A critical and constructive reading of the UNFCCC paper on artificial intelligence and climate action, one year later<\/p>\n\n\n\n
In July 2025, the Technology Executive Committee (TEC) of the United Nations Framework Convention on Climate Change (UNFCCC), together with the United Nations Industrial Development Organization (UNIDO), published a 170-page technical paper entitled \u201cArtificial Intelligence for Climate Action: Advancing Mitigation and Adaptation in Developing Countries.\u201d It remains one of the most comprehensive documents to date on the intersection of artificial intelligence and climate change from the perspective of developing countries.<\/p>\n\n\n\n A year later, the AI landscape has continued to accelerate. New models, new commitments, new promises. But the fundamental questions raised by that paper remain open\u2014and some it didn't address have become even more urgent.<\/p>\n\n\n\n At DelPlata Green, we work at the intersection of sustainability, technology, and impact measurement. We do so guided by a conviction that underpins everything we do: artificial intelligence must serve the common good. Not as a slogan, but as an evaluation criterion. And it is precisely from this perspective that I want to revisit this document: to acknowledge what it does well, point out what it lacks, and formulate the question that I believe should be at the heart of the conversation but which the paper never explicitly poses.<\/p>\n\n\n\n AI for good, yes. But\u2026 for whose good?<\/strong><\/em><\/p>\n\n\n\n We must start with what is appropriate: recognizing that this document fills a real void.<\/p>\n\n\n\n Before this paper, the conversation about AI and climate change was dominated by two opposing and incomplete narratives. On one hand, there was Silicon Valley's techno-optimism: AI will solve everything. On the other, there was widespread distrust: AI consumes too much energy, AI is biased, AI is part of the problem. The UNFCCC\/UNIDO paper does something more useful than simply taking sides: it maps the landscape.<\/p>\n\n\n\n The central chapter of the document (chapter 4) covers ten areas of AI application for climate action in developing countries: early warning systems, Earth observation, climate simulation and prediction, natural resource management, energy management, transportation, disaster risk reduction, emerging applications of language modeling (LLMs), education and community engagement, and an overview with reference tables.<\/p>\n\n\n\n But where the paper truly distinguishes itself is in what it had the courage to include: an entire chapter dedicated to risks and challenges (Chapter 6). The document points out that the same AI that optimizes renewables is also used to make oil and gas extraction more efficient. It mentions the gender bias in climate datasets. And it mentions the risk that generative AI will amplify climate misinformation.<\/p>\n\n\n\n All of this deserves recognition. It is a serious document, honest in its limitations, and necessary.<\/strong><\/p>\n\n\n\n That said, several of the risks identified in the paper are accurately described but insufficient. It names them, but doesn't address their underlying causes.<\/p>\n\n\n\n Energy and water consumption: more than just a technical data point<\/strong><\/p>\n\n\n\n The paper correctly states that AI has a significant energy and water footprint and recommends using Small Language Models (SLMs) instead of large models whenever possible. However, it doesn't provide enough quantification for a decision-maker to understand the magnitude of the problem in their context. Training a large model can consume as much water as a small community needs for months.<\/p>\n\n\n\n The digital divide: it's not just "lack of internet"\u201c<\/strong><\/p>\n\n\n\n The document describes the digital divide as a combination of limited connectivity, a lack of computing power, and a shortage of skilled professionals. This is correct, but incomplete. What the paper describes as a \u201cdivide\u201d is, in reality, something more structural: a new form of technological dependence. If the models, training data, cloud infrastructure, and talent are concentrated in the United States, China, and Europe, developing countries not only \u201carrive late\u201d to climate AI\u2014they arrive in a position where they consume solutions designed by others, with data from others, and optimized for other contexts.<\/p>\n\n\n\n Beyond what the document says without going into sufficient detail, there are risks that are simply not mentioned \u2014 and which, from my experience in sustainability consulting, I consider central.<\/p>\n\n\n\n Institutional technological solutionism<\/strong><\/p>\n\n\n\n When an international organization or government adopts AI for climate action, there is a risk that it will replace investment in proven and cheaper solutions\u2014basic infrastructure, local governance, ecosystem restoration\u2014rather than complement them. The question \u201cIs AI the best tool for this problem?\u201d should precede any implementation. And the answer, often, will be no.<\/p>\n\n\n\n Data extractivism<\/strong><\/p>\n\n\n\n Developing countries possess something that AI models desperately need: unique climate, ecological, and social data. There is a real risk that this data will be collected by external actors, used to train proprietary models, and that the benefits will not return to the communities that generated it. It is the same extractive pattern we know well in Latin America with natural resources: the resource leaves, the added value is generated elsewhere, and what returns\u2014if it returns at all\u2014are finished products at market price. Only this time the resource is not lithium or copper. It is data.<\/p>\n\n\n\n Accelerated obsolescence<\/strong><\/p>\n\n\n\n AI is evolving at a speed unprecedented in other technologies. For a country with limited resources that has invested significantly in a climate AI solution, this obsolescence is no minor inconvenience. It represents a serious financial and operational risk. And it raises an uncomfortable question: are we building local capacity, or are we creating dependence on technological cycles we don't control?<\/p>\n\n\n\n AI greenwashing<\/strong><\/p>\n\n\n\n \u201cWe implemented artificial intelligence to monitor our emissions\u201d sounds impressive in an ESG report. But if the data is bad, if the model isn't calibrated for the local context, if the results don't translate into decisions, if no one verifies the outputs\u2014what we have is declarative sustainability, not demonstrable sustainability.<\/p>\n\n\n\n All of the above converges on a question that I consider central and that the UNFCCC\/UNIDO paper never explicitly formulates: for whose benefit?<\/p>\n\n\n\n I propose five questions that should accompany any AI project applied to climate action:<\/p>\n\n\n\n The last question is the most uncomfortable \u2014 and the most necessary.<\/strong><\/p>\n\n\n\n The UNFCCC\/UNIDO paper of July 2025 was a valuable and necessary contribution. It mapped the territory, documented real-world applications, and had the honesty to name risks that other institutional documents prefer to ignore. It is worth reading, especially Chapter 4 as a technical reference and Chapter 6 as a starting point for a deeper discussion.<\/p>\n\n\n\n But a year later, that conversation remains unfinished. AI for climate action continues to be deployed\u2014with more resources, with greater urgency, with more promises. And the questions that aren't being asked remain unanswered.<\/p>\n\n\n\n Because artificial intelligence for the common good is not defined by the technology used. It is defined by whom it serves, who decides, and who reaps the benefits.<\/strong><\/p>\n\n\n\n At DelPlata Green<\/strong> We help companies move from declared sustainability to demonstrable sustainability, with the technology of Zero Carbon One<\/strong>.<\/p>\n\n\n\n \ud83d\udce9 info@delplatagreen.com<\/a><\/p>\n\n\n\n \ud83c\udf10 delplatagreen.com<\/p>\n\n\n\n
By Silvina \u2014 DelPlata Green \u00b7 AI for the common good<\/em><\/p>\n\n\n\nA much-needed map<\/strong><\/h2>\n\n\n\n
The risks he mentions but doesn't fully address<\/strong><\/h2>\n\n\n\n
The silences: what the paper doesn't say<\/strong><\/h2>\n\n\n\n
The fundamental question: AI for the common good as a criterion, not as a slogan<\/strong><\/h2>\n\n\n\n
\n
One year later: the unfinished conversation<\/strong><\/h2>\n\n\n\n
Main source<\/strong><\/h2>\n\n\n\n