Mozilla

Mozilla is focused on what we can directly influence: the tools we build and support. In the #StateOfMozilla, learn how that work is shaping the future of AI. The web’s next chapter is being written now. https://lnkd.in/gidPHDPX

The last decade of AI was defined by scale. The next will be defined by adaptability. Today, Mozilla Ventures is proud to back adaption, a new AI company building systems that evolve with people — not the other way around. As AI becomes more expensive and centralized, Adaption is taking a different path: efficiency-first, context-aware AI that adapts across languages, cultures, and real-world needs. Founded by Sara Hooker and Sudip Roy, Adaption reflects a shared belief in open, community-driven technology and AI that works for more people. Everything intelligent adapts. AI should too. https://lnkd.in/gahwXzHN

Mozilla-sponsored hackathon on Saturday! Raffi Krikorian, Mozilla's CTO will be speaking. Build open source AI with us and apply today. Hack-Nation's Global AI Hackathon https://lnkd.in/gxpcjUTK

New Indonesian Language Datasets are now available on Mozilla Data Collective! 🎉 Our community members have just released new TTS and ASR datasets for Javanese, Tolaki, and Mandar language research: • TTS Central Javanese - 440.11 MB of TTS data for Javanese, focusing on the Ngoko variety of speech formality and containing instances of Indonesian and English code-switching. https://lnkd.in/esziuB8A • Mandar Spontaneous Speech - 534.34 MB of Spontaneous Speech ASR data, primarily containing dialects used in Majene and Polewali Mandar. https://lnkd.in/ewshNKiU • TTS-Tolaki - 249.04 MB of Tolaki Language data containing cultural heritage stories. https://lnkd.in/eMVaAhRA • TTS Javanese-Lumajang Dialect - 684.32 MB of TTS data created in response to 690 prompts in the Lumajang dialect of Javanese with code-mixing in Indonesian. https://lnkd.in/eagWhXQF 🔬 Browse all datasets: https://lnkd.in/e7Xp3dqd #OpenData #AI #ASR #TTS #NLP #MachineLearning #SpeechRecognition #LanguageTech

New Virtual Session: Encryption and Feminism — Reimagining Child Safety Without Surveillance
🗓️ Feb 10 | 4:00 PM GMT | Online Missed this session at Mozilla Festival? Couldn’t make it to Barcelona or didn’t get into the room? We’re bringing the conversation online so more people can experience it in full. We’ll revisit key panel insights and explore emerging Feminist Encryption Principles, shaped by participant questions and collective discussion. Panelists include:
Hera Hussain (CHAYN), Georgia Bullen (Superbloom), Lucy Purdon (@Courage Everywhere), Gerda Binder (UNICEF), and Mallory Knodel (IX). Moderated by Ramma Shahid Cheema Shahid and Audrey H. (Internet Exchange) If you believe child protection should not come at the cost of privacy and that privacy is a right, not a privilege, we hope you’ll join us. Register here:
https://lnkd.in/ess7rXpf

When I was wrapping up my fellowship project at Mozilla on multilingual AI safety evaluation (www.multilingualailab.com), my practitioner mind was telling me: 💡 Now what? Yes, I built a tool to evaluate and document multilingual inconsistencies in large language models. But if models behave differently across languages and contexts, shouldn't context and language-aware evaluation lead to the design of context and language-aware safety guardrails? Thanks to Lisa Gutermuth and Davide Eynard, I was introduced to mozilla.ai's Daniel Nissani any-guardrail, a unified interface for working with AI safety guardrails. Together with Daniel, we used insights from my fellowship project's evaluation results in a humanitarian context to define safety criteria and test existing, customizable guardrails supported by any-guardrail. We asked: ▪️ How do LLM safety guardrails -- often LLM-powered themselves -- behave when LLM responses are in non-English languages? ▪️ Does the language a policy is written in (for example, Farsi vs. English) affect guardrail decisions? ▪️ What are the real-world safety implications of this behavior in a humanitarian use case? We tested three different guardrails using English and Farsi policy variants and humanitarian scenarios. We observed similar flaws we see in LLMs themselves -- hallucinations, biases, and multilingual inconsistencies -- even when policies and prompts were semantically equivalent across languages. 💡 Another, now what? Based on these findings, we developed a set of recommendations and a path forward for this work, including: the importance of guardrails having agentic capabilities such as access to factuality checks, retrieval, and search; and guidance on writing policies in ways that minimize language-based inconsistencies and the noise and randomness of results. In light of the growing recognition of the need for custom, language-specific evaluation and benchmarks, I'm sure/hopeful that in 2026 we'll see more work on the evaluation-to-guardrail-design pipeline.

AI is being forced across the web, often without a clear choice.  At Mozilla, we believe the user should always be in control. Starting with Firefox 148, we’re introducing AI controls, a single place to manage, customize, or entirely block AI features. When AI is blocked, Firefox hides AI entry points and stops AI-related prompts, meaning there are no pop-ups, reminders, or nudges to use AI features. Whether you want to use AI for better tab group suggestions or prefer to keep your browsing environment traditional, the choice is yours. https://lnkd.in/gqAVKRbY

The Mozilla Manifesto has guided our work for over two decades. As AI reshapes the internet, we’re doubling down on the principles that matter most right now — protecting human agency, supporting open and decentralized innovation, and ensuring public benefit keeps pace with commercial power. This is how we approach the future of AI. https://lnkd.in/gGAq8Yne

AI is being forced across the web, often without a clear choice.  At Mozilla, we believe the user should always be in control. Starting with Firefox 148, we’re introducing AI controls, a single place to manage, customize, or entirely block AI features. When AI is blocked, Firefox hides AI entry points and stops AI-related prompts, meaning there are no pop-ups, reminders, or nudges to use AI features. Whether you want to use AI for better tab group suggestions or prefer to keep your browsing environment traditional, the choice is yours. https://lnkd.in/gqAVKRbY

The concerns about AI are valid. One solution? Better data that’s open and free. In 2026, Mozilla Foundation is expanding the Mozilla Data Collective—a product that enables responsible collection and sharing of training datasets. Built as an extension of Common Voice, which to date has created open datasets in 358 languages, it supports open, high-quality data designed in the public interest. Learn more about Mozilla Data Collective and sign up to become a contributor here: https://lnkd.in/expufMvB And dig into the full State of Mozilla here: https://lnkd.in/e6W5cetQ