Artificial Intelligence, real consequences

Front matter

AI* Juliana E. Gonçalves ARTIFICIAL INTELLIGENCE REAL CONSEQUENCES How AI is changing the way we live

Artificial Intelligence, Real Consequences AI* Perhaps the first images that come to mind when thinking about artificial intelligence (AI) are machines and robots that look and behave like humans. Machines that are self-aware, conscious, and autonomous, just like us. This version of AI is undoubtedly the one most exploited in sci-fi movies and books: The “grand dream” of creating machines that have cognitive capacities as humans do. However, many scientists and critics challenge this grand dream. First, many believe such a dream is too grand to be feasible. Second, many argue that pursuing this grand dream is not desirable at all. T wo questions then arise: What is feasible in AI? What is desirable for AI?

Although these questions remain open, AI is increasingly becoming more influential in our society. This influence may seem unnoticeable for most people, but it has profound societal consequences. Opportunities for AI are promising, but threats are looming. On the one hand, AI brings with it a fairly untapped potential for positive societal impact. AI may help to ensure that operations run more efficiently, improving services for people and reducing operational costs. For example, AI has been an important tool in several efforts to tackle the impacts of the COVID-19 pandemic. Among other tasks, AI has been used in outbreak prediction and tracking, contact tracing, early diagnosis, and curbing the spread of misinformation.

Chapter 1

Historical accounts of AI usually bring us back to the famous scientist Alan T uring. T uring (1912 - 1954) was an English mathematician often cited as the father of computer science and artificial intelligence. Among his many contributions to these fields, in 1950, T uring published a landmark research article in the British journal, Mind. In the article “Computing Machinery and Intelligence” , T uring describes a test to verify whether a machine is intelligent. The test consists of a parlour game in which a man and a woman are in two separate rooms and communicate with an interrogator via chat. The interrogator may ask any question that can be communicated via chat to identify the man and the woman. The man tries to convince the interrogator that he is the woman, and the woman tries to convey her real identity. At some point, the man is replaced by a machine. The machine “passes the test” if the interrogator cannot distinguish the machine from the woman. T uring named his test the Imitation Game.

T uring’s article has generated a lot of commentary and controversy throughout the years and remains relevant in the field. T oday, the Imitation Game is called the T uring T est. It is conducted differently: a single room with either a person or a machine, and the interrogator determines whether they are communicating with a real person or a machine. T uring’s article is indeed an important marker in the history of AI, but only possible thanks to previous developments. Arguably, AI has developed building upon previous work both in the field of computer science and in philosophic work reflecting on how humans think, bringing us back to names like Ada Lovelace, the first computer programmer, and Leibniz, Spinoza, Hobbes, Locke, Kant and Hume, famous W estern philosophers.

When T uring published “Computing Machinery and Intelligence” , the term “artificial intelligence” did not exist yet. It was only in 1956 that a group of American computer scientists coined the term for a conference organised at Dartmouth College. The conference is considered the “official birthday of artificial intelligence” and reunited scientists and students working in the field, who became the founders and leaders of AI research. It was at this conference that the first artificial intelligence computer program was presented to the scientific community by Allen Newell, Herbert A. Simon, and Cliff Shaw. From this point, the field received much attention and funding. Many important developments occurred between 1950-1970. However, progress slowed in the 1970s, when a series of disappointments in AI led to cutbacks in government funding. The generally held feeling was that AI did not deliver on the expectations around it and, the following What is AI? years became a difficult period to obtain AI-related funding. This period marks the first AI winter.1 W e can say that AI is a field of ups and downs, with both winters and summers. W e are currently experiencing an AI summer, with major breakthroughs in the field and, again, a lot of expectations about future AI developments. However, some believe that another winter is just around the corner. Others argue that AI winters happen because we expect too much from AI: the higher the fly, the harder the fall. The hopes for AI have always been high, right from the beginning, as T uring himself believed computers would have consciousness by 2000. It is 2023, and we have clearly disappointed T uring. Further, AI experts believe we may never achieve T uring’s expectations simply because it may not be possible to create machines that think. The Dutch computer scientist Edsger W. Dijkstra wrote, "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim” . The British-born American writer, Pamela McCorduck, had similar lines, comparing the question of “Can a machine think?” to whether a chicken has lips. In addition, critics claim that trying to create machines that think should not be AI's goal at all. And yet, this has not stopped efforts to create human-like machines1.

Media often portrays AI as the “grand dream” - the idea of creating machines that can think and behave like humans - sometimes depicted as a utopia and other times as a dystopia, but always with strokes of exaggeration. These representations bring an unrealistic picture of both strengths and flaws of AI that do not match reality at all. In fact, many of the current applications of AI are based on techniques developed in the 20th century but are only now being implemented 1 McCorduck, P., & Cfe, C. (2004). Machines who think: A personal inquiry into the history and prospects of artificial intelligence. CRC Press. thanks to the increasing availability of data and computational power. Notwithstanding the media exaggeration and misrepresentation of AI, there are high hopes for AI and huge concerns about a future with AI.

If AI is not what we see in the media and movies, what is AI? W ell, AI is notably difficult to define. Even researchers and engineers working on AI for many years have not agreed on a unique definition. AI encompasses intelligent systems and machines that emulate human thinking and behaviour and perform tasks that would otherwise require human intelligence. More officially, the European Union describes AI as “systems that display intelligent behaviour by analysing their environment and taking actions - with some degree of autonomy - to achieve specific goals”2. Under this definition, AI applications are multiple. They can be simple and already influencing our daily routines, like AI-powered systems that block spam. They can also be rather sophisticated, like AI-powered image analysis to support cancer diagnosis. Furthermore, AI can be fully integrated into machines in the form of advanced robots and autonomous cars, or operate mostly in the virtual world, like voice assistance, search engines, and speech and face recognition systems. Sometimes AI is equated with smartness, but a smart device does not necessarily have AI. 2 European Commission. (2018). Coordinated Plan on Artificial Intelligence.

HOW DOES AI WORK?

Chapter 2

AI is an umbrella term for techniques computers use to perform multiple tasks, including cognitive tasks (such as learning or self-correction), which in the past have only been performed by humans. Many techniques and technologies belong to the AI domain, including machine learning, deep learning, natural language processing, among others. Most of these techniques work by analysing (huge amounts of ) data to identify patterns and learn about the world. Generally speaking, the learning process of an AI system or machine is no different from a child being taught about the world. For example, children are only able to differentiate between a horse and a zebra once they know that a zebra has black and white patterns, and a horse does not. Similarly, an AI system needs to learn what a horse is and what a zebra is to differentiate between them. While humans are good at extrapolating and reasoning, learning from a relatively small dataset, AI systems require much more data to learn the same amount, or arguably less. Another difference is that AI can process highly complex data in a short amount of time, a task impossible for humans.

A fundamental step in developing AI systems is thus teaching the system to understand the environment in which it will operate so it can correctly perceive information and make accurate decisions based on the information provided. This is called training and requires training data. The training process teaches AI systems to differentiate a zebra from a horse, a bus from a car, or an intense cry from joyous laughter. In the first two cases, the training data has to contain multiple pictures of zebras, horses, buses, and cars and, in the third case, multiple voice samples of people crying and laughing, all needing humans to label the trained samples as horses or zebras, and so on. The quality of the training data highly influences the accuracy of the AI system. For example, if the training data only include front pictures of zebras and horses, the AI system will not be able to recognise side or back pictures. The same happens for voice samples of people yelling and laughing. Even humans sometimes have a hard time saying if some people are laughing or crying. If the training data is suitable, the AI system can be used for the task it has been trained for. Without training, an AI system will be, at best, pointless or inefficient and, at worst flawed and dangerous.

Generally speaking, the learning process of an AI system or machine is no different from a child being taught about the world. For example, children are only able to differentiate between a horse and a zebra once they know that a zebra has black and white patterns and a horse does not. Similarly, an AI system needs to learn what a horse is and what a zebra is to differentiate between them.

The training process teaches AI systems to perceive information and make decisions about it.

To differentiate between zebras and horses, the training data contains images of zebras and horses, properly labelled by humans.

After training, the AI system can be used to identify front pictures of zebras and horses.

A fundamental step in developing AI systems is thus teaching the system to understand the environment in which it will operate so it can correctly perceive information and make accurate decisions based on the information provided. This is called training and requires training data. The training process teaches AI systems to differentiate a zebra from a horse, a bus from a car, or an intense cry from joyous.

Algorithm An algorithm is a series of step-by-step computer operations that can solve a problem in a finite number of steps. Algorithms are at the core of all AI systems. Machine Learning Machine Learning (ML) is a subfield of AI, although the terms are often used interchangeably. ML enables machines to learn from patterns in the data without being explicitly programmed. ML is often based on simpler mathematical methods, like linear regression. ML is everywhere: identifying spam in our mailboxes, providing song and movie recommendations in streaming services, and giving us warnings about suspicious transactions in our bank accounts.

Deep Learning Deep Learning is a specific type of ML that allows AI to learn and improve by processing highly complex data. Deep Learning also works by identifying patterns in the data without being explicitly programmed for it. The difference is that Deep Learning uses artificial neural networks that mimic biological neural networks in the human brain to process information. Deep Learning systems are more complex and require less human intervention; higher complexity and less intervention means more data is required for training. Deep Learning is behind more complex and autonomous technologies, like self-driving cars or robots that perform advanced medical surgery.

A small glossary Artificial Neural Networks Artificial Neural Networks follow the operation principles of biological networks of neurons in the human brain, allowing AI systems to process large and complex datasets. Natural Language Processing Natural language processing (NLP) is also a type of ML. NLP allows computers to recognise, analyse, and interpret written or spoken languages. NLP is at the core of technologies that interact with humans in some way, either via text or audio inputs, including the famous virtual assistants Siri and Alexa and business chatbots. Computer Vision Computer Vision is one of the most prolific uses of AI technologies. It combines pattern recognition and deep learning to review and interpret the content of an image. Computer Vision is behind daily applications like face recognition and advanced image analysis in cancer treatment.

Cognitive Computing Cognitive Computing refers to AI systems deployed to imitate the human thought process by simulating the human cognition process. T o achieve this ambitious goal, Cognitive Computing combines multiple AI methods: self-learning algorithms, pattern recognition by neural networks, and natural language processing. Advances in AI have brought a lot of new words in our daily life. A whole new vocabulary includes terms like algorithms and machine learning. Understanding AI jargon is key to facilitating discussion about the real-world applications of this technology. Before going deeper into the complex world of AI, we provide a small glossary below.

Artificial intelligence has already begun to transform many aspects of our lives and is expected to continue to do so in the future. In this chapter, we describe current applications of AI and opportunities for future developments across multiple sectors like healthcare and education. As AI technologies continue to develop, it is important to consider the potential benefits and opportunities it presents, but also the ethical considerations that must be taken into account to use these technologies in the right way and with the right purposes. Chapter 3 APPLICATIONS & OPPORTUNITIES Chapter 3

Many AI applications have the ultimate goal of helping humans to make better decisions. These applications are known as Intelligent Decision Support Systems (IDSS).

IDSS use artificial intelligence (AI) to help humans make decisions by providing them with relevant information and recommendations. The feature that makes recommendations in streaming services like Netflix or Spotify is an example of IDSS. IDSS help bridge the gap between analysing a situation (which movies or songs are available for the genres I like) and making a decision based on this analysis (which movie or song I will choose).

Of course, the application of IDSS goes beyond making movie and song recommendations.

IDSS can be embedded in various other systems, like a chatbot offering advice on financial investments or a navigation app providing the fastest route by analysing all possible access points and monitoring traffic in real-time. IDSS are also used in healthcare to analyse patient history, medical images, and specific protocols, and provide suggestions for treatment.

One of the reasons why AI is attracting so much attention is the increasing use of AI to make decisions in the public realm. Governments are gradually integrating AI to deliver and improve their operations, with much-untapped potential, particularly concerning policymaking and monitoring. The development and monitoring of public policies require the ability to sense patterns, develop evidence-based programs, forecast outcomes, and measure effectiveness, all concentric on AI capabilities. By leveraging the increasing amounts of data being collected about daily life, AI can help governments to build a comprehensive evidence-based foundation that transcends the barriers between public sectors. Going beyond sectoral barriers, AI can reveal complex patterns and interlinkages between sectors and thus provide comprehensive evidence for decision-makers to design better public policies. However, introducing AI systems into public decision-making raises multiple questions, concerning liability and accountability, transparency, safety and security, impacts on jobs, and ultimately threats to fundamental rights and democracy1. 1 European Parliament (2022). Artificial intelligence: threats and opportunities. Supporting Human Decisions

AI is already a reality in healthcare, pharma, and life sciences industries to support diagnosis, medical image analysis (like X-ray images), and the development of new drugs. AI helps analyse, present, and understand medical and healthcare data. These data can include, for example, patient history and medical images. AI can learn from a huge set of medical images to diagnose diseases like cancer, a task impossible for humans. AI algorithms are already able to detect 95% of skin cancers using medical images1. Using patient data, AI can review their medical history to identify patterns and recommend interventions targeting unique symptoms and characteristics of people. For example, AI systems can suggest specific exercises patients should perform to improve injury recovery. Advances in AI also bring the potential to boost patient care. One example is using so-called carebots - robots that support patient care. Images of nurse-like robots pop up when searching for carebots on the web. However, in reality, carebots look less like robots and more like invisible chunks of computer code connected to cameras. This way, carebots can monitor 1 https://www.karger.com/Article/FullText/504785 patients' health conditions, providing medications when needed or giving reminders about when patients need to take their medicine. They can also help lift and move patients and may serve as social companions. Carebots help human caregivers, alleviating significant manual or repetitive work. AI can also hold ground in places with a severe deficit of healthcare workers by taking over some diagnostic and care duties.

Carebots thus offer many opportunities for the health and care support of vulnerable groups like the elderly and people with physical and mental impairment. Despite the huge potential of AI applications in the healthcare sector, AI must be integrated into existing medical processes rather than substituting current practices. Doctors can achieve enhanced medical care by integrating AI insights into conventional medical practices that include both patient interaction and medical team discussions. This integrated approach can ensure encompassing diagnostic and healthcare processes, in a productive way that leverages both human and machine knowledge Healthcare Environment & Climate Change In today’s world, environment preservation and combating climate change are of main concern for nearly all governments or public institutions. AI can provide helpful tools to identify patterns and correlations in large amounts of environmental data, providing insights that would otherwise would not have been able to be detected by humans. For example, concerning how wild species are disappearing and with them, the biodiversity that supports life on earth, an AI-powered system called WildT rack1 is exploring 1 https://www.wildtrack.org the use of artificial intelligence in biodiversity preservation by analysing footprints the way indigenous trackers do and protecting these endangered animals from extinction. In this way AI can be used to monitor wildlife populations and their habitats. Additionally, it can track deforestation and desertification through analysis of satellite imagery or predict the occurrence of extreme climate-related events with higher accuracy. These applications can provide valuable information to government bodies to design required policy responses intended to cope with these challenges.

As our energy systems are the largest contributors to carbon emissions globally, it comes as no surprise that we attempt to leverage AI technologies to minimise the carbon emissions of the energy sector. The amount of electricity generated from renewable energy sources such as wind and solar has increased from around 30 000 GWh in the year 2000 to almost 2 500 000 GWh in 2020, averaging a yearly growth rate of almost 25%1. This means that progress is happening, but we are still far from a sustainable energy system. When thinking about the energy transition, the first images that come to mind are probably not of AI. However, AI is already very much present in our energy systems, helping to accelerate the transition towards a sustainable energy system. AI has brought improvements in many aspects of the energy transition, including forecasting of solar and wind energy, grid operation and optimisation, and coordination and management of distributed energy assets, like solar panels. Many countries already use AI systems to optimise the system’s performance and prevent power outages.

One of the main use cases for artificial intelligence in our current energy systems is within energy planning, demand and control. By analysing vast amounts of unstructured data on energy supply and consumption, AI 1 IEA (2022), Energy Statistics Data Browser, IEA, Paris https://www.iea.org/data-and-statistics/data-tools/energy-statistics-data-browser. technologies can enable more efficient operation of power companies. Identification of patterns of electricity demand and production results in more accurate predictions of future demand and necessary supply. These predictions can serve to shift flexible demand to reduce grid-strain, highlight optimal moments for increased levels of electricity storage or determine the optimal combination of electricity generators to reach the necessary supply2.

Another important applications of AI in energy systems are grid fault detection, data security for smart grid and forecasting of electricity prices. As the amount of data produced by our energy systems increases along with general computational power, AI will continue to become more embedded in and vital to energy systems. Future advances in AI in supporting the energy transition look promising, but its full potential is yet to be realised. Examples of such advances are the prevention of cyber attacks on energy systems, which is of course crucial to national security, the integration of electric vehicles and to smart grids to serve as storage capacity, and improved optimisation of energy yield2. 2 Ahmad, T., Zhang, D., Huang, C., Zhang, H., Dai, N., Song, Y ., & Chen, H. (2021). Artificial intelligence in sustainable energy industry: Status Quo, challenges and opportunities. Journal of Cleaner Production, 289, 125834.

Energy Systems

The use of AI has exploded in the automotive industry. Perhaps the most evident application of AI in the mobility sector is the advent of autonomous vehicles (or self-driving cars1). AI is currently being used in voice recognition, voice search, image recognition and detection in autonomous vehicles. An autonomous vehicle can drive itself from a starting point to a predetermined destination in “autopilot” mode. This is possible thanks to the various sensors installed in these vehicles. Autonomous vehicles are fitted with cameras, sensors, and communication systems to generate massive amounts of data. When such data is used to train AI systems, it enables the vehicle to sense the environment and make decisions. In all cases, autonomous vehicles still require a human driver to override the system when necessary and are not truly autonomous.

AI also supports the management of transport systems at a larger scale. AI improves overall mobility by analysing various types of data relevant to transport operations, like weather forecast, topology of the roads, and information about 1 https://medium.datadriveninvestor.com/artificial-intelligence-and-autonomous-vehicles-ae877feb6cd2 past incidents and traffic bottlenecks. For example, if a road accident happens or a road gets flooded due to heavy rain, the system can direct the flow of cars to less affected roads. Furthermore, this type of AI application is not restricted to terrestrial transportation. In fact, AI has already reached our skies. One example is the use of AI to predict the arrival times of long-haul flights to optimise ground handling and shorten passenger queues. The AI system used by the Singapore’s Changi Airport can make such predictions with almost 95% accuracy2. AI is also present in mobility systems in less sophisticated ways, for example, by providing recommendations of the fastest routes and counting the number of cars parked in a restricted area to prevent overcrowding. AI is also present in mobility systems in less sophisticated ways, for example, by providing recommendations of the fastest routes and counting the number of cars parked in a restricted area to prevent overcrowding. 2 https://cmp.smu.edu.sg/ami/article/20190522/ai-getsreal-singapores-changi-airport-part-1 Transportation Water Management Access to water is one of the most basic human needs for health and well-being. However, we are still far from universal coverage, and progress is slow. According to the United Nations, between 2015 and 2020, the population with access to safe drinking water services increased from 70 to 74%, and the population with access to handwashing facilities in the home increased from 67 to 71%. T o achieve universal coverage in 2030, the rate of progress would need to quadruple. This is a big challenge given that water demand is rising due to population growth, urbanisation and increased water use in agriculture, industry, and energy sectors. In addition, countries are facing growing challenges linked to water scarcity caused by climate change1.

The application of AI tools in the water sector contributes enormously to addressing multiple global challenges, in terms of water quality, management, and monitoring. For instance, acoustic sensors and gas tracers are being deployed to identify water leaks and detect contaminants in the water. These sensors and tracers are connected to AI systems to offer real-time monitoring of water quality, allowing for early detection of leaks and contamination, and thus preventing further damage. Furthermore, AI is increasingly 1 https://sdgs.un.org/goals/goal6 being used in agriculture to optimise water use. AI can be used to monitor soil conditions and estimate water demand. Combined with complex systems to predict weather conditions, AI will help develop ideal crop management strategies and streamline the use of water used in agriculture making it more efficient. This is extremely important given that 70% of freshwater use today is in agriculture and the efficiency of use is less than 50%2. 2 https://inc42.com/resources/how-ai-istransforming-the-water-sector/

Food Systems AI has also found applications in arts and culture. One example is the AI system called DALL·E 21, which creates images and art from a simple combination of words. Starting from a simple set of words like “an astronaut” , “riding a horse” , and “photorealistic style” , the system creates multiple realistic images that depict an astronaut riding a horse. By changing “riding a horse” to “playing basketball” , a new set of images is generated, now with an astronaut playing basketball instead of riding a horse. Then, by changing the style from “photorealistic style” to “style of Andy W arhol” , the system generates a new set of images of an astronaut riding a horse that resemble Andy W arhol’s famous creations. All of this in a manner of seconds. AI systems working in similar ways can also generate other cultural and art pieces, from songs and guitar tabs to fiction stories and essays. All artistic material generated by AI is called synthetic media and covers audio, text, and video.

Like in the cultural sector, the potential applications and opportunities for AI within education and academia are plentiful2. 1 https://openai.com/product/dall-e-2 2 Luckin, Rose; Holmes, Wayne; Griffiths, Mark and Forcier, Laurie B. (2016). Intelligence Unleashed: In the education sector, AI can assist teachers in various ways to provide a more efficient learning process. Examples of possible tools are personalised learning algorithms that can focus on areas where a particular student is struggling. A popular platform making use of such algorithms is Duolingo3, a company producing learning applications and providing language certification. Other examples are virtual tutors that can provide immediate feedback to individual students, or intelligent grading and assessments systems that can free up time for teachers to spend their time on higher level tasks like mentoring.

However, there are also serious drawbacks and threats to introducing AI into the classroom. AI can easily provide answers to students, stopping them from thinking critically and solving problems themselves. This is illustrated by the recent introduction of OpenAI’s large language model ChatGPT4, able to take over and perform many student assignments, like essays and reports. This aggravates increased dependence on technology, An argument for AI in Education. Pearson Education, London. 3 https://www.duolingo.com 4 https://openai.com/blog/chatgpt preventing students from thinking and learning independently. Additionally, increased use of AI-driven systems in education could reduce personal interaction and relationship-building, which is especially important for younger students in classrooms. New ways of teaching with percolating AI-based applications are urgently necessary. In academia, too, AI can support research activities. T raditional academic ways of analysing acadtext cannot keep up with the massive amounts of textual data generated at unprecedented rates. AI technology can screen vast amounts of textual data faster and more efficiently than humans. Recently, AI started to be used more extensively in the analysis published articles, supporting researchers and scholars in drawing connections between different fields of research and finding new insights and hypotheses. A concrete example is the use of AI to search databases of molecules and find the ones with desired properties for a specific application5. 5 https://www.iotforall.com/artificial-intelligence-in-research#:~:text=Researchers%20use%20 AI%2Dbased%20algorithms,would%20take%20 an%20expert%20scientist.

Culture, Education & Academia AI is turning out to be the secret ingredient for creating more efficient and sustainable food systems. According to the United Nations, the number of people going hungry and suffering from food insecurity has been gradually rising since 2014. The COVID-19 pandemic has led to a spike in this trend. In 2020, between 720 and 811 million persons worldwide were suffering from hunger - as many as 161 million more than in 2019. In the same year, over 30% of the world’s population was moderately or severely food-insecure, lacking regular access to adequate food. The war in Ukraine has since aggravated this scenario, triggering food shortages that affect the most vulnerable. AI-driven systems are already being deployed in various ways across the agricultural industry. The combination of AI systems with remote sensors, such as satellite imagery and drones, can help in tackling challenges such as pest management, nutrient deficiency, and irrigation issues. Other applications include the optimisation of irrigation schedules and the prediction of crop yields1. The application of AI to solve food challenges goes beyond the agricultural sector. AI is being used in food processing to analyse the quality of food and determine whether it is suitable for consumption, thereby reducing food waste. 1 R. Sharma, "Artificial Intelligence in Agriculture: A Review," 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), Madurai, India, 2021, pp. 937- 942, doi: 10.1109/ICICCS51141.2021.9432187.