Google unveils new 'open' AI models emphasizing safety

Google has introduced a trio of new, “open” generative AI models — Gemma 2 2B, ShieldGemma, and Gemma Scope — which it claims are safer, smaller, and more transparent than most available models.

These additions to Google's Gemma 2 family, first launched in May, aim to enhance safety in AI applications.

Gemma 2 2B is a lightweight model designed for generating and analyzing text, compatible with various hardware including laptops and edge devices. It's available for certain research and commercial uses through Google’s Vertex AI model library, Kaggle, and Google's AI Studio toolkit.

ShieldGemma is a set of “safety classifiers” built on top of Gemma 2, designed to detect and filter out toxic content such as hate speech, harassment, and sexually explicit material.

Gemma Scope allows developers to delve into specific aspects of a Gemma 2 model to make its operations more interpretable. According to Google, it uses specialized neural networks to unpack complex information, helping researchers understand how Gemma 2 identifies patterns and makes predictions.

This release follows the U.S. Commerce Department's preliminary endorsement of open AI models, which it said can make generative AI more accessible to smaller companies, researchers, and nonprofits while also highlighting the need for monitoring these models for potential risks.