One day we may look back at this decade for its technological breakthroughs in generative artificial intelligence (AI). So much unchecked progress with tools like OpenAI's ChatGPT has understandably created anxiety among people from all walks of life about whether our societies are prepared for the widespread use of AI. The feared risks include the generation of misinformation and deepfake images, threats to people's jobs as formerly complex tasks become automated, and even a dystopian “replacement” of humans by machines. However, AI is still in the infancy of its own “Cambrian explosion”, a period of unparalleled emergence and rapid innovation. At this point in time, we can only imagine where its future evolution will lead us.

Ecologists have been using AI in research for decades (machine-learning is a more boring name for it), and today it is not uncommon for ecology graduate students to run their statistics using iterative, problem-solving AI algorithms. At its core, AI-based prediction is simply an automated version of the scientific method, designed to be an iterative learning process that becomes more refined with each iteration based on feedback and experience. In machine learning, selection for an optimized solution occurs with every iteration, somewhat similar to how natural selection operates on each generation of a species. With each iteration, the model attempts to minimize differences between its output and what it was trained to believe should be the “correct” output. Ultimately, humans control the inputs and impose artificial selection pressures (such as model parameters, thresholds, and goals for the training) that drive evolution of the outputs in a desired direction. A relevant question is whether humans can sensibly guide this evolution in a manner that parallels evolution and adaptation by natural selection.

Those worried about AI fear that we humans will end up on the wrong side of this selection process, in a zero-sum game between biology and technology. But the reality is that selection is driving biology and computing more closely together, toward an obligate symbiosis rather than a divergence. One could argue that this coevolution has already commenced and that we are already part human, part machine. For example, many of us have instant and unrestricted access to the vast knowledge of the internet via smartphones in the palms of our hands. It is relatively easy to imagine that humans will become more integrated with and dependent on AI in the future, because AI can help humans optimize solutions for complex problems (whether for morally good or bad reasons). If a hypothetical tipping point is crossed in which AI surpasses human intelligence and gains some degree of autonomy and sentience, it is unlikely that AI will annihilate humans, because that would be akin to attacking itself.

Instead, the more likely risk is that humans are becoming, and will continue to become, something new. Who better to understand the limits than ecologists, with their understanding of the fundamental principles of adaptation and evolution? In The Origin of Species, Darwin described natural selection as a process analogous to selective breeding in domesticated pigeons and horses, and this analogy can be further generalized to our coevolution with AI. If humanity becomes entangled within a mutualistic association with AI, its outputs and capabilities will be refined and its early forms will eventually either become extinct or morph into better adapted versions. This evolution is likely to be slow, though punctuated by moments of rapid and drastic change.

Are there risks? Of course, but they are more likely of the variety that we currently face. Just as any successful technological innovation increasingly becomes a part of daily life, there will be initial winners and losers. Even the best adapted species of AI code and bioengineered networks will still be susceptible to disease and disorders, malfunctions, and inefficiencies. However, over time, selection will drive the evolution of better adapted forms of AI. We simply argue that this process will more closely resemble coevolution, rather than an existential battle, between humans and machines.

Ecologists should find comfort in knowing that we will not soon become subject to our AI-enhanced machine overlords. We should find familiarity in the idea that human–machine coevolution will likely be guided by the same principles and processes that govern the natural systems that we study. An improved understanding of how AI can help us better use, conserve, repair, and build the natural world is where we are heading. Ecologists are well-positioned and uniquely qualified to lead in this endeavor.

有一天，我们可能会回顾这十年，看看它在生成人工智能（AI）方面的技术突破。 OpenAI 的 ChatGPT 等工具取得如此多的不受控制的进展，自然会引起各行各业的人们对我们的社会是否为人工智能的广泛使用做好准备的焦虑。令人担忧的风险包括错误信息和深度伪造图像的产生、随着以前复杂的任务变得自动化而对人们工​​作的威胁，甚至是机器对人类的反乌托邦“取代”。然而，人工智能仍处于“寒武纪大爆发”的起步阶段，这是一个空前崛起和快速创新的时期。此时此刻，我们只能想象它未来的演变将把我们引向何方。

几十年来，生态学家一直在研究中使用人工智能（机器学习是一个更无聊的名字），如今，生态学研究生使用迭代的、解决问题的人工智能算法来运行统计数据并不罕见。从本质上讲，基于人工智能的预测只是科学方法的自动化版本，旨在成为一个迭代学习过程，根据反馈和经验，每次迭代都会变得更加完善。在机器学习中，每次迭代都会选择优化解决方案，这有点类似于自然选择在物种的每一代中的运作方式。在每次迭代中，模型都会尝试最小化其输出与训练后认为应该是“正确”输出之间的差异。最终，人类控制输入并施加人为选择压力（例如模型参数、阈值和训练目标），从而驱动输出朝所需方向演化。一个相关的问题是，人类是否能够以与自然选择的进化和适应相平行的方式明智地引导这种进化。

那些担心人工智能的人担心我们人类最终会在这个选择过程中站在错误的一边，陷入生物学和技术之间的零和游戏。但现实是，选择正在推动生物学和计算更加紧密地结合在一起，走向必然的共生而不是分歧。有人可能会说，这种共同进化已经开始，我们已经一半是人类，一半是机器。例如，我们中的许多人可以通过手掌上的智能手机即时且不受限制地访问互联网上的大量知识。相对容易想象，未来人类将与人工智能更加融合和依赖，因为人工智能可以帮助人类优化复杂问题的解决方案（无论是出于道德上的好还是坏的原因）。如果跨越一个假设的临界点，即人工智能超越人类智力并获得一定程度的自主性和感知力，那么人工智能不太可能消灭人类，因为这类似于攻击自身。

相反，更可能的风险是人类正在并将继续成为新事物。谁能比生态学家更了解这些限制，因为他们了解适应和进化的基本原理？在《物种起源》中，达尔文将自然选择描述为类似于家养鸽子和马的选择性育种的过程，这种类比可以进一步推广到我们与人工智能的共同进化。如果人类陷入与人工智能的互惠关系中，其产出和能力将得到完善，其早期形式最终要么灭绝，要么演变成更适应的版本。这种演变可能是缓慢的，尽管不时会出现快速而剧烈的变化。

有风险吗？当然，但它们更有可能是我们目前面临的那种。正如任何成功的技术创新日益成为日常生活的一部分一样，也会有最初的赢家和输家。即使是最适应的人工智能代码和生物工程网络物种仍然容易受到疾病和紊乱、故障和低效率的影响。然而，随着时间的推移，选择将推动更适应的人工智能形式的发展。我们只是认为，这个过程将更类似于人类和机器之间的共同进化，而不是一场生存之战。

生态学家应该感到安慰的是，我们不会很快受到人工智能增强机器霸主的控制。我们应该熟悉这样一个想法：人机协同进化很可能遵循我们所研究的自然系统的相同原则和过程。加深对人工智能如何帮助我们更好地使用、保护、修复和建设自然世界的理解是我们前进的方向。生态学家处于有利地位并且具有独特的资格来领导这一努力。