中国人工智能发展对城市能源效率的影响

1^。2015年，联合国启动了17个可持续发展目标（SDG），敦促各国将经济增长与环境责任保持一致，并在2030年之前取得实质性进步，以确保一个更安全，更公平的未来2。^{在这些目标中，能源效率在实现可持续发展方面起着关键作用}3，，，，^{4。面对地缘政治紧张局势的上升}5^{加强气候变化}6^{不确定性日益严重，能源的有效利用对于国家发展至关重要。}广泛的研究已经确定了能源效率的多个决定因素，包括金融发展

7，，，，^{8碳排放交易9}，，，，^{10，，，，11财政权力下放12}制造集聚¹³，，，，^{14和政策不确定性15}。^{这些因素以不同的理论框架为基础，通过不同的机制影响能源效率。}最近，数字技术引起了学术对能源系统的变革性影响的越来越多的关注。技术进步正在重塑能源生产，运营和传输过程。在更广泛的可持续发展背景下，数字创新正在加速向更绿色，低碳实践的过渡¹⁶。这些技术已被证明可以提高能源效率，现代化能源基础设施和重组能源消耗模式^{16，，，，17，，，，18}。利用技术创新理论，几项研究强调了数字化转型的有益效果，强调了其在升级工业结构，优化能源使用和提高整体效率的作用¹⁹。但是，一些学者警告说，数字基础设施的发展也可能带来不利的环境后果，可能破坏可持续性目标⁶。在新兴技术中，人工智能（AI）是未来创新的变革驱动力。在2022年，每年授予的AI相关专利的数量超过了62,000次，超过了2018年记录的数字的七倍以上。到2023年底，AI系统在诸如图像分类，基本阅读理解，自然语言推理，多语言理解，多语言理解和视觉理解和视觉理解等领域的标准化评估中已经等于或超过了人类绩效。只有在视觉常识性推理和复杂的数学问题解决方面，AI仍然落后于人类能力（人工智能指数报告，2024年）。

越来越多地将AI整合到能源部门中，以加速向可持续能源的过渡并提高能源效率。通过先进的数据分析，预测性建模和自动化，AI促进了更明智的决策，优化能源分配并促进了可再生能源技术的采用，从而有助于更具弹性和可持续的能源系统^{20，，，，21，，，，22}。支持者认为，人工智能正在重塑能源管理和可持续性的景观，推动高质量的城市发展并产生可衡量的环境利益，包括提高能源效率和增强的生态成果^{23，，，，24}。但是，尽管有这些优势，AI技术的能源密集型性质仍会为总体能源效率带来潜在的风险。培训和操作AI模型需要大量的计算资源²⁵这会导致二氧化碳排放量增加和环境降解^{19，，，，26}。此外，AI系统对大量数据集的依赖性间接扩大了信息技术领域的碳足迹。例如，仅美国的数据中心约占国家能源消耗的2％。在全球范围内，预测表明，到2030年，信息和通信技术（ICT）行业的能源消耗可以达到总能源使用的20％，强调其在全球能源景观中的作用不断增长^{17，，，，27}。

鉴于AI的快速发展及其对转化和挑战可持续能源使用的双重潜力，因此对其影响的实证研究至关重要。这项研究以中国为中心，为人工智能发展与能源效率之间的关系提供了宝贵的见解。根据工业和信息技术部（中国）的数据，到2023年，该国的AI行业超过了5000亿元人民币，包括超过4,500家企业。作为世界上最大的能源消费者和领先的发展中国家^{28，，，，29}中国对能源效率的追求对全球可持续性产生了重大影响，尤其是对其他发展中经济体。中国AI的快速增长为评估其影响并探索将AI融入可持续发展策略的途径提供了独特的机会。

先前的研究主要取决于指标，例如工业机器人的数量或与AI相关的专利申请，以评估AI开发^{30，，，，31，，，，32}通常专注于生产过程和技术创新。尽管有益，但这些指标倾向于强调特定的技术及其对能源消耗的直接影响，从而有可能忽略AI在整个部门之间的广泛和更系统的影响。为了解决这一差距，本研究使用来自中国城市的县级数据采用固定效应模型来评估AI发展与能源效率之间的关系。

这项研究以四种关键方式推进了文献：（1）与使用诸如工业机器人或专利计数之类的代理的先前研究不同，我们评估了AI企业的影响，对AI实施的实际和商业维度提供了更强有力的反映。（2）除了能量与GDP比率之类的简单指标外，我们采用了基于CharnesCooperâRhodes（CCR）方法的数据包络分析（DEA）模型，从而使跨多个输入和产出的能源效率进行细微的评估。（3）我们研究了AI影响能源效率的途径，从而深入了解驱动这些影响的中介机制。（4）通过研究AI在影响能源效率方面的功能边界，我们对其潜在范围提供了更清晰的了解。通过解决这些研究差距，该研究对AI对能源效率的直接影响提供了关键的见解，尤其是通过解开将AI开发与可持续能源结果联系起来的机制。这些发现为政策制定者和行业利益相关者提供了可行的策略，从而在AI技术的一体化中促进了基于证据的决策，以支持全球可持续性目标。

本文的结构如下：部分： 2对文献进行了全面的综述，并介绍了理论假设。部分 3描述数据源和研究方法。部分 4报告经验分析的结果。部分 5提供扩展的分析，重点是调节效应。最后，部分 6总结了关键结论，并讨论了研究的政策含义和局限性。

能源效率的主要驱动力

通过加强环境治理，有效且可再生能源系统已被广泛认为是可持续发展的重要因素

33^。许多研究研究了影响能源效率的因素。其中，环境考虑在制定财务策略方面发挥了关键作用，尤其是通过绿色金融政策，该政策将资本分配给环境可持续的倡议，并为改善能源绩效奠定了基础^{7，，，，8}。在此基金会的基础上，诸如碳交易系统等机构机制提供了基于市场的激励措施，以减少排放并促进更有效的资源分配^{9，，，，10，，，，11}。此外，工业聚集的作用引起了人们的关注。研究人员发现，聚类行业可以创造规模经济，并产生积极的外部性，例如知识溢出，从而简化生产过程并增强能源利用率^{13，，，，14}。

数字技术的影响

最近，由于数字技术的快速发展，全球向数字化转型的转变已重塑了经济范例。人工智能和其他边界技术的突破已成为各个部门能源使用的关键推动者¹⁹。这些进步不仅促进了经济和社会转型，还促进了进一步的技术进步。结果，日益增长的学术兴趣现在致力于了解数字化如何影响生产系统并改变能源消耗模式。

在学术论述中，关于数字技术对能源过渡的影响，已经出现了两种不同的观点。积极观点的支持者认为，数字技术在带来可衡量的环境福利的同时促进了高质量的城市发展³⁴。增强的创新能力³⁵通过优化人力资本，将先进的制造与现代服务行业整合并实施降低成本策略来提高总要素生产率^{36，，，，37}，最终导致能源效率的提高^{23，，，，24}。但是，对比的观点警告说，由于资源密集型数字基础设施的构建，技术创新也可能会驱动碳排放，从而对生态环境产生不利影响^{19，，，，26}。

AI的影响

人工智能（AI）是数字技术的基础元素，有可能深刻地改变能源利用并提高效率，尤其是在制造业中^{21，，，，22，，，，38}。与其他数字创新不同，AI的核心功能可以分为六个关键领域：学习，感知，预测，互动，适应和推理³⁹。例如，AI的学习能力使系统可以随着时间的推移提高性能，从而提高运营效率和有效性。感知能够解释复杂的数据集和动态环境，支持更明智的决策⁴⁰。预测能力有助于准确的结果预测，这对于战略规划至关重要⁴¹。互动促进了人与机器之间的无缝沟通，增强用户参与度和系统响应能力⁴²。适应使AI系统可以响应新的条件和任务，并保持其相关性和功能^{40，，，，43}。最后，推理使AI得出逻辑推断并解决了复杂的问题，从而有效地增强了人类认知能力。实证研究强调了AI对能源效率的影响。例如，AI驱动的通风控制系统在美国的商业建筑中可节省约26％⁴⁴。与传统的数字工具不同，AI自主做出决策，识别模式并展示了类似人类的科学推理的能力，可以使生产过程更加细微的管理，从而支持更绿色的制造实践⁴⁵。总的来说，这些功能将AI建立为在各种工业和技术领域的创新和效率的多功能和强大的推动者。然而，AI系统的开发和部署是能源密集型的，尤其是在制造，模型培训和操作阶段，可能导致大量的碳足迹²⁰。

越来越多的文献探讨了AI与能源效率之间的复杂关系，从而产生了一系列观点。几项研究突出了AI提高能源性能的潜力，特别是在以高水平绿色创新为特征的地区²⁵，在AI集成更高级的高性能组织中⁴⁶。但是，这些好处与AI模型开发和培训相关的巨大能源需求抵消了这些好处。用于算法开发，培训周期和数据中心冷却的高性能计算会消耗大量的计算资源，从而有助于碳排放量升高和环境退化^{19，，，，25，，，，26，，，，47}。数字基础设施的扩展，尤其是大规模的扩展，加剧了这些挑战，因为它需要大量的构建和维护能源投入^{17，，，，28，，，，48，，，，49}。

尽管有这些担忧，AI还是通过预测分析，实时监控和自动化控制系统来解决与能源相关的挑战的变革机会。随着AI技术在制造，运输和公用事业等领域越来越多地采用，它们提供了复杂的机制来检测能源消耗的效率低下，并实现了确切的数据驱动干预措施，以提高运营效率⁵⁰。这些应用产生直接和间接的环境利益。例如，AI可以优化及时输入生产因素，以最大化资源利用率⁵¹。此外，大规模工业生态系统中智能自动化的整合使生产线变得更加灵活，自适应，自我意识，自我调节，并且能够自主优化^{52，，，，53，，，，54}。这些进步大大降低了资源消耗并促进可持续性，从而在环境绩效和工业效率方面取得了可观的增长。

文学差距

该文献综述综合了当前的能源效率奖学金，并对其影响因素进行了全面分析。尽管大量研究研究了数字技术和AI对能源效率的影响，但仍然存在一些关键差距。值得注意的是，许多现有文献将AI与更广泛的数字技术混为一谈，有限的研究隔离了AI来检查其独特的影响。此外，先前的研究经常通过诸如工业机器人密度或专利数据等代理来量化AI。相反，本文强调了AI技术的实际应用^{55，，，，56，，，，57}。此外，普遍的研究通常使用一维框架来评估能效，例如能源输入与GDP输出的比率。这项研究通过采用CCR模型采用了一种更强大的方法，该模型可以对能源效率进行多维评估^{25，，，，56}。最重要的是，AI影响能源效率的基本机制仍未得到充实。这些研究差距强调了对进一步的实证研究的必要性，尤其是在中国，在中国，对AI在能源消耗中的作用的细微了解可以增强其在发展中经济体中的适用性。

理论背景和假设

如上所述，AI和能源效率之间的关系是多方面的，涵盖了机遇和挑战。当AI技术从专利注册过渡到现实世界部署时，它们的能量使用动态就会发展。实际实施通常会导致系统优化和减少能源消耗。这意味着AI的提高能源效率的能力是通过主动应用而不是仅凭理论建模来实现的。在这方面，AI通过将能源效率与适应性，互动性和创造力相结合，提供了独特的优势²²。这种现象与索洛悖论一致，这表明从技术创新中获得的生产力并不明显，但随着时间的推移逐渐出现⁵⁸。创新理论进一步支持这样的观点，即广泛采用AI可以导致生产力和运营效率的显着提高。鉴于AI显示出减少能源消耗和破坏常规能源使用模式的潜力，这项研究提出了以下假设：

H1：AI开发可以提高能源效率。

绿色技术创新是指旨在保护能源，减少排放，减轻气候相关的环境损害以及增强生态益处的企业驱动的创新。这些创新也有助于生产技术的现代化⁵⁹。经验证据表明，AI发展对绿色创新成果产生积极影响⁹。从技术创新的角度来看，绿色创新可以减少化石燃料的排放，同时促进采用可再生能源¹¹。尽管绿色创新可以增加研发支出，但它同时提高了生产率，并可以有效地管理制造过程中废水，排气排放和固体废物。实施严格的环境法规来刺激技术创新可以大大降低环境污染，而不会损害生产效率⁶⁰。因此，这项研究提出了假设2：

H2：AI开发通过促进绿色技术创新提高了能源效率。

先前的研究已经确定，能源强度在工业领域各不相同，结构转化会影响总能源效率⁶¹。值得注意的是，工业结构是碳排放的关键决定因素。行业的配置和组成，尤其是它们对能源密集型技术和过程的依赖，基本上影响了整体碳足迹。因此，工业重组已成为可持续发展和气候变化策略的越来越重要的⁶²。升级工业结构可以通过减少高能消费型部门的份额，同时扩大低碳和可再生能源行业的份额来减少总碳排放。这种结构性转变不仅减少了排放，而且还促进了更具弹性和可持续的经济体系，与全球努力打击气候变化并提高环境可持续性的努力保持一致⁶³。人工智能（AI）通过优化和重塑工业配置在推动工业升级方面起催化作用。预计AI的广泛部署将彻底改变传统技术，从而实现更理性，高效和创新驱动的工业结构。AI和相关的创新技术通过通过采用，整合和扩散新技术范式来促进工业聚类和提高运营效率，从而加快工业适应和进步^{64，，，，65，，，，66，，，，67}。技术进步还产生了重大的资本重新分配和创新驱动的转型。一方面，包括物理，人类和机构在内的资本倾向于从高能密集型部门流入知识密集型的低碳行业。另一方面，以卓越的能源效率和环境绩效为特征的新兴行业的进步有助于总体改善能源利用。基于此基本原理，本研究提出了以下假设：

H3：AI开发通过合理化工业结构来提高能源效率。

在探索AI影响能源效率的机制时，研究其影响可能会有所不同的上下文条件同样重要。利用环境调节理论，先前的经验研究表明，这种法规对两家公司对创新的意图及其实际的创新行为产生了明显的影响。这表明AI在提高能源效率方面的有效性可能取决于监管环境以及激励公司采用绿色技术的程度。环境法规提高了企业对生态关注的认识，并促进参与绿色创新活动。例如，严格的政策可以增强公司对环境责任重要性的认识，从而增加他们追求绿色创新的动力⁶⁸。在发展中国家，正式的监管框架通常是弱或不一致的非正式环境监管（IER），这是一种诱人的机制⁶⁹。我们认为，在这种情况下，AI对能源效率的影响在以较高水平的IER为特征的环境中可能更为实质。作为非正式监管的关键驱动力，社会压力和公众审查可以激励公司采用减少排放并改善运营可持续性的AI技术。这些监管动力可能会促进采用AI的更有益环境，从而扩大其提高能源效率的潜力。在此基础上，我们提出以下假设：

H4：非正式环境调节扩大了人工智能对能源效率的积极影响。

城市的发展阶段显着影响AI可以提高能源效率的程度。在中国，大约44％的城市地区被归类为基于资源的城市中心，其经济在很大程度上依赖于矿物，水和森林等自然资源的提取和处理⁷⁰。根据资源诅咒理论，这些资源的有限本质意味着持续的提取最终会导致其枯竭⁷¹。基于资源的城市通常遵循四个发展阶段：成长，成熟，下降和再生。每个阶段都提出了能源使用和技术采用的不同挑战和机会。不断增长的基于资源的城市拥有丰富的资源，并且处于工业扩张的早期阶段。这些城市越来越多地探索可持续实践以及资源开发。经过多年的密集剥削，成熟的城市已经达到了峰值资源的产出，面临着平衡经济增长与环境保护的挑战。实现这一平衡对于长期可持续性至关重要，需要刻意的努力来减轻环境降低，同时保持经济绩效⁷²。基于资源的城市下降面临资源耗尽和经济不稳定的复杂挑战。这些城市必须从依赖资源的经济过渡到多元化和可持续的经济。这种过渡需要创新的政策措施和战略干预措施。AI可以通过确定效率低下，优化剩余资源并促进向可持续实践的转变来在这种转变中发挥关键作用。在再生城市中，资源提取基本上已经停止，努力侧重于生态恢复和可持续的城市发展。对于这些城市，AI对于支持智能城市计划，能源管理系统和环境监测至关重要，所有这些都是实现可持续城市规划的关键⁷³。实证研究表明，数字经济对资源依赖性城市的能源效率有明显的积极影响。但是，在多元化经济体的城市中，这种影响不太重要。基于资源的城市的响应能力提高可能归因于其集中式工业结构，最初降低了能源效率基线⁷⁴。因此，这些城市既表现出更大的能力，又有更迫切的需求来利用AI和其他创新技术来优化能源使用。随着自然资源不可避免地会减少，基于资源的城市的优先级在不同阶段发展。在早期阶段，重点可能是提高提取效率，而在以后的阶段，重点转向可持续性和生态恢复。因此，对提高能源效率的AI应用的需求在各个阶段各不相同，在经历资源消耗和经济下降的城市中，需求变得最急剧。基于这种情况，我们提出以下假设：

H5：基于资源的城市中资源开发的阶段调节了人工智能对能源效率的影响。

数字1显示概念框架。

显示概念框架。

数据源

我们研究的重点是AI对中国城市的能源效率的影响。为了凭经验检验我们的假设，我们使用五个主要来源的数据采用固定效应模型。由于城市级别的能源消耗数据尚未公开，我们遵循既定方法，并使用NOAA的夜间光强度数据作为能源消耗的代理^{75，，，，76}。官方的能源数据（以大量的标准煤炭为标准化）是从中国能源统计年鉴中汲取的，以促进计算和比较。从中国城市统计年鉴中提取了其他控制变量，包括GDP，工业结构，收入水平，教育程度，人口密度和人口组成。使用将关键字提取技术应用于Tianyancha Enterprise数据库中列出的业务范围描述。此外，来自CNIPA的绿色专利数据提供了对技术创新水平的见解。所有连续变量均经过对数转换以稳定方差和标准化分布。在数据收集和预处理之后，最终数据集包含4177个观测值，这些观察结果涉及2006年至2020年。

模型设置

利用现有的研究，我们采用双向固定效应面板模型来控制个人（城市级别）和特定时间的效果²⁵。该模型通过有效减少省略的可变偏差并增强因果推断的精度来处理复杂面板数据。基线规范如下：

我在这里表示这一年。因变量EE代表能源效率，而主要的解释变量AI捕获了人工智能发展指数。控制变量包括人口密度，工业结构，人均GDP（PGDP），二氧化硫排放（SO2）和能源消耗。城市和年份的固定效果由\（\：{{\ upmu \：}} _ {\ text {i}}} \）和\（\：{{\ upgamma \：}} _ {\ text {t}}} \）分别。

变量的描述

Explained variable

Energy efficiency is a multifaceted concept with no universally accepted metric.Broadly, it refers to achieving the same level of output or service with reduced energy input—for example, using the ratio of energy consumption to GDP output as a proxy²⁵。However, such single-factor indicators provide a limited, one-dimensional view that neglects interactions among multiple inputs and the presence of multiple outputs^{77，，，，78}。

These indicators typically focus solely on the energy–output relationship, overlooking the influence of other production factors such as labor and capital.To address this limitation, we adopt the CCR model within the Data Envelopment Analysis (DEA) framework, which evaluates relative efficiency using a non-parametric approach.This model incorporates a broader set of inputs—namely energy, labor, and capital—and a single output (GDP) to estimate each decision-making unit’s (DMU’s) position relative to a production frontier⁷⁴。This methodology provides a holistic view of energy efficiency by comparing the performance of a DMU with peers that optimize input use for a given output or maximize output for a given level of input⁷⁹。By integrating both technical and scale efficiency under the assumption of constant returns to scale⁸⁰the CCR model captures input–output interactions more comprehensively, providing a more accurate and holistic assessment of energy efficiency.Therefore, we construct our measure of energy efficiency using the CCR model rather than relying on single-factor indicators.For robustness checks, we further construct a measure of total factor energy efficiency using the SBM (Slack-Based Measure) model^{81，，，，82，，，，83}。

Explanatory variable

Prior studies have commonly measured the level of AI development using two main approaches.The first relies on industrial robot data from the International Federation of Robotics (IFR)^{30，，，，31，，，，32，，，，84，，，，85}。While widely used, this method offers only a partial view and fails to capture the full spectrum of AI applications.The second approach uses AI-related patent data as a proxy for AI development.As a direct reflection of technological innovation, patent data enables precise identification of AI-specific technologies and has been increasingly employed to assess technological advancement^{86，，，，87}。Nevertheless, the number of AI-related patents is often used as a proxy for the level of R&D activity and the cumulative technological achievements in the AI domain.However, when examining energy consumption, the number of AI enterprises may serve as a more direct and relevant indicator.This is because enterprise count reflects the degree of industrial agglomeration and the extent to which the AI sector has matured in a given city.A growing number of AI enterprises typically signals the translation of R&D efforts into real-world applications and market-driven innovations^{25，，，，88}。Therefore, compared to patent counts, the number of AI enterprises offers a more comprehensive measure of both innovation output and its practical implementation.

Mediating variables

Green technology innovation (GTI).At the city level, green technology innovation is measured by taking the natural logarithm of the number of green patent applications plus one, to account for the potential of zero values and to normalize the data⁸⁹。

Industrial Structure Rationalization (ISR).Following prior research, this study adopts the Theil index to evaluate the rationalization of industrial structure by assessing both sectoral coordination and resource allocation efficiency⁶²。The Theil index quantifies disparities in output and employment across sectors, capturing the extent of structural imbalance at the city level, as shown in Eq. (2）：

In Eq. (2），TLdenotes the Theil index,with Y和lrepresenting total industrial output and labor force, respectively.是的和Lirefer to the output and employment of sector我， 在哪里我ranges from 1 to n, the total number of sectors.A lower Theil index approaching zero indicates a more rational industrial structure.This study calculates ISR for 284 cities over the period 2006–2020.

Moderating variables

Informal environment regulation (IER).To assess the moderating role of informal environmental regulation, the study employs a composite index constructed from indicators such as per capita income, population density, age structure, and educational attainment.These variables jointly capture the socio-economic conditions that influence public awareness and pressure for environmental protection, thus serving as proxies for informal regulatory mechanisms operating outside formal institutions.

To explore how informal environmental regulation moderates the impact, this research draws on the methodology of selecting a series of indicators such as income level, educational background, population density, and age structure to measure the extent of informal environmental regulation in cities^{90，，，，91}。

Stage of Resource City (SRC).Based on the National Resource-Based City Sustainable Development Plan (2013), resource-based cities are classified into four developmental stages: Growing, Mature, Declining, and Regenerating.In this study, these stages are numerically coded from 1 to 4, respectively, following the classification framework proposed by previous research⁷⁰。

Control variables

In addition, we control for several variables commonly identified in the literature as influencing energy efficiency: (1) Population Density (density): Areas with higher population density often exhibit greater energy demand²⁶。Population density is calculated based on the year-end total population^{57，，，，92}。(2) Industrial Structure (structure): This is measured by the share of tertiary industry value added in GDP, capturing the economic composition and its implications for energy use patterns.(3) GDP per Capita (pgdp): Reflecting regional economic development, GDP per capita is associated with higher living standards and greater awareness of sustainability concerns⁹³。It is measured as the per capita GDP of urban residents.(4) Total Energy Consumption (energy): This variable captures the absolute scale of energy usage within a city, serving as a key control for evaluating energy efficiency⁹⁴。(5) Environmental Pollution (soâ‚‚): Given that highly polluted regions often allocate more resources to environmental management, sulfur dioxide (soâ‚‚) emissions—one of the most prominent industrial pollutants—are used as a proxy for environmental pressure and are included as a control variable⁹⁴;therefore, this study controls it, considering the environmental pollution status, and measures it by the amount of sulfur dioxide emissions.

Table 1outlines the definitions and calculation methods of all variables.

Benchmark regression analysis

Due to the Hausman test results, with a p-value of 0.000, we choose the fixed effects model to control for unobserved individual heterogeneity.The analysis begins by examining the impact of AI development on energy efficiency (EE), with regression results presented in Table 2。Column (1) reports baseline estimates without control variables, while column (2) introduces controls.Column (3) further refines the model by clustering standard errors at both the city and calendar year levels to mitigate potential intra-group error correlation.The progressive increase in R-squared values across columns (1) through (3) indicates improved model fit with the inclusion of controls and robust error adjustments.In all specifications, the coefficient of AI on EE remains positive (0.049) and statistically significant at the 1% level, providing preliminary evidence that AI development positively influences energy efficiency.

Regarding the control variables, the coefficient on population density is positive and significant at the 1% level, suggesting that higher population density is associated with greater energy efficiency.This may be attributable to the intensified economic activity and more diversified industrial structures typically found in densely populated regions.Additionally, such regions often adopt stricter energy efficiency regulations, incentivizing firms to implement advanced energy-saving technologies.Likewise, the coefficient on per capita GDP is both positive and significant at the 1% level, indicating that higher income levels are linked to enhanced energy efficiency.This relationship likely reflects the increased capacity of wealthier regions to invest in technological innovation and R&D, along with heightened public awareness of environmental sustainability and energy conservation.

In contrast, the coefficients for soâ‚‚ emissions and total energy consumption are negative, implying that greater pollutant emissions and energy usage are detrimental to energy efficiency.These findings are consistent with empirical patterns observed in real-world contexts.

Robustness checks

Robustness test

To verify the robustness of these results, the study undertakes three validation approaches.First, it replaces the CCR model-derived measure of energy efficiency with a single-factor energy efficiency (EE) measure⁸¹。This shift allows us to assess energy efficiency through a more straightforward metric, calculated as the ratio of energy input to economic output¹⁰。Although this approach oversimplifies the multidimensional nature of energy performance, it aligns the analysis with conventional metrics frequently used in existing literature, enabling comparability across studies.As shown in column (1) of Table 3, the positive and significant relationship between AI development and EE persists, lending further credibility to the baseline findings.

Second, a subsample analysis is conducted to account for regional heterogeneity.The cities of Beijing, Shanghai, Guangzhou, and Shenzhen—China’s leading megacities—exhibit advanced economic development, robust infrastructure, and strong capabilities in AI innovation, placing them well ahead of cities in central and western China, which face structural and infrastructural limitations.To mitigate potential bias introduced by these outliers, a regression is performed after excluding the four first-tier cities.The results, presented in Column (2) of Table 3, are consistent with those of the full sample, further reinforcing the robustness and generalizability of the core conclusions.

Third, the effects of AI development on energy efficiency may exhibit a time lag due to delays in policy interpretation and implementation.To account for this, we introduce one- and two-period lags of AI development as robustness checks, as shown in columns (3) and (4) of Table 3， 分别。The results remain consistent with our baseline findings and provide further empirical support for Hypothesis 1, reaffirming that AI development is positively associated with improvements in urban energy efficiency.

Endogenous processing

To address potential endogeneity, we adopt an instrumental variable (IV) approach.Drawing on established methodologies in the literature, we construct two instruments: (1) a one-period lag of AI development, and (2) a Bartik shift-share instrument.The latter is derived by interacting the lagged first-order AI index with its first-order difference, thereby generating a theoretically grounded IV^{25，，，，95，，，，96}。Table 4presents the results of the IV regressions, which include the same set of control variables used in the baseline model to mitigate confounding effects.The results demonstrate that AI remains positively and significantly associated with energy efficiency under both IV strategies.Specifically, columns (1) and (3) report the first-stage regression outcomes, confirming a strong correlation between the instruments (IV1 and IV2) and AI.Columns (2) and (4) show the second-stage estimates, where the coefficients on AI are 0.051 and 0.049, respectively, both significant at the 1% level.These findings confirm the robustness and consistency of the main regression results.

Mechanism test

Green technology innovation

Table 5reports the mediating role of green technological innovation (GTI) in the AI–energy efficiency relationship.Column (1) presents the total effect of AI on energy efficiency, while column (2) shows that AI significantly promotes GTI at the 1% level.This suggests that AI development exerts a substantial and positive influence on the advancement of green technologies, underscoring its capacity to catalyze sustainable innovation.In column (3), when both AI and GTI are regressed on energy efficiency, the coefficient for GTI remains positive and statistically significant at the 5% level.This indicates that GTI contributes to enhancing energy efficiency and functions as a partial mediator.The clustering of AI enterprises appears to foster green innovation, which in turn drives improvements in energy efficiency.These results support Hypothesis 2 regarding the mediating role of GTI.

Rationalization of industrial structure

Table 6presents the mediating mechanism analysis involving the rationalization of the industrial structure (ISR).Column (1) shows the total effect of AI on energy efficiency, while column (2) reveals that AI significantly promotes ISR at the 1% level, suggesting that AI contributes to industrial upgrading and structural optimization.In column (3), when AI and ISR are jointly regressed on energy efficiency, both variables show positive and statistically significant coefficients at the 1% level.These findings indicate that ISR serves as a significant mediating pathway through which AI enhances energy efficiency.Thus, the results provide empirical support for Hypothesis 3, confirming that AI facilitates energy efficiency improvements by driving the rationalization of the industrial structure.

After establishing the fundamental connection between artificial intelligence (AI) development and energy efficiency, this study proceeds to examine the nuanced dynamics of this relationship under varying boundary conditions⁹⁷。Specifically, we investigate how environmental regulations and the developmental stage of resource-based cities moderate the impact of AI on energy efficiency.This focus is grounded in the recognition that environmental regulation significantly shapes the technological and operational landscape of cities⁵⁷。Previous research suggests that technological innovation tends to be more effective in regions with stringent environmental oversight⁶⁰。Likewise, the economic trajectory of resource-based cities—categorized as growing, mature, declining, or regenerative—is closely tied to the energy sector, and each stage exhibits distinct characteristics⁵⁶。These structural differences create divergent incentives and capacities for integrating AI to enhance energy efficiency.By exploring these moderating factors, we aim to refine our understanding of the contextual conditions under which AI contributes most effectively to energy efficiency.

Moderate role of informal environment regulation

We begin by analyzing the role of informal environmental regulation (IER).Table 7, column (1), presents the moderating effects of IER on the relationship between AI development and energy efficiency.The interaction term between IER and AI development is both positive and statistically significant at the 1% level, indicating that AI’s impact on energy efficiency is more pronounced in areas with strong informal regulatory mechanisms.These findings support Hypothesis 4.

Moderate role of the stage of resource development

Next, we consider how the developmental stage of resource-based cities influences the effectiveness of AI in improving energy efficiency.Our analysis reveals that AI exerts a stronger positive effect in declining and regenerative cities compared to growing and mature ones, likely due to their heightened need for transformation and greater potential for system optimization.In Table 7, column (2), the moderation analysis indicates that mature cities are less energy efficient than growing cities, with declining and regenerative cities performing even worse.However, the interaction term between AI and mature cities is not statistically significant, suggesting no meaningful difference in AI’s impact on energy efficiency between mature and growing cities.In contrast, the significantly positive interaction terms for declining and regenerative cities indicate a more substantial effect of AI in these contexts.This may reflect the targeted application of AI technologies to enhance efficiency amid resource depletion in declining cities, and to optimize energy management in regenerative cities transitioning from historical resource dependency.These results empirically validate our assumption that AI’s contribution to energy efficiency varies across different stages of resource-based urban development.

Amid growing global emphasis on sustainable development, the role of AI in advancing energy efficiency has garnered considerable attention from both policymakers and researchers.

This study assesses the extent to which AI development facilitates sustainable practices in China, particularly in the energy domain.Three key findings emerge from our analysis.First, AI development positively influences energy efficiency.Second, this effect operates primarily through two channels: green innovation and industrial structure rationalization.Specifically, AI promotes the development and deployment of green technologies while also enabling a more balanced and efficient allocation of industrial resources.Third, the moderating analysis reveals that AI’s impact on energy efficiency is significantly amplified in cities with strong informal environmental regulations, and is comparatively muted in cities with looser regulatory environments.Additionally, AI has proven particularly effective in enhancing energy efficiency in resource-based cities that are either in decline or undergoing regeneration, relative to their growing or mature counterparts.Our study’s results align with existing research, emphasizing the significant impact of AI on energy efficiency.By comparing our findings with other studies, we have identified the mechanisms through which AI affects energy efficiency and the conditions under which these effects are most pronounced.This comparison not only validates our results but also highlights the unique contributions of our study.

These findings contribute significantly to the scientific understanding of AI’s role in sustainable development by offering empirical evidence of its impact on energy efficiency.By elucidating the mechanisms linking AI to energy efficiency, our research provides critical insights for policymakers and industry stakeholders, underscoring the importance of incorporating AI into strategic frameworks to advance sustainable development.The practical implications are particularly salient for China, given its energy scarcity and uneven resource distribution.In light of China’s ambitious targets to peak CO₂ emissions by 2030 and achieve carbon neutrality by 2060, harnessing AI to enhance energy efficiency and drive industrial transformation in resource-based cities is imperative.This strategy not only facilitates the attainment of domestic sustainability goals but also strengthens China’s position as a global leader in sustainable development.

Policy implication

Based on the above findings, several key policy implications emerge for promoting sustainable development through AI adoption.First, the positive impact of AI on energy efficiency is significantly amplified in cities with robust informal environmental regulations.Policymakers should therefore prioritize the reinforcement of environmental governance by implementing formal regulations, enhancing public awareness, supporting environmental NGOs, and cultivating a culture of sustainability at the community level.Second, AI exerts a stronger influence on energy efficiency in declining and regenerating resource-dependent cities.Targeted investments in AI infrastructure and applications during these critical transition stages can support industrial restructuring.Simultaneously, efforts should be made to cultivate conducive environments for AI adoption in growing and mature cities.Early integration of AI technologies in these areas can yield long-term benefits by embedding sustainable practices and mitigating future transition risks.Third, AI enhances energy efficiency primarily through green technological innovation and industrial structure optimization.Policymakers and industry leaders should leverage AI to drive innovation and streamline industrial operations, thereby promoting sustainability and unlocking the full potential of AI-driven transformation.

Limitations and future research

Our analysis is based on data from China, which may limit the generalizability of the findings to other regions or developed countries.Future research should incorporate cross-national datasets to validate and expand upon our conclusions, enabling a broader understanding of AI’s role in global energy efficiency improvements.Furthermore, this study does not delve into the differential impacts of various AI subtypes.For instance, generative AI may present unique implications worth exploring in greater depth in subsequent research.

作者和隶属关系

1. School of Economics and Management, Sichuan Normal University, Chengdu, 610101, Sichuan, China

   Jun Zeng & Tian Wang

贡献

Jun Zeng and Tian Wang designed the study, drafted the article, acquired the data, completed the analysis, critical revision for important intellectual content together.

Corresponding author

Correspondence to天王。

竞争利益

作者没有宣称没有竞争利益。

Publisher’s note

关于已发表的地图和机构隶属关系中的管辖权主张，Springer自然仍然是中立的。

开放访问本文在Creative Commons Attribution-Noncormercial-Noderivatives 4.0国际许可下获得许可，该许可允许任何非商业用途，共享，分发和复制以任何媒介或格式的形式，只要您提供适当的原始作者和来源的信用，请符合原始作者和来源，并提供了与Creative Commons的链接，并指示您是否修改了许可的材料。You do not have permission under this licence to share adapted material derived from this article or parts of it.The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material.If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this licence, visithttp://creativecommons.org/licenses/by-nc-nd/4.0/。重印和权限

关于这篇文章