Google DeepMind

Are LLMs stubborn or oversensitive to pushback? Both — at once. Our new paper in Nature Machine Intelligence identifies two competing biases in how LLMs handle their own confidence. First: LLMs become more confident in their initial answers simply because they gave them before — a choice-supportive bias established in human cognition, but striking in a stateless model with no memory of having provided a confidence rating before. Second: when challenged, LLMs markedly overweight opposing advice, updating 2–3× more strongly than a Bayesian ideal observer — and changing their minds far more often than warranted. Notably, this is asymmetric — they don't comparably overweight advice that agrees with them, distinguishing this from simple sycophancy. These biases coexist, pull in opposite directions, and generalise across multiple models — from factual queries to math problems. Joint work with Google DeepMind and the UCL Institute of Cognitive Neuroscience. 📄 Open access: https://rdcu.be/feOjz

Decoupled DiLoCo is our latest approach helping train AI models across multiple distant data centers. This process normally relies on identical chips staying in near-perfect synchronization. If a single chip fails, the entire training run can stall. With Decoupled DiLoCo, we explored a way to train across a global network. Here are some of our results: 🔘 We trained a 12B parameter model simultaneously across four US regions - so  we are no longer constrained by the size of a single centre. 🔘 The system seamlessly mixes older and newer chip generations without slowing down, unlocking more value from existing hardware. 🔘 If hardware breaks mid-run, it isolates the failure and keeps training. We look forward to continuing to evolve our systems into more resilient, useful tools - helping us develop the next generation of AI. Find out more → https://goo.gle/4mNE36q 

Only 25% of organizations have moved AI into production at scale. We’re working to change that. 🛠️ Accenture, Bain & Company, Boston Consulting Group (BCG), Deloitte, and McKinsey & Company are combining our advanced research with their expertise to bring AI innovation to more industries responsibly. 🤝 Find out more → https://goo.gle/42oa37A

Deep Research and Deep Research Max are our latest autonomous research agents powered by Gemini 3.1 Pro. They can safely navigate both the web and your custom data, like internal docs and specialized financial information, to create professional-grade, fully cited reports. 🔵 Deep Research: Optimized for speed and efficiency. Perfect for interactive apps needing quicker responses. 🔵 Deep Research Max: It uses extra time to search and reason. Ideal for exhaustive context gathering and tasks happening in the background. Now featuring arbitrary MCP support to securely connect and analyze your own or third-party data. Plus, it’s our first research agent to natively generate presentation-ready visuals that bring data to life. Start building via the Gemini API → https://goo.gle/4tAJBDC

Announcing the AI for Science European Summer School https://lnkd.in/e3xeWXvY Join us in Athens, Greece on July 16 and 17, 2026 for an exciting two day program at the intersection of AI and Scientific Discovery. The workshop will be hosted at the beautiful Acropolis Museum. The summer school is organized by Hellenic Institute of Advanced Studies, sponsored by Google DeepMind and will feature a stellar lineup of leading researchers and innovators. We warmly invite applications from early career scientists and industry members eager to engage with cutting edge developments in AI for Science. 👉 Apply by May 10th https://lnkd.in/ecZYcjd7 🎉 A limited number of fellowships are available to cover travel and attendance costs for young scientists.

Gemini 3.1 Flash TTS is our most controllable text-to-speech model yet. With new Audio Tags, you can easily direct vocal style, delivery, and pace through text commands. Updates and features: 🔵 More natural sounding speech 🔵 Support for 70+ languages like Hindi, Japanese, and German  🔵 SynthID watermarking on all outputs Developers can start building in preview via the Gemini API and Google AI Studio, rolling out to enterprises in preview in Vertex AI. Rolling out to everyone in Google Vids. Find out more → https://goo.gle/4mrqnxy

We’re giving robots a better understanding of the physical world with Gemini Robotics-ER 1.6. 🤖 It’s significantly better at spatial and physical reasoning, meaning: 👉 It can use pointing to precisely identify objects ✅ Agents can intelligently choose where to retry a task or move forward 👓 Robots are able to read and interpret a variety of instruments. Gemini Robotics-ER 1.6 is also our safest robotics model yet, and able to recognize physical constraints like avoiding liquids or items over 20kg when following instructions. It is also 10% better at detecting human injury risks in videos. The model is available now on Google AI Studio, and the Gemini API. Find out more. ↓ https://goo.gle/4cuBh1f

What’s new in Gemma 4? Gemma 4 is our newest family of open models, bringing the same world-class research and technology from Gemini 3 directly to your local hardware. This means you can now run advanced reasoning, native vision and audio, and agentic tool-use on anything from high-end workstations to mobile phones. Learn more → https://goo.gle/4cb8LBE

Gemma 4 is here. 💻 We’ve built a new family of open models based on the same world class research and tech as Gemini 3. “Open” means the model weights are yours to download, customize, and run on your own hardware. ⚖️ Four sizes: High-performance versions for workstations (31B Dense & 26B MoE) and highly optimized “Edge” versions (E4B & E2B) built specifically for mobile. 🧠 Advanced reasoning: Capable of multi-step planning and deep logic with native vision and audio support.  🤖 Built for agents: Native tool use lets you build autonomous systems that can actually do things, like search databases or trigger APIs. 🔒 Apache 2.0 License: Complete flexibility to build, fine-tune, and deploy however you want. Start building with Gemma 4 now in Google AI Studio. You can also download the model weights from Hugging Face, Kaggle, or Ollama. Find out more → https://goo.gle/4cb8LBE

We’re sharing new research into how AI might be misused to exploit emotions or manipulate people into making harmful choices.  Our study of 10,000 people reveals that AI manipulation is dependent on domain. While models demonstrated high influence in finance, they hit a wall in healthcare, where existing guardrails helped block false medical advice. By identifying red flag tactics – like using fear – we can build stronger protections. We’ve built an empirically validated, first-of-its-kind toolkit to measure AI manipulation in the real world – to better understand how it can occur and help protect people. Find out more → https://goo.gle/4bMruTP