AI 时代的韧性

OpenAI 基金会旨在确保通用人工智能惠及全人类。

我们一直在不懈努力，跟上 AI 自身快速发展的步伐。今年 4 月，基金会宣布⁠在“生命科学与疾病防治”领域提供首批 1 亿美元资助，旨在借助先进 AI 帮助预防和治疗阿尔茨海默病等疾病。上周，我们宣布⁠了“就业与经济未来”计划，希望了解并塑造工作和经济繁荣对未来世代的意义。

今天，我们将进一步阐述下一项重大计划的愿景：随着 AI 能力不断增强，社会驾驭 AI 的能力也必须同样快速提升。我们将这项工作称为“AI 韧性”：这是一种在生态系统层面降低 AI 风险的方法，旨在让社会最大程度受益。

这项工作已经启动。自启动以来短短几个月内，基金会正通过 AI 韧性计划，敲定向各类组织提供的总额超过 1.3 亿美元的资助。相关信息将很快公开，后续还会有更多进展。^1⁠

变革性技术的演进规律

回顾那些深刻塑造人类历史的过往技术，是理解 AI 韧性重要性的最佳方式。

每隔一段时间，就会有一项技术出现，从根本上重塑社会。经济学家称之为“通用技术”。火。印刷机。电力。互联网。这些技术都经历了相似的发展轨迹：飞速创新、切实风险，以及相关机构的奋力追赶。但每个例子也表明，要让一项强大的技术变得安全，需要付出怎样的努力。

火让人类文明成为可能。它为我们取暖、烹煮食物，也保护我们免受猛兽侵害。它也曾将城市付之一炬。随着时间推移，社会逐步建立起韧性：耐火材料、消防栓网络、专业消防队和建筑规范。一个生态系统，就这样一层层建立起来。

电力也走过类似路径。1882 年，爱迪生的珍珠街电站点亮曼哈顿后，电力也引发了火灾、触电事故和公众恐慌。正是由于当时缺乏绝缘电线、断路器和规范等防护措施，全美各地城市中都曾发生过工人和路人触电伤亡的事故。各城市开始讨论是否应彻底放弃这项实验。但我们没有放弃。随着技术进步，我们建立了 Underwriters Laboratories（美国保险商实验室）等独立测试机构，制定了《国家电气规范》（National Electrical Code）等行业标准，并通过公共投资，把电力带到那些市场未能覆盖的社区。每一层建设都让电力更安全、也更容易获得；如今，电力已安全到孩子轻轻按下开关，灯光就会亮起。

这就是韧性建设做得好的样子。

AI 需要韧性生态系统

AI 正在沿着以往技术的轨迹发展，但速度前所未有。

我们仍处于 AI 发展的早期阶段，但益处已经清晰可见：它正在降低创业门槛、扩大教育可及性、加速科学发现，并改变医学的发展与实践。

与此同时，风险也以同样速度显现，而且几乎与 AI 的益处互为镜像。催生新行业的增长动力，同样可能颠覆现有行业，打乱人们的职业路径。那些能够帮助年轻人学习和创作的系统，也可能带来不良行为风险。加速生物学研究的工具，也可能降低制造有害病原体的门槛。AI 编写代码的能力一旦被恶意使用，也可能威胁关键基础设施。

早期的 OpenAI 团队认为，确保 AI 造福社会主要取决于解决技术对齐问题。这仍然至关重要，也是我们工作的核心，但我们现在认为，它只是整个拼图的一部分。随着 AI 在各行业和各个国家/地区不断普及，社会还需要独立研究、公共基础设施、行业协调，以及全新的专业领域。简而言之，我们需要建立 AI 韧性。

我们选择将初期工作聚焦在 4 个领域^2⁠，它们既关乎重大的近期风险，也有望产生直接影响：

1. 生物韧性，帮助预防未来可能出现的工程化大流行病；

2. 网络韧性，致力于保障全球关键系统安全；

3. AI 模型安全，加强人类对所创建模型的控制；以及

4. AI 对年轻人的影响，帮助让技术成为造福未来世代的积极力量。

我们的工作才刚刚开始。我们计划分享更多关于各个领域的策略和首批资助的信息，并逐步拓展至其他领域。

Bio-resilience

AI will enable biological research to move at unprecedented speed, helping develop new cures and public health improvements that enable us to all live healthier and longer. However, these same capabilities could also be misused by malicious actors, lowering the barrier to designing harmful pathogens.

The age of AI requires a renewed focus on biosecurity. Because advanced AI systems could be misused by bad actors to help create a wide range of biological threats, we will prioritize pathogen-agnostic biosecurity solutions. This will require investments across prevention, detection, and defense. We need to make it harder for malicious actors to access the expertise, equipment, and materials to create biological threats, improve our ability to identify and track novel outbreaks early, and strengthen the technologies—such as protective equipment, indoor air cleaning systems, and medical countermeasures—needed to respond quickly and effectively.

Cyber-resilience

AI has begun to rapidly reshape the cybersecurity landscape. The work that once required specialized teams can now be assisted or automated by capable models. At the same time, rapidly-improving AI capabilities can also be used to accelerate cyberdefenders, including by identifying and patching vulnerabilities and accelerating response.

Many large companies and private actors can spend heavily on cyber to secure their own systems, including with new advances in AI. We anticipate focusing significant resources on securing other important societal actors that are less resourced and will have a much harder time deploying AI-ready cyberdefenses as quickly as needed. In parallel, we are also focused on preparing for novel security challenges that artificial general intelligence will ultimately bring.

AI model safety

AI model safety focuses on the behavior of the systems themselves—whether they are truthful, reliable, and aligned with human intent. In a world where this goes awry, models can break out and behave in unpredictable ways, deceiving us or pursuing goals beyond their design. Getting this right becomes increasingly important as AI systems grow more autonomous and approach—and eventually surpass—human-level intelligence.

AI companies are investing substantial resources in model safety. However, the importance of this challenge calls for a broader, more robust ecosystem: independent institutions to evaluate model safety, public infrastructure to verify models’ safe deployment in practice, and continued advances in alignment science that advance the field broadly.

AI’s impact on young people

Young people are often the earliest adopters of new technologies, using them to learn, create, communicate, and explore the world. AI is no exception. But as these tools become an increasing part of young people’s daily lives, it is critical that we develop a stronger evidence base to understand its impacts.

Families, schools, policymakers, and community organizations are all grappling with questions about how and when young people engage with AI—including its impact on human connection, learning, and development. Our initial focus will be on advancing independent research to help guide those decisions—to better understand where AI can support development, the risks it may introduce, and the contexts that shape those effects.

These insights should drive broad safety standards and design principles that guide how any AI product is developed, how schools choose to deploy them, and if and how families decide to incorporate these technologies into their lives.

The work ahead

There is one critical difference between AI and the technologies that came before it: speed.

Fire resilience took millennia. Electricity resilience took decades. AI resilience is evolving in a matter of years. The systems that make it safe, reliable, and broadly beneficial must be built alongside it.

If we get it right, AI can become part of the foundational infrastructure of modern life—expanding access to knowledge, accelerating discovery, and improving lives at a global scale.

But that outcome isn’t guaranteed. No general purpose technology ever made itself safe.

Resilience is a permanent discipline that requires many people and institutions to build, invest, and collaborate. That is the work ahead, and it is one of the defining challenges of our time. We hope you’ll join us.