谷歌发布 AI 科学进展综述:支持 300 种语言
谷歌高级副总裁 James Manyika 发表文章,总结其 AI 技术在健康、灾害预测及教育领域的长期应用成果。核心亮点是宣布其技术现已支持超过 300 种语言,覆盖全球 86% 的人口。这是一篇侧重宏观愿景与既有成就回顾的观点性文章,而非新技术或产品的即时发布。
报道正文
我们正在探索 AI 在健康、自然灾害与气候韧性、学习以及经济机遇方面所能实现的可能性。
[James Manyika
研究、实验室、技术与社会高级副总裁](https://blog.google/authors/james-manyika/)
您的浏览器不支持音频元素。
收听文章 [[duration]] 分钟
Google 正利用 AI 加速科学突破,从而惠及全球民众。这些工具已经能够更早地检测疾病,预测洪水、野火等自然灾害,并打破语言障碍。它们还在帮助学生更高效地学习,并为面临职业转型的劳动者提供支持。通过与世界各地的专家合作,Google 希望借助这项技术解决一些世界上最棘手的问题。
摘要由 Google AI 生成。生成式 AI 仍处于实验阶段。本文内容
今天,我们达成了一个重要里程碑,这是数十年 AI 研究与进步的成果见证:Google 的技术现已支持超过 300 种语言,覆盖 70 亿人口——占全球人口的 86%。为了帮助我们了解这些工具如何推动现实世界的机遇,我们今天还通过我们的 AI & Economy ATLAS(AI 与经济图谱)发布了 新的交互式洞察,这是关于全球真实用户如何使用 AI 的最全面分析。此外,仅在过去几周内,我们在科学领域也取得了一系列关键的 AI 进展,以造福人类:
- 我们利用 AlphaGenome Atlas 绘制了人类基因组中所有 90 亿种可能的单字母遗传变异的预测影响图谱,并向研究人员开放获取。
- 我们推出了 WeatherNext 3,这是我们最先进、最精准的全球天气模型,其提前一天或更长时间的降水预报准确率提高了 50%——且该模型已应用于我们的产品之中。
- 我们将全球健康、粮食安全和社会经济数据整合到一个统一的“行星预测引擎”(Planetary Prediction Engine)中,以预测行星危机——该引擎已在刚果民主共和国持续的埃博拉疫情中得到应用,并在美国用于通过 21 项 CDC 健康指标识别脆弱社区。
- 我们扩展了 AI 研究规模,助力降低航空业的气候影响——相关技术目前已应用于英国(与政府合作)及亚洲地区。
将所有这些工作串联起来的是什么?是这样一个信念:AI 的进步能够加速科学探索,从而直接改善人们当下和未来的生活。这是驱动我们在 AI 领域工作的核心动力之一。我们正聚焦于最关键的重点领域:实现疾病的可检测、可治疗和可预防,预测自然灾害,拓展学习机会,并为更多人释放经济潜力。
尽管前景令人振奋,但 AI 带来的益处并非理所当然。要将这些益处变为现实——并缓解其中的挑战与风险——需要全社会共同努力。虽然仍有大量工作待完成,但 AI 的进展已经让我们有底气去追求那些能造福人类的宏大目标,去提出并解决那些曾被认为不可能攻克的难题。

我们正在利用 AI 改进疾病检测与诊断,并更深入地理解健康状况,取得了以下进展:
-
深化科学发现: 2024 年,Demis Hassabis 和 John Jumper 因其在 AlphaFold 方面的工作共同获得诺贝尔奖。AlphaFold 已预测出科学界已知的所有 2 亿种蛋白质结构,为理解和研究疾病提供了新的基础。目前,来自 190 个国家的 400 万研究人员正在使用它,应用领域涵盖药物研发到理解被忽视的疾病(如 恰加斯病和利什曼病)。AlphaMissense 正帮助研究人员识别致病的基因突变。现在,我们在 AlphaFold 和 AlphaMissense 的基础上构建了 AlphaGenome Atlas,为科学家提供关于遗传变异如何改变细胞行为的预测性洞察。
-
更早检测: 我们与帝国理工学院及英国 NHS 合作开展的最新 乳腺癌研究 表明,AI 能够检测出在 17.5 万名女性的乳腺 X 光检查中此前被漏诊的 25% 的间期癌症。与此同时,我们在开发辅助检测 肺癌、结直肠癌 以及 肿瘤细胞中的基因突变 的工具方面也取得了实质性进展。
-
全球筛查: 针对结核病——全球约 40% 的感染者未被诊断——我们的胸部 X 光模型(由 Nexus Intelligence使用)已在六个国家的 40 个地点完成了超过 25,000 例 X 光片筛查。同时,我们正利用生物声学模型,通过咳嗽声检测结核病,相关技术基于 Health Acoustic Representations。此外,我们与合作伙伴共同开发的糖尿病视网膜病变模型已支持全球超过 115 万次筛查,并计划在未来十年内将筛查规模扩大至 600 万次,以助力发现这一可治疗但日益普遍的可预防性失明原因。
-
扩大服务覆盖范围: 我们正在率先探索利用日常智能手机和可穿戴设备进行早期检测,涵盖心血管疾病、胰岛素抵抗、高血压、脉搏缺失以及被动心率监测。我们正与阿肯色州的相关负责人合作,协助制定改善农村地区健康结果的蓝图。
[注:本文篇幅超长,以上为核心前篇精译,后续内容保留原文呈现]
- Tools for scientists and health practitioners: Collaborative AI tools like Co-Scientist are helping researchers accelerate and expand core steps of the scientific method, like generating and validating novel hypotheses (such as identifying new therapeutic applications for existing drugs for acute myeloid leukemia). We have also open-sourced AI tools like DeepConsensus, DeepVariant, and DeepPolisher. Over the last decade, these tools have assisted in understanding the human genome, drafting the first pangenome, and enabling ongoing work as part of the Human Pangenome Reference Consortium, better representing human genetic diversity and allowing experts to more accurately diagnose and treat diseases. Through AMIE (Articulate Medical Intelligence Explorer), we are advancing toward longitudinal disease management, building prospective evidence in real-world settings, collaborating with academic and medical institutions (e.g., Beth Israel Deaconess Center), and conducting a first-of-its-kind nationwide trial in real-world care settings. AMIE assists those providing care on the frontlines — freeing up doctors to spend more time with their patients.

To protect people’s safety and livelihoods, we need accurate predictions of natural disasters and weather. To make accurate predictions, we must understand the physical earth. Here’s how our technical and scientific progress is already making a difference in crisis prediction and detection:
-
Extreme weather and earthquakes: Last year, authorities in Jamaica used WeatherNext to accurately predict Hurricane Melissa’s path, securing early disaster funding and enabling life-saving emergency preparations long before the storm made landfall. Just a few months ago, our Earthquake Alert system alerted millions of people in Venezuela ahead of an earthquake. And we’re making progress in other areas, including cyclones and extreme heat.
-
Monsoons and floods: The scale of this work is encouraging. In 2025, our monsoon predictions provided information for 38 million farmers in India. And in just a few years, our forecasts on Flood Hub — now including both riverine floods and flash floods — have grown to cover 2 billion people across more than 150 countries in areas at risk for significant flood events.
-
Wildfires: Our AI tools have helped predict wildfire boundaries in the U.S. and 33 other countries. We’re also working with partners toward a FireSat constellation of satellites to detect wildfires previously too small to detect anywhere on earth. In 2025, we generated more than 520 crisis alerts on Google Search that provided timely wildfire information to over 75 million users around the world.
-
Next-gen prediction: With WeatherNext 3, we’re combining real-time satellite observations with AI, delivering high-resolution, hourly forecasts without immense supercomputing power — a major step forward for data-sparse regions that have been long underserved by high-resolution forecasting.
-
Planetary crisis mapping: We recently introduced the Earth AI Planetary Prediction Engine (PPE), an autonomous AI system that uses simple language instructions to almost instantly predict global crises like disease outbreaks, food shortages, and climate risks, helping humanitarian responders act faster during emergencies. In the ongoing Ebola outbreak in the Democratic Republic of the Congo, PPE successfully pinpointed 83% of emerging hotspots ahead of time, beating current forecasting systems. In Nigeria, PPE doubled food security forecasting accuracy at the local district level. And in the U.S., PPE outperformed traditional models in identifying vulnerable communities across 21 CDC health indicators. We’re continuing to expand prediction horizons to make these tools even more useful.

Learning is the bedrock of economic and societal progress. The world has made huge strides in education: 90% of primary school-aged children globally are enrolled in school, and approximately 87% complete their primary education. But 2022 research showed uneven access to education, differences in education quality, and limited content availability across languages, among other issues. Global learning outcomes have declined over the last two decades across 81 countries.
AI offers a profound opportunity to expand access to knowledge and support learners and educators alike:
- For learners: AI can enable personalized learning, remove language barriers, and make complex subjects more accessible through tools like Guided Learning. It can also expand access and support for learners and teachers, for example with free access to student tools in multiple countries.
- For educators: AI can help with research, preparation, and administrative tasks, freeing up time to spend with students.
Tackling challenges and focusing on responsible design: AI presents real challenges for learning, including questions on accuracy, safety, critical thinking, cheating, and learning loss. That’s why we have partnered with ISTE+ASCD to provide AI literacy training to all 6 million educators in the U.S., while conducting ongoing and rigorous research to ensure we design our educational tools with teachers in the lead, and AI as the assistant. We still have much to work through together.

Across the economy, AI can drive innovation and productivity for workers, businesses, and organizations across sectors. Its true value isn’t in what it does on its own, but in how people choose to use it and what they build with it:
- For small businesses: AI can help deliver new products, services, and economic growth. Small businesses are scaling custom marketing campaigns with Pomelli and bringing together complex regional sales data from different product lines with the Gemini Enterprise App.
- The future of work: The economic benefits of AI are neither automatic nor guaranteed. Like other new technologies, we expect AI to reshape work in three major ways: Some jobs are likely to decline, some jobs will grow, and many more jobs will change as AI increasingly assists workers. To understand these shifts, we’ve launched the AI & Economy ATLAS, which provides empirical insights on how people are using AI tools at work and in their daily lives, and translated the data into an interactive, open-access experience.
- Skilling: Training and skill building must be accessible to all regardless of a person's geography, language, or access to capital. To date, Google has provided more than $1 billion globally in training and skilling initiatives, helping over 100 million people gain essential digital and AI skills, from electricians and other skilled tradespeople, to rural healthcare workers, to manufacturing workers. More tailored worker assistance and transition support will likely be needed.

Grounded in our company’s founding mission, we are making important strides in breaking down languages and accessibility barriers with products, research, and models that understand the world’s languages exactly as they are expressed:
- Language tools: AI and machine learning have been part of Google Translate since its launch in 2006. Since then, advances in AI have helped us expand our technologies and products to support nearly 300 languages, spoken by 7 billion people — representing 86% of the world’s population, and helping more than 1 billion people translate around 1 trillion words each month. Our goal is to support the world’s 1,000 most-spoken languages.
- Inclusive speech technology: Access is broader than translation alone. For example, automatic speech recognition uses AI to understand spoken words and convert them into text. To help build inclusive speech technology, we partnered with Makerere University, University of Ghana, and Digital Umuganda to create an open dataset for 27 African languages spoken by over 100 million people**.** One language alone may have 100 dialects, each with its own nuance, which is why we’ve also partnered with IISC-Banglore and ARTPARK on Project Vaani to open-source speech and image datasets for 109 Indic languages.
- Accessibility: Beyond spoken languages, we've been making our tools more accessible for people with disabilities, including with our recent sign-language-to-text (SL2T) model for sign-to-text dictation.

Bold and responsible, together
We believe that such bold ambitions for the societal benefits of AI must always be accompanied by an equal commitment to responsibility. This is the basis of our bold and responsible approach to AI. Beyond our ongoing work to address the immediate and future risks of AI (detailed in our annual updates), the application of AI and science presents distinct challenges: On the one hand, we need the right detection, reporting, and mitigation tools in place to address risks around chemical, biological, radiological, and nuclear (CBRN) mis-use. At the same time, it is important to avoid the risk of missed use — that is, not applying AI in places and communities where it would provide benefit and, in some cases, worsening existing societal challenges and gaps.
We believe advancing scientific discovery and bringing those benefits to communities requires that we collaborate deeply with many other researchers and scientists working to tackle similar issues, including academia, national laboratories, philanthropic foundations, non-governmental organizations, governments, other companies, and local partners. The examples of advances cited above have benefited from such research collaborations and our own Google.org grant-making to accelerate scientific discovery and to address society’s greatest challenges. That’s why as part of our longstanding commitment to supporting science and discovery in academia, Google has given more than $1 billion across 1,700+ research institutions worldwide since 2006.
As Nobel laureate Joel Mokyr noted, new technology drives scientific advances, which in turn drive new technologies, creating "a positive feedback loop that can keep growing with no end in sight.” And it does so in ways that have improved people’s wellbeing and propelled societal progress. AI is the most powerful engine for that loop we have ever built — and we’re just getting started.