本文检视了大型科技公司反复宣称生成式 AI 将带来重大气候效益，重点放在 Google 于 late-2023 的说法：AI 可在 2030 前将全球温室气体排放降低 5 to 10 percent，幅度大致相当于欧盟每年规模的减排。研究者 Ketan Joshi 追查这个被广泛重复的数字，发现其经由一份 Google-BCG 文件可回溯到 2021 年 BCG 分析，而该分析主要依据客户经验、非公开的实证建模；同时，Google 自身 2023 年永续报告后来也承认，AI 驱动的扩张正在提高其企业排放，但公司在 2025 年前的政策倡议中仍持续引用同一个标题级估算。

Joshi 在 February 2026 的报告检视了 more than 150 项公开主张，并分析了 154 条把 AI 视为净气候效益的具体陈述，结果只有约 25 percent 引用了学术研究，且超过 33 percent 完全没有引用任何公开证据。文章并将此与美国资料中心市场的基础设施趋势并列：燃煤退场放缓、数百 gigawatts 的新燃气发电在开发中，且其中接近 100 gigawatts 明确连结到资料中心，而主要 AI 领导者仍持续提出高信心的气候承诺。文中受访专家指出，许多主张混淆了较旧、能耗较低的机器学习应用与较新的大型生成式系统之间的差异，后者需要高得多的运算与电力。

核心意涵是，对未来气候上行效益的信心正跑在透明证据之前，而缺失的基线揭露使生成式 AI 难以进行稳健的成本效益核算。研究者主张，有效的气候 AI 已存在于更狭窄且高效率的模型中；而新的比较研究也显示，较小模型有时可在更低训练成本下达到与较大系统相当的表现，挑战了「越大越好」的叙事。政策上的关键但书是衡量不透明：在领先公司公开的能源与排放资料有限的情况下，提议的补救措施聚焦于强制且细致的申报，例如每年用电成长总量，以及生成式 AI 负载的明确拆分（例如，6 terawatt-hour 增量中有 2 terawatt-hours 可归因于生成式工作负载）。

The article examines Big Tech’s repeated claim that generative AI will deliver major climate gains, centering on Google’s late-2023 assertion that AI could cut global greenhouse gas emissions by 5 to 10 percent by 2030, roughly comparable to annual EU-scale emissions reductions. Researcher Ketan Joshi traced this widely repeated figure through a Google-BCG paper to a 2021 BCG analysis based largely on client experience rather than disclosed empirical modeling, while Google’s own 2023 sustainability reporting later acknowledged that AI-driven expansion was increasing its corporate emissions even as the company continued citing the same headline estimate in policy advocacy through 2025.

Joshi’s February 2026 report reviewed more than 150 public claims and analyzed 154 specific statements about AI as a net climate benefit, finding only about 25 percent cited academic research and over 33 percent cited no public evidence at all. The article pairs this with infrastructure trends in the US data-center market: coal retirements slowing, hundreds of gigawatts of new gas generation in development, and nearly 100 gigawatts specifically linked to data centers, while major AI leaders continued high-confidence climate promises. Experts quoted in the piece argue many claims blur distinctions between older, less energy-intensive machine-learning applications and newer large generative systems that demand far higher compute and power.

The core implication is that confidence in future climate upside is outpacing transparent proof, with missing baseline disclosures preventing robust cost-benefit accounting for generative AI. Researchers argue useful climate AI already exists in narrower, efficient models, and new comparative work suggests smaller models can sometimes match larger systems’ performance at lower training cost, challenging a bigger-is-better narrative. A central policy caveat is measurement opacity: with limited public energy and emissions data from leading firms, proposed remedies focus on mandatory, granular reporting such as annual power growth totals and explicit splits for generative AI loads (for example, a 6 terawatt-hour increase with 2 terawatt-hours attributable to generative workloads).