Artificial intelligence for good

• Artificial Intelligence (AI) comprises a rich set of methods and disciplines, including vision, perception, speech and dialogue, decisions and planning, problem solving, robotics and other applications that enable self-learning. AI is best viewed as a set of technologies and techniques used to complement traditional human attributes, such as intelligence, analytical ability and other capabilities.
• AI has the potential to drive progress across all the UN Sustainable Development Goals, advancing sustainable and economic development, but its use also entails extensive social, economic and ethical implications. Many governments and organizations are therefore preparing for the widespread adoption and use of these technologies.  Developments in AI are closely tied to data policies, including data protection and privacy legislation.
• Notable breakthroughs are occurring in research labs, corporate boardrooms and industry (robotics and communication networks), as well as with the consumers using AI. The extent to which AI affects different countries and regions will depend on their economic structure and extent of digitization in key sectors including health, agriculture and manufacturing.
• ITU is engaged in a body of work about how AI can influence telecommunication and radiocommunication networks, as well as the broader information and communication technology (ICT) environment.
• A robust enabling environment is necessary for driving innovation and trusted use of AI technologies. Development of policy must take into account the needs of specific user groups to avoid discrimination and ensure that everyone can experience the benefits of AI (including poorer communities, children, persons with disabilities and indigenous peoples).​
  

Artificial Intelligence (AI) comprises a set of widely different technologies, which can be broadly defined and grouped together as 'self-learning, adaptive systems'.  There are various approaches to defining AI:

• In terms of technologies, techniques and/or approaches (e.g., a neural network approach to machine translation);
• In terms of purpose (facial recognition, image recognition).
• In terms of functions (e.g., the ability to understand language, recognize pictures, solve problems, and learn, according to the Cambridge Dictionary).
• In terms of agents or machines or algorithms (e.g., robots, self-driving cars).

AI comprises a rich set of methods and disciplines, including vision, perception, speech and dialogue, decisions and planning, problem-solving, robotics and other applications that enable self-learning. AI is best viewed as a set of technologies and techniques used to complement traditional human attributes, such as intelligence, analytical ability and other capabilities. AI, Machine Learning (ML) and modern data techniques have been greatly enabled by recent advances in computer processing, power and speed, and advances in AI depend in turn on advances in data techniques. 

Across many sectors, AI offers advantages of new and innovative services, and the potential to improve scale, speed and accuracy. AI extends and combines many of these advantages with insights from statistics and big data. Based on trend analysis, AI helps move business and policy models and regulatory approaches from descriptive analysis and trendspotting to more sensitive, proactive predictive and evidence-based models and approaches. For example, AI is being used to spot patterns in health vulnerabilities and insurance risks, among many other applications.

The use of AI tools and techniques is driving new opportunities across many diverse domains. AI and other algorithms are used extensively in online search, entertainment, social media, self-driving cars, visual recognition, translation tools, smart assistants/speakers, voice-to-text and many other applications.​

Policy and regulatory frameworks for AI remain at an initial, formative stage. Key policy questions that have arisen relate to:

• Use, accuracy and methods used by AI tools, including in relation to humans, including the development of bias in machine learning models and the data used to train them;
• Accountability and responsibility accompanying the use of AI models;
• Purposes for which they are used; as well as
  
• Datasets used to train them, and the methods used to collect (or 'scrape') data.

Major questions arise in relation to the quality and representativeness of the datasets that have been used to train AI. Researchers are also working to improve the accuracy of software tools and algorithms, amid concerns they magnify racial and socioeconomic biases. For example, while the COVID-19 pandemic has, in many countries, disproportionately affected minorities, AI-based prediction models may not always include other relevant health disparities and thus may not always correctly assess risks for each person or group.

Data ownership has emerged as a major issue. Data must be continually aggregated to help keep every model valid, accurate and effective in predicting outcomes. There is an increasing proliferation of deep fakes (e.g. convincing programmed videos of high-profile personalities saying or doing things the video creator has requested) and other AI-generated materials. Indeed, deep fake technologies have been used to generate misleading videos in the mainstream media, as well as to animate photos of long-dead celebrities. Aside from crucial ethical questions about use and accuracy, who owns the copyright to these “new" works?

AI has extraordinary potential to act as a force for good. However, considerable challenges persist: 

1. Fundamental trust and the transparency of models: It is frequently unclear how deep learning models arrive at their conclusion and the models may be opaque and not very transparent. Depending on the purpose, although researchers want AI to make accurate predictions, some researchers may still prefer simple yet explainable AI models to more accurate, but more opaque models. Some people are willing to 'trust' machines with complex systems and tough decisions, while others may fundamentally prefer to retain some degree of human involvement. 
2. Bias: While AI can be used for extremely useful purposes, it can also inadvertently generate poor or inappropriate purposes or unintended outcomes. There is growing concern about issues of racial, disability and gender bias in AI and machine learning algorithms, and their wider impact on society at large. The accuracy of an AI ML model depends on the quality and the amount of data that an AI model is trained on. In real life, data is often poorly labelled. Standardization of data sets is needed. Data are also often biased. Training courses on the ethical applications of AI are needed, and not just for computer engineering students. 
3. Data availability and ownership: Getting data is very difficult. Best practices need to be defined under which circumstances data can be made available and to whom, whilst respecting ownership and explicit promises of confidentiality for certain types of data. 
4. Data privacy and security: Security breaches due to cyber-attacks can have horrific consequences. Techniques such as federated learning can reduce the risks by enabling AI models to be trained across devices that hold data locally, without exchanging them,  while privacy-preserving technologies help ensure personal data protection. 
5. Limited know-how: AI can tackle many problems, but there is only a limited pool of experts who know how to apply AI ethically. Many researchers point to the need to involve sociologists and policy-makers in discussions, rather than assume that AI designed by a narrow pool of 'technologists', computer engineers and data scientists will be used ethically. Education is key to learn about the responsible use of AI. 
6. Equitable uses of AI: AI research is computationally intensive. Unequal access to computing power and to data deepens the divide between a few companies and elite universities which do have resources, and the rest of the world which does not.
   

The Potential of AI to be used for Good

AI has many important applications to help accelerate progress towards achieving the UN's Sustainable Development Goals (SDGs). AI makes new services possible in many domains important for the SDGs – for example:

Figure: Machine Learning Models in Health​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

1. In education for SDG4, AI is being used to monitor pupils' attention or to carry out emotional surveillance to determine how comfortable children are learning certain subjects, identifying students who are struggling before their test results become available. In many countries, AI is being used to develop personalized testing tools, to identify areas of weakness and help students improve.
2. In finance, AI commonly provides insights and assistance with accounting and investment work, including automating routine tasks and uncovering new data patterns that could help with micro-investments to combat poverty (SDG1) or introduce new financial services and infrastructure (SDG9).
3. In manufacturing, industry and sustainable economic growth (SDG8), the use of automation, fifth generation (5G) mobile telephony, the Internet of Things (IoT) and more extensive robotics has transformed factories, supply depots and warehouses throughout Asia and Europe and the Americas, enabling more efficient and effective manufacturing, production and distribution.
4. Online translation and publishing software has transformed online publishing, media, and the distribution of text and materials, including books and websites. Many industries now employ chatbots and intelligent assistants to cope with routine customer queries and concerns.
5. In transport, AI is helping facilitate fully autonomous vehicles and autonomous driving systems (ADS), which steadily improve their driving and navigation skills through self-learning programs, as well as for real-time traffic management through urban spaces.
6. In agriculture, AI can be used for farm management and predictive analytics based on data from crop, soil, and weather monitoring to support decision-making and to optimize the use of resources (water, fertilizers, etc.). It can help detect pests and diseases by analysing images of plants and data on the behaviour of livestock. Agricultural robots and automation are saving labour in many resource-consuming tasks.
   

ITU is engaged in a wide range of work relating to new and emerging trends in AI, as well as helping ITU Members, Member States and stakeholders prepare for its wide-ranging policy and regulatory consequences.

AI for Good as a platform

The AI for Good platform [https://aiforgood.itu.int/programme/] focuses on uses of AI to help fulfil the essential needs of humanity, including achieving the 17 SDGs set out by the UN to be achieved by 2030, as an all-year, always online programme. The goal of the Summit is to identify practical applications of AI to advance the sustainable development goals and scale solutions for global impact. The Summit is organized by ITU in partnership with 38 UN sister agencies and co-convened with Switzerland. 

The AI for Good YouTube channel hosts hundreds of videos highlighting interviews with AI leaders and innovators, innovations and demos showcasing AI solutions to accelerate the SDGs as a one-stop shop to catch up on emerging trends in AI for Good. Subscribe to the channel and join online for new updates and exclusive content as they go live on to explore ideas, insights and active discussions around AI to achieve the SDGs. The channel features keynotes, webinars, perspectives, an Innovation Factory and social media.

Policy and Regulation

Through annual regulatory surveys and monitoring (https://1f8a81b9b0707b63-19211.webchannel-proxy.scarabresearch.com/itu-d/sites/regulatory-market/), ITU tracks the growth of national AI strategies and policies. Machine Learning models are trained and fed by vast quantities of data, so it is vital to consider national policies in data privacy, regulation and data protection as well as approaches to the Internet of Things (IoT), sensor networks and the 5G networks making data transmission possible, when considering national approaches to AI.

According to ITU's latest Telecommunication/ICT Regulatory survey, some 18 countries had prepared specific strategies on AI by 2019, although more countries have AI sector-specific strategies, which has risen to 49 countries in 2021. However, AI encompasses diverse set of technologies, and few national strategies consider the field in total. Countries also have to consider the treatment of data flows now generated by IoT and sensor networks which feed ML models and AI technologies. Several countries, including the United States and Saudi Arabia, have prepared strategies on all three topics (5G, IoT and AI). According to the United Nations Conference on Trade and Development (UNCTAD; 2020), some two-thirds of all countries have developed policies for data protection, including AI for development.

Figure: Numbers of countries with strategies for emerging technologies, 2020​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

Setting Standards

Moving forward, international standards—the technical specifications and requirements needed for AI and other technologies to perform well—can help address real and perceived risks by setting clear boundaries and making machine learning (ML) predictable, reliable and efficient.

AI and ML are gaining ground in ITU's standardization work, with research, analysis and stakeholder discussions focusing on network orchestration and management, multimedia coding, service quality assessment, and various aspects of telecom management, operation and services, as well as cable networks, digital health, environmental efficiency, and autonomous driving.  

AI in Radiocommunication Standards

ITU Radiocommunication (ITU-R) study groups and forthcoming reports examine the use of AI in radiocommunications:

• ITU-R Study Group 1 covers all aspects of spectrum management, including spectrum monitoring. Question 241/1 looks at “Methodologies for assessing or predicting spectrum availability". 
• ITU-R Study Group 6, dedicated to broadcasting services, is studying AI and ML applications: 
  
• ​ Question ITU-R 144/6, “Use of AI for broadcasting", considers the impact of AI technologies and how can they be deployed to increase efficiency in programme production, quality evaluation, programme assembly and broadcast emission. 
• Recommendation ITU-R BS.1387: “Method for objective measurements of perceived audio quality" about AI in the field of broadcasting.
  
• Report ITU-R BT.2447, “AI systems for programme production and exchange", discusses current applications and near-term initiatives, revised regularly to reflect the latest progress on AI for the applications in broadcasting.

ITU-T standards addressing AI and Machine Learning

ICT companies in the networking business are introducing AI and ML to optimize network operations and increase energy and cost efficiency. New ITU standards provide: an architectural framework to integrate ML into 5G and future networks (ITU-T Y.3172); an evaluation framework for intelligence levels across different parts of the network (ITU-T Y.3173); and a framework for data handling in support of ML (ITU-T Y.3174. These standards originated in discussions by the ITU-T Focus Group on 'Machine Learning for Future Networks including 5G'.

The ITU-T AI/ML in 5G Challenge, introduced in 2020, rallied like-minded students and professionals from around the globe to study the practical application of AI and ML in emerging and future digital communication networks. The first edition attracted over 1,300 students and professionals from 62 countries, competing for global recognition and a prize fund of USD 36,000. By mapping emerging AI and ML solutions, the Challenge fosters a community to support the evolution of ITU standards. See the Challenge GitHub (Clickable link: https://github.com/ITU-AI-ML-in-5G-Challenge).

The ITU-T Focus Group on 'Environmental Efficiency for AI and other Emerging Technologies' aims to benchmark best practices and describe pathways towards a standardized environmental framework. 

The ITU-T Focus Group on 'AI for Health', convened jointly with the World Health Organization (WHO), is working towards a framework and processes for performance benchmarking of AI for health solutions, including in response to COVID-19. It represents an open platform open to all stakeholders from different fields. The Focus Group works at the interface of multiple fields (e.g., ML/AI, medicine, regulation, public health, statistics) and includes decision-makers who value a standardized benchmarking framework.

The ITU-T Focus Group on 'AI for Autonomous and Assisted Driving' is working to establish international standards to monitor and assess the behaviour of the AI 'drivers' in control of automated vehicles. 

The Global Initiative on 'AI and Data Commons', established in January 2020, assembles key resources for AI projects aligned with SDGs, supports rapid implementation and aims to help bring AI for Good projects  to global scale. 

AI and ML are widely used to construct models for the qualities of speech and other audio-visual (AV) data. An ITU-T working group on 'AI-enabled multimedia applications' (ITU-T Q5/16)  is discussing standard requires for the quality assessments in AV streaming, in progressive-download and adaptive-bitrate AV (ITU-T P.1203) and video streaming (ITU-T P.1204).

New ITU-T standards address intelligent network analytics and diagnostics (ITU-T E.475) and the creation and performance testing for ML models to assess the impact of the transmission network on speech quality for 4G voice services (ITU-T P.565). Others address environmental sustainability, cable networks, and operational aspects of service provision and telecom management. 

Other new ITU standards describe a datacentre infrastructure management (DCIM) system based on Big Data and AI technology (ITU-T L.1305), aiming to reduce the energy needs of datacentres, and provide the framework for a premium cable network platform to support industry in offering advanced multimedia services (ITU-T J.1600) for AI-assisted cable networks.

AI and ICT for development issues

ITU is gathering and disseminating information about effective and sustainable AI solutions to equip relevant stakeholders with evidence and knowledge to adopt and leverage relevant AI applications. FAO and ITU publish the e-Agriculture in Action Report: AI for Agriculture, which identifies informative case studies of AI uses in agriculture, with valuable insights about implementation, success factors, and lessons learned. 

In addition, ITU engages directly in deployment and testing of some promising AI applications to support the provision of SDG-related services and tools. In Senegal, ITU, in close collaboration with WHO and the Ministry of Health and Social Action, is leading the work on pilot testing an AI application for automatic detection of diabetic retinopathy to improve the coverage and accessibility of screening. This solution could support ophthalmologists in analyzing digital images of retinas. 

ITU has released the AI and big data for development 4.0 report which highlights opportunities and outlines good policy and regulatory practices for implementation, with suggestions in managing and overcoming barriers. The report describes the building-blocks of a national AI and data system for development, including governance, regulation, ethical considerations, digital and data skills, the technological innovation landscape and opportunities for international collaboration.​

Last update: Jan 2023​