What Google is in AI
Google — through its research division Google DeepMind — is one of the most active AI organizations in the world. It builds the Gemini family of frontier models (its current flagship is Gemini 3.1 Pro), releases open-weight models under the Gemma brand that anyone can download and run, and deploys AI across its own products like Google Search, Google Workspace, and NotebookLM. It also supplies AI infrastructure and models to other companies, including competitors.
Why it matters
Google sits at almost every intersection of the AI world. It is simultaneously a model builder, a cloud provider, an open-source contributor, a research lab, and a product company with billions of users. That breadth means decisions Google makes — about what to release, how to price it, what safety rules to follow — ripple across the entire industry.
The Gemini model family
Google's main model line is Gemini. The current flagship, Gemini 3.1 Pro, competes at the top of the market against OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8. For faster, cheaper tasks, Gemini 3.5 Flash is the workhorse — and it recently gained computer use, meaning it can control software on a screen autonomously, the same kind of capability that lets an AI book a flight or fill out a form without human hand-holding.
Google also released Gemini 3.5 Live Translate, which handles real-time translation across 70+ languages, and integrated Gemini into NotebookLM, its research assistant tool.
Open models: Gemma
Not everything Google builds stays behind a paywall. The Gemma 4 family is a set of open-weight multimodal models — meaning they can understand both text and images — that developers can download and run on their own hardware. Gemma 4 uses a unified architecture that eliminates the separate vision component most multimodal models require, making it more efficient. A specialized 27-billion-parameter Gemma-based model was also released for single-cell biological analysis, contributing to the discovery of a potential cancer therapy pathway.
The Apple partnership
One of the biggest recent developments for Google is its deepening relationship with Apple. Apple rebuilt its on-device AI architecture around Google Gemini models, meaning Gemini now powers AI features on iPhones and Macs. Apple's new Foundation Models (AFM 3) family was distilled from Gemini and runs locally on Apple silicon. This is a significant win: it puts Gemini inside one of the world's most widely used device ecosystems.
Scientific research tools
Google DeepMind has been particularly active in applying AI to science. AlphaGenome is an open-weights model that interprets the roughly 98% of human and mouse DNA that doesn't code for proteins but regulates how genes are expressed — a frontier problem in biology. It matched or exceeded prior models in 47 out of 50 evaluations and correctly predicted gene expression changes linked to a type of leukemia.
The Aletheia agentic system used Gemini 3 Deep Think to tackle unsolved mathematical problems from the Erdős problem set, producing 4 genuinely novel solutions out of 200 evaluated.
CodeMender, announced by DeepMind, is an AI agent focused on identifying and fixing critical software security vulnerabilities in real-world codebases.
Safety, regulation, and legal exposure
Google is navigating a more complex regulatory environment than it faced even a year ago. On the government side, it agreed to submit models to the NIST TRAINS task force — a new U.S. multi-agency body that reviews frontier AI models for cybersecurity, biosecurity, and chemical weapons risks before public deployment. Google is also a member of Project Glasswing, Anthropic's consortium of technology companies working to proactively patch software vulnerabilities discovered by powerful AI models.
On the legal side, a German court ruled that Google can be held liable for defamatory false statements generated by its AI Overview feature — the AI-generated summaries that appear at the top of Google Search results. The court found that AI Overview outputs count as "independent, new, and substantive statements," rejecting Google's argument that it was just an intermediary passing along information. Google plans to appeal, but the ruling could trigger similar cases across Europe.
Separately, research found that a gray-market proxy network was selling access to a fake "Gemini-2.5" API in China at steep discounts — with the proxy version scoring only 37% on a medical benchmark versus 84% for the real model, suggesting widespread model substitution.
Where things are heading
Google's trajectory points toward deeper integration across devices (via Apple), more capable agentic tools (Gemini 3.5 Flash with computer use, Spark background agents), and continued scientific applications. At the same time, the German court ruling and the U.S. government's growing appetite for pre-deployment oversight signal that Google — like all frontier AI companies — will face increasing scrutiny over what its models say and do in the world.
