From Paris to Practice: Making UNESCO's AI Ethics Framework Actually Work

In November 2021, all 193 member states of UNESCO gathered in Paris and did something that rarely happens in international affairs: they agreed on a single document. The Recommendation on the Ethics of Artificial Intelligence became the world's first global normative framework for how AI should be designed, deployed and governed. Every country signed on to the same four core values—respect for human rights and dignity, support for peaceful and just societies, protection of diversity, and environmental stewardship. Ten principles followed: proportionality, transparency, human oversight, fairness, accountability. The document was specific enough to matter. It said AI systems should never be used for social scoring or mass surveillance. It insisted that decisions affecting human lives must remain in human hands.

Nearly five years later, the harder work began. Chatbots now draft legal documents. Algorithms screen job applicants and flag welfare beneficiaries. Students submit essays written with machine assistance. The principles that seemed so clear in Paris have collided with the messy reality of deployment. Agreeing on ethics, it turned out, was the easier task. Building institutions that keep those principles working for decades—that is where the real challenge lives.

The UNESCO framework addresses two kinds of sustainability, and both demand attention. The first is environmental. Training and running large AI models consumes staggering amounts of electricity and water. Data centres are multiplying across continents. The International Energy Agency projects that electricity demand from data centres could more than double by 2030. Cooling these facilities draws heavily on freshwater, often in regions already facing acute shortages. A 2025 study released by UNESCO with University College London found that practical measures—using smaller models built for specific tasks, keeping prompts short—could cut energy use sharply without meaningful loss of performance. For India, this is not abstract. Power grids in several states already run close to capacity during summer. Groundwater levels in cities attracting data centre investment, including parts of Maharashtra, Telangana and the National Capital Region, are under strain. Growth in AI infrastructure has to be planned alongside water and energy policy, or one will undercut the other.

The second meaning of sustainability concerns institutional durability. Ethics frameworks have a habit of fading once the launch event ends. Committees meet a few times and then go quiet. A sustainable approach embeds ethical review inside the routine work of ministries, regulators, universities and companies, so that it continues when political attention shifts elsewhere. Ethics works best as infrastructure—built once and maintained continuously.

UNESCO developed two practical instruments to help countries move from agreement to action. The Readiness Assessment Methodology allows a government to examine its own preparedness for AI across legal, social, cultural, scientific and economic dimensions. Dozens of countries across Africa, Asia, Latin America and Europe have completed or begun this exercise. The results give policymakers an honest picture of where laws are missing, where skills are thin and where public trust is weak. The Ethical Impact Assessment is meant for public bodies before they buy or deploy an AI system. It asks direct questions: Who could be harmed? Whose data trains the system? Can affected citizens challenge a decision? How will errors be found and corrected? A municipal body introducing facial recognition, or a state department using an algorithm to identify welfare beneficiaries, would answer these questions before signing a contract.

Colombia offers one of the clearest examples of a country turning principles into policy. It completed UNESCO's Readiness Assessment, then fed the results into national planning. In February 2025, the country approved CONPES 4144, its National Artificial Intelligence Policy, with 106 actions and an investment of roughly 116 million US dollars running until 2030. Ethics and governance stand first among the policy's six pillars. The Colombian approach is sustainable because ethics runs through every pillar—the policy treats AI as a driver of growth and social inclusion that must also respect the environment, and it gives clear responsibility for delivery to named agencies. Colombia's Constitutional Court added another layer in 2024, ruling that judges may use AI tools as support provided they disclose the use, check the output and keep the final decision in human hands. For a middle-income country with a large rural population and sharp regional inequality, the lesson is practical: ethics lasts when it is tied to a budget and to an agency answerable for results.

India supported the Recommendation in 2021 and has since assembled parts of its own governance architecture. The Digital Personal Data Protection Act of 2023 sets rules on consent and data handling. The IndiaAI Mission, approved in 2024, aims to expand public computing capacity, build national datasets and support startups. In January 2025, the Ministry of Electronics and Information Technology and UNESCO began stakeholder consultations under the Readiness Assessment Methodology, with sessions in several cities including Bengaluru and Hyderabad. The India AI Readiness Assessment Report was released at the AI Impact Summit in New Delhi in February 2026, the first summit in this series held in the Global South. For the first time, India has a systematic account of where it meets the UNESCO standard and where it falls short. India's sustainability model rests on shared public infrastructure. Under the IndiaAI Mission, startups, researchers and universities can rent high-end computing power at subsidised rates from a common pool, avoiding the energy cost of every institution building its own facility. Bhashini, the government's language translation platform, makes AI available in Indian languages to people who do not read English. These choices extend the digital public infrastructure approach that India used for payments and identity into the field of AI. The country's circumstances make the UNESCO principles especially relevant. With 22 scheduled languages and hundreds of dialects, AI systems trained mostly on English data can misread or exclude large sections of the population. Welfare schemes increasingly depend on digital identity and automated verification, and a single error can cut a family off from rations or pensions. Agriculture, which employs nearly half the workforce, is seeing early use of AI for crop advice and credit scoring, where biased data could harm small farmers.

Yet despite this progress, the distance between commitment and capacity remains wide. A small number of countries and corporations control most of the world's advanced computing power, training data and AI talent. Many developing nations signed the Recommendation without the technical staff, funding or regulatory bodies needed to enforce it. Global regulation is also fragmenting. The European Union's AI Act entered into force in August 2024, with obligations phasing in through 2026 and 2027. The Council of Europe opened its Framework Convention on Artificial Intelligence for signature in 2024. The United Nations agreed in 2025 to establish an independent international scientific panel on AI and a global dialogue on its governance. China, the United States and others follow their own approaches. Companies operating across borders now face different and sometimes conflicting rules. Smaller countries risk becoming rule-takers, importing systems built elsewhere and governed by standards they had no hand in shaping. The UNESCO framework, because it carries the consent of every member state, offers a shared reference point that can help align these separate efforts. Experts point to several measures that would give it lasting force: permanent institutions with statutory backing and clear mandates; mandatory impact assessments for public procurement of high-risk AI; green AI standards requiring disclosure of energy and water consumption; sustained investment in public literacy so judges, police officers, teachers, doctors and district officials can question AI outputs and recognise when a system is wrong; inclusion designed in from the start, with systems tested for performance across languages, genders, castes, regions and disabilities before deployment; and international cooperation extending to shared computing resources, open datasets and joint research among countries of the Global South. The real test will come far from conference halls. It will come when a pensioner in rural Bihar asks why an algorithm stopped her payments and receives a clear answer. It will come when a student in Srinagar uses an AI tutor that understands her language, or when a city chooses a smaller, efficient model over a power-hungry one because the rules require it to count the cost. The 193 nations that adopted the Recommendation set the direction. Whether they build the courts, regulators, classrooms and power grids to follow it will decide what the agreement means for the next generation.