Google announced on Friday that the latest in-depth research feature of its Gemini platform is being extended to support 40 additional languages.
This advanced research capability, introduced earlier this month, is accessible to users subscribed to the Google One AI premium plan. It functions as an AI-driven research assistant that employs a multi-step process: beginning with creating a structured research plan, it gathers relevant data and performs repeated searches to refine insights. The output is then consolidated into a comprehensive report.
The expanded list of languages supported by Gemini now includes 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.
One of Google’s primary challenges is ensuring the accuracy of summarized information while preserving grammatical integrity in native languages. HyunJeong Choe, the director of engineering for Gemini, acknowledged in a December interview with TechCrunch that summaries in languages like Hindi sometimes exhibit inaccuracies, despite efforts to train the model with reliable data sources.
Choe explained that the system relies heavily on native language sources and integrates Google Search to validate information. Before deploying updates, Google conducts evaluations and fact-checks using native data. “Factual accuracy is a recognized challenge in generative AI,” Choe stated. While the model’s pre-training equips it with extensive knowledge, the focus remains on refining its ability to use this information effectively.
Jules Walter, Gemini’s product lead for international markets, highlighted ongoing quality assurance efforts. He described programs that gather native perspectives to ensure high-quality outputs, noting that locally sourced teams review datasets used to train the model.
Separately, a report earlier in the week revealed that Google’s contractors, who are tasked with assessing Gemini’s responses, were instructed to rate all responses without skipping prompts, regardless of their expertise. A Google spokesperson later clarified that these contractors evaluate not just content accuracy but also aspects like style and formatting, ensuring well-rounded assessments.