菲尔兹奖得主、著名数学家陶哲轩（Terence Tao）近日对人类日益依赖人工智能发出警告。他担心，随着人们逐渐将思考和解决问题的过程交给AI，人类正在失去对问题本身及其解决方案的理解，而这一趋势可能很快在现实世界中产生严重后果。
这番言论来自陶哲轩10月16日在加州大学洛杉矶分校举办的Science x AI Summit上的演讲。科学与人工智能研究基金会（SAIR）近日在YouTube发布了题为《Math 2.0》的完整演讲视频，相关片段随后在海外社交平台X上引发关注。
“我们正在把自己的思考委托给AI，由它们替我们解决问题。但与此同时，我们也正在失去一种保障——无法确信AI给出的解决方案，是否真的是我们想要的。”陶哲轩表示，“我真的非常担心，我们很快就会看到这件事在现实世界中产生的后果。这实际上是我在数学领域之外的一项担忧。”
陶哲轩指出，人工智能在数学领域取得的进展，正在改变人类解决问题的传统方式。过去，数学家为攻克复杂问题，需要长期研究、理解相关理论，并在这一过程中积累新的知识。但如今，AI能够快速给出答案，人类却未必能够理解答案背后的推理过程。
他认为，这种变化在纯数学领域的风险相对可控，因为数学证明通常可以通过形式化工具进行验证。但当同样的方法被用于现实世界，情况就会复杂得多。
陶哲轩特别提到了电网、航空交通管制、药物研发和芯片设计等领域。他担心，未来人们可能直接将复杂问题交给AI处理，却缺乏对底层模型和现实约束条件的充分理解。
在这种情况下，AI给出的方案可能在数学模型中完全成立，却并不符合现实世界的真正需求。 AI甚至可能找到模型中的漏洞或边界条件，在满足表面指标的同时偏离人类原本希望实现的目标。
陶哲轩强调，问题并不在于AI技术本身，而在于人类如何使用AI。他主张推动数学研究进入所谓的“Math 2.0”时代，让AI不仅帮助人类解决问题，更能增强人类对问题的理解，帮助研究人员识别模型的局限性和潜在风险。
10月22日，陶哲轩还在《自然》杂志发表文章，呼吁数学界重新审视AI时代的研究目标。他指出，随着AI生成数学证明的速度超过人类理解和消化这些成果的能力，数学研究正在面临新的挑战。
陶哲轩认为，AI工具完全有可能增强人类的理解能力，但这需要研究方式和激励机制发生改变。AI时代真正需要保留的，不只是获得正确答案的能力，更是人类理解、检验和判断答案的能力。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The Fields Prize winner and famous mathematician Terence Tao has recently warned against the increasing reliance on artificial intelligence. He feared that, as the process of thinking and solving problems was gradually handed over to AI, humanity was losing its understanding of the problem itself and its solutions, a trend that could soon have serious consequences in the real world.
The speech was given at Science x AI Summit at UCLA on October 16. The Science and Artificial Intelligence Research Foundation (SAIR) recently presented a full presentation video entitled " Math 2.0 " on YouTube publish, which was followed by attention on social platform X overseas.
“We are entrusting our thinking to AI, which will solve our problems. But at the same time, we are also losing a guarantee — we cannot be sure that the solution that AI offers is really what we want.” Ta Zhrau said, “I am really worried that we will soon see the consequences of this in the real world. It's actually a concern outside of math."
Tao Chul-hyun noted that the progress made by artificial intelligence in the field of mathematics is changing the traditional ways in which humans solve problems. In the past, mathematicians had struggled with complex problems, requiring long-term research, understanding of relevant theories and building new knowledge in the process. But now that AI is able to give the answer quickly, human beings do not necessarily understand the reasoning behind the answer.
In his view, the risk of such a change was relatively manageable in the field of pure mathematics, since a mathematical certificate could usually be validated through formalization tools. But when the same approach is applied to the real world, the situation is much more complex.
Special mention was made of areas such as electricity grids, air traffic control, drug development and chip design. He feared that, in the future, people might refer complex issues directly to AI but lacked a full understanding of the bottom model and realistic constraints.
In this case, the programmes offered by AI may be fully established in a mathematical model but do not meet the real needs of the real world. AI may even find loopholes or boundary conditions in the model that deviate from what humans would otherwise wish to achieve while meeting surface indicators.
Tao Chul-hyun stressed that the problem is not about AI technology per se, but about how humans use AI. He advocated moving mathematical research into the so-called “Math 2.0” era, so that AI would not only help humans to solve problems, but also enhance human understanding of problems and help researchers to identify the limitations and potential risks of models.
On 22 October, Tao Chul-hyun also published an article in Nature magazine calling for a fresh look at the research objectives of the AI era in mathematics. He noted that mathematics research was facing new challenges as AI generated mathematical evidence faster than humans could understand and digest those results.
According to Tao Chul-hyun, the AI tool has the full potential to enhance human understanding, but this requires a change in the way research is conducted and in the incentive mechanisms. The AI era really needs to be preserved, not just the ability to get the right answers, but the ability of humankind to understand, test and judge them.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.