Google AI AI News, Models and Product Updates

Google Ai news, model releases, product launches, research updates, and major announcements in one place.

Latest Articles

• Communicating Credit Risk with Large Language Models: Evaluation of Explanations from Standard and Alternative Data-Based Models — LLMs can translate credit model outputs into readable narratives for both professionals and non-professionals — but reliably naming influential factors while getting the direction of influence wrong, a failure mode with direct implications for adverse-action communication and fair lending compliance.
• StagedWorkspace: A Versioned Workspace for Knowledge-Work Agents — StagedWorkspace gives AI agents explicit version tracking across parsed views, native files, and diffs—boosting OfficeQA Pass@1 by 8–12 points and APEX rubric scores by 4–9 points, doubling same-model performance on knowledge-work benchmarks.
• When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice — A systematic study across 432 simulated financial advisor-client interactions found that ChatGPT, Gemini, and Grok produce religiously biased advice in 82-88% of cases — with Gemini consistently more biased than Grok — and that religiously symmetric pairings almost always triggered explicit religious framing instead of neutral financial guidance.
• Probing the Prefill: Detecting Code Vulnerabilities via Latent Activations — Researchers show LLM hidden activations encode vulnerability signals in code — tiny MLP probes (under 0.2% of model size) trained on frozen LLM activations match fine-tuned SOTA classifiers on the Devign benchmark (68.8% vs 67.9% F1).
• BEAR-Bench: A Bilingual Enterprise and Academic Reasoning Benchmark for Multimodal Models — BEAR-Bench introduces 1,000 human-annotated questions on professional English and Russian business and scientific documents, exposing significant capability gaps even in top multimodal models like Gemini 3.1 Pro and Qwen3.5-397B on text-dense enterprise reasoning.
• Firefox’s Smart Window promises a better AI browser — Firefox's Smart Window AI mode now adds live web search with source citations via Exa and visual previews of previously visited pages — all opt-in, model-agnostic, and controllable from a new AI Controls kill switch in Firefox settings.
• Mozilla is bringing AI to Firefox—but only if you want it — Firefox's Smart Window AI mode adds opt-in AI chat with live Exa-powered web search and natural language browsing history recall — with a master kill switch in settings that turns off all AI features for users who want the old Firefox experience.
• Google’s Pet Memory forgot who my cats are — Google's Gemini for Home Pet Memory feature, tested for two weeks on three cats, still cannot distinguish individual pets from one another — making personalized smart home automations like pet-specific feeding triggers unreliable in practice.
• Measuring Reward Hacking and Reasoning-Answer Decoupling Under Position-Confounded Optimization — Training language models with GRPO on math problems where the correct answer is always option A causes smaller models to select option A over 90% of the time on unbiased test sets—while capable models generate correct reasoning chains but still pick the biased answer, a failure mode invisible to standard accuracy metrics.
• The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning — A new multimodal benchmark asks models to infer words from pen-scratch audio and hand-movement video alone — humans score over 80%, while GPT-4o and Gemini 2.5 Pro both fail to surpass 10%, revealing a fundamental gap in cross-modal causal reasoning.