Google’s AI overviews in native languages like Hindi tend to have inaccuracies in summaries.
“We generally rely on native sources of the data, and we also use Google search on the back end to ground that information. Additionally, we run evaluations and fact-checks in native language data before rolling out the model,” Choe said.
“Factuality or getting correct information is a well-known research problem for generative AI in general. While the model has a lot of information in the pre-training mode already, we are focusing on training the model to use the information in the right way,” Choe said.
Jules Walter, product lead for international markets for the Gemini app, said that the company has testing programs to get quality checks from native perspectives. He mentioned that the company generates data to train models. Plus, local teams review those datasets too.
Earlier this week, TechCrunch reported that a contracting firm working to improve Gemini by rating responses passed down guidelines from Google that contractors were no longer allowed to skip prompt responses, irrespective of their expertise.
After that report was published, a Google spokesperson said that contractors not only rate answers for content but also look at style, format, and other factors.
TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.
