The artificial intelligence (AI) developed by Google has managed to learn a foreign language without human help, Novaya nauka (New Science) reports.

In an interview with CBS, Google CEO Sundar Pichai told why the company approaches AI cautiously. The Mountain View giant has indeed shelved several AI-related projects in recent years out of an abundance of caution. In particular, language models were delayed, with Google citing security and image risks.

Partly because of this caution, OpenAI and Microsoft have pulled ahead of Google in the AI “race.”

“We all have to be responsible in each step along the way,” says Pichai.

In an attempt to be bolder in response to competitors, Google took a similar approach with the launch of Bard. But instead of rolling out the chatbot worldwide right away, the company opted for a phased rollout, starting with the US and UK.

Originally, Google relied on the LaMDA (Language Model for Dialogue Applications) model. But gradually, the company added more powerful language model capabilities, namely the Pathways Language Model (PaLM), which has 540 billion parameters, while LaMDA has only 137 billion.

According to Pichai, sometimes generative AIs behave in completely unexpected ways. According to Google's findings, language models can sometimes learn new skills on their own without the knowledge of the programmers.