Lecture AI Economics Ethics Ludovic PDF

Understanding AI in Business Contexts Ludovic Dibiaggio KTO & SCAI Intelligence Artificielle (IA) Technology is an autonomous force that determines society (e.g. Martin Heidegger , 1962), Jacques Ellul (1964), Habermas (1968) Intelligence...

Understanding AI in Business Contexts Ludovic Dibiaggio KTO & SCAI Intelligence Artificielle (IA) Technology is an autonomous force that determines society (e.g. Martin Heidegger , 1962), Jacques Ellul (1964), Habermas (1968) Intelligence Asimov’s (1950) three laws of robotics Artificielle (IA) 1. A robot may not injure a human being or, through inaction, allow a human being to come to harm. 2. A robot must obey the orders given by human beings except where such orders would conflict with the First Law. 3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. Technologies are shaped by human interests and values, and open to human choice (e.g; Latour & Woolgar, 1979; Winner, 1980) Guide the development of AI in line with a range of values (e.g. High-Level Expert Group on AI, 2019) based on responsible AI, Humane AI, explainable AI, and meaningful human control (e.g. , Floridi et al., 2020) Intelligence Artificielle (IA) Technology and society co-evolve; some aspects are hard to change or unmalleable, but technology creates novelty and unexpected (and unintended) consequences. Artificial intelligence (AI) Définitions Darmouth Proposal, August 1955 ✓ “ The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can made to simulate it.” ✓ “…how to make a machine […] solve kinds of problems now reserved for humans, and improve thelsemves.” John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. Intelligence Artificielle (IA) ‘Good old-fashioned artificial intelligence’ ‘Artificial neural networks’ (GOFAI) “Within a generation, the problem of creating "artificial intelligence" would be substantially solved.” Marvin Minsky, 1967  2012 Hinton , Srivastava, Krizhevsky, Sutskever and Salakhutdinov Intelligence AlexNet’s decisive and unexpected victory in ImageNet’s annual competition to automatically Artificielle (IA) recognise objects and scenes from natural images  Deep Learning Revolution The promises of Artificial Intelligence « The AI moment » Introduction of deep learning Reduction in error rate during the annual ImageNet competitions from 2011 to 2017. Source: Electronic Frontier Foundation  The promises of Artificial Intelligence Evolution of the number of AI scientific publications Deep Learning revolution « The AI moment » 2012 Source: The AI Index Report (Stanford)- SCOPUS, Elsevier, 2020 Artificial intelligence (AI) Who are these guys? Artificial intelligence (AI) Deep learning 2018 ACM Turing Award Laureates the “Nobel Prize of the Computer Industry” « revolution » ✓ The trio were honored “for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing.” ZDNet ✓ They developed the foundations of deep learning. Geoffrey Hinton of the Yann LeCun of Meta Yoshua Bengio of University of Toronto and Canada's MILA institute vice president and engineering fellow at Google Artificial intelligence (AI) “Take any old problem where you have to predict Définitions something and you have a lot of data, and deep learning is probably going to make it work better than the existing techniques.” Geoffrey Hinton (2016) ✓ The theory and development of computer systems able to perform tasks normally requiring human intelligence (Oxford English Dictionary) ✓ “A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments” (OECD, 2019). Technical systems that interact with their environment to improve their capacity to autonomously solve problems or answer (pre-defined) questions Intelligence Artificielle (IA) An economists’ definition of AI ✓ Systems that increase the power and reduce the cost of prediction (Agrawal et al., 2018). Autonomous predictive artificial system that solve problems based on existing data. Intelligence Artificielle (IA) Autonomy… what are we talking about? The level of autonomy depends on how open is the pre-defined question/problem ✓ Autonomous action (given predefined goals and deliberation rules i.e., how should the system respond to different contingencies) ✓ Autonomous deliberation/decision (given predefined goals) ✓ Autonomous goal-making (given predefined values/rules ) ✓ Autonomous identity – or self organizing systems (definition of values given specific practices and contexts) – Requires a theory of mind. Intelligence Artificielle (IA) Weak AI Can learn and autonomously perform tasks in a predefined context Strong AI Can autonomously learn and acquire new abilities to handle situations and perform tasks in contexts different from those initially intended Superior AI Surpasses human capacities to achieve a point of singularity (Raymond Kurzweil) What about General “It’s pretty apparent to anyone who’s paying attention that generative AI opens the door to computerization of a lot of kinds purpose AI (e.g. of tasks that we think of as not easily automated,” (David Autor) Generative AI)? Expected effects of AI on economic growth ✓ 2017 PwC report, predicts that AI’s contribution to the global economy will be $15.7 trillion in 2030 (more than the current output of China and India combined). ✓ The amount of compute (computing power) used to train cutting-edge AI systems has been doubling every six months over the past decade (Sevilla et al., 2022) ✓ Generative AI has broad applications that will impact a wide range of workers, occupations, and activities. LLMs could affect 80% of the US workforce in some form (Eloundou et al., 2023) ✓ Software engineers can code up to twice as fast using a tool called Codex, based on the previous version of the large language model GPT-3. (Kalliamvakou, 2022) ✓ Many writing tasks can also be completed twice as fast (Noy and Zhang, 2023) AI and economic growth: Labor productivity growth trends the productivity paradox So… where are productivity gains?  Growth in real GDP per capita, 1300-2100 Actual and hypothetical path (Gordon 2012) Why does it matter? Année Gordon (2012) Secular stagnation A pessimistic view Gordon (2012) contends that the growth potential of digital technologies is already exhausted Gordon (2012) Source: UN/DESA estimates, based on Maddison Project. General purpose technology (GPT) Reduce the cost Reduce the cost Reduce the cost of energy of computation of prediction Definition ✓ Pervasive: widely used with a cumulative economic impact The drop in cost of prediction will (…) increase the value of complements (data, judgement, and action) and ✓ Cost reduction reduce the value of substitutes (human prediction) ✓ Capable of ongoing quality improvement (Agrawal, Gans and Goldfarb, 2018) ✓ Enable complementary innovations in application sectors => many follow up inventions Examples of GPT: printing, bronze, the waterwheel, steam power, electricity, the internal combustion engine, railways, motor vehicles, lasers, the Internet (Lipsey et al., 1998)  The example of the electricity Lessons from history GPT as a source of industrial revolutions Electrification of the economy Cost reduction of energy Substitution effect: lowers relative costs and increases the propensity to adopt the new technology Complementarity effect: lowers prices and increases the value of complementary resources/products AI and economic growth AI as a GPT (Key enabling Technology) New techniques Reduction of prediction costs AI and economic growth AI as a GPT (Key enabling Technology) New techniques Functions AI and economic growth AI as a GPT (Key enabling Technology) New techniques New functions Applications So… From Steam power to electricty where are productivity gains? Adoption takes time!  Adoption and labor productivity growth of electricity compared to IT and labor productivity growth Adoption & diffusion matter! AI is the continuation (the achievement?) of the digital revolution ChatGPT effect The boost of open (source) AI Strategic challenges « The imitation game »  Patent production by country USA Strategic China challenges UE 27 Japan South Korea International competition Germany UK France Canada Taïwan Netherlands Sources : (1) PATSTAT édition 2022. (2) OECD.AI (2021), powered by EC/OECD (2021), database of national AI policies, accessed on 11/02/2023, https://oecd.ai , authors’ calulation AI Technologies and key players (SKEMA, 2022)  VC investment Strategic Scientific publications AI Patents (cumulative challenges Academic journals -1 sum $ million) -2 (cumulative sum 2022) (2) Countries Public Private International competition USA 183 362 39 385 540 382 361 080 China 279 429 19 206 93 518 197 870 Japan 31 576 5 711 133 004 7 427 South Korea 29 538 33874 96 020 8 326 Germany 39 769 3 857 63 496 16 845 UK 42 938 3 317 32 458 25 029 France 26 063 4 767 21 910 9 820 Canada 26 861 2 029 21 250 12 425 Sources : (1) PATSTAT édition 2022. (2) OECD.AI (2021), powered by EC/OECD (2021), database of national AI policies, accessed on 11/02/2023, https://oecd.ai  Patent production by country, per million inhabitants Strategic challenges Country participation in innovation incorporating AI Source: AI Technologies and key players (SKEMA, 2022)  Strength ( ) and Weaknesses ( ) Standards and regulation USA China Europe Others Algorithms AI is an Google Meta Tencent Baidu, Alibaba, ecosystem Microsoft, Universities Xiaomi Data Google Meta Tencent Baidu, Alibaba, Microsoft Amazon etc. Xiaomi International competition Cloud platforms China : Three-Year Action Plan (2017): $150 Amazon AWS, billion, after their «Spoutnick moment» Alibaba, Hawei, Google, Microsoft Tencent, Baidu, Restrictive standards Azure US : Chips and Science Act (2022): $280 billion Super computers Japan Alliances w/ taïwanese & South Global HPC leadership depends on staying at the Push Korean Cies to invest and cutting edge of both HPC systems development as produce in the US well as their application and use. EU : Digital sovereignty plan: €42 billion Hardware RGPD, Data governance Act, Data act, Taïwan IA Act Manufacturing (TSMC, Design (NVIDIA, AMD UMC, Vanguard, and S. Korea et INTEL) Powerchip)  Total estimated investments in AI start-ups ($ billion), 2011-2017 Strategic challenges International polarization Number of AI Private Investment Country Startups (2013-2022) (2013-2022) #1 United States 4643 $249B #2 China 1337 $95B #3 United Kingdom 630 $18B https://sciencebusiness.net/news-byte/us-and-china-lead-investments-artificial-intelligence-start-ups #4 Israel 402 $11B #5 Canada 341 $9B #6 France 338 $7B #7 India 296 $8B #8 Japan 294 $4B #9 Germany 245 $7B #10 Singapore 165 $5B https://www.visualcapitalist.com/sp/global-ai-investment What about jobs? GoldmanSachs' cash equities trading floor at the firm's headquarters in 2000 – Head Count 600 GoldmanSachs' cash equities trading floor at the firm's headquarters in 2008 – Head Count 2 "Automated trading programs have taken over the rest of the work, supported by 200 computer engineers.“ Marty Chavez Deputy Chief Financial Office AI and employment The Horse Parable Leontieff (1982) Horses became unnecessary with the advent of tractors, automobiles and trucks. “After all, this doesn’t precipitate a political problem, since horses don’t vote” AI and employment The Dog and Man Parable Warren Bennis “The factory of the future will have only two employees: a man and a dog The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment” Average productivity extremely high, marginal productivity of labor = 0 Wage = 0 AI and employment No job destruction (so far) AI and employment Source. Bureau of Economic Analysis. Report prepared for the G20 Employment Working Group Antalya, Turkey, 26-27 February 2015 Labor share has constantly decreased with automation in the USA since 2000s Source: Daron Acemoglu & Simon Johnson Annual WIDER Lecture, October 2022  Changes in the labor share and in income inequality in OECD countries, 1990s to mid-2000s AI and employment Labor share has an impact on income inequalities  Change in Task Content of Production, 1947–1987 AI and employment Displacement effect: Automation changes the task content of production adversely for labor because capital destroyed jobs. Some of the tasks performed by white-collar workers in accounting, sales, logistics, trading, and some managerial occupations are being replaced by specialized software and artificial intelligence. Reinstatement effect: Automation changes the task content of production and reinstates labor into a broader range of tasks that are labor intensive => increase in the labor share as well as in labor demand. Change in task content = displacement + reinstatement. AI and employment The substitution effect captures the substitution between labor- and capital- intensive tasks within an industry The composition effect arises from the reallocation of activity across sectors with different labor intensities. Productivity effect: Automation technology also increases productivity, and contributes to the demand for labor in non-automated tasks “Technological ‘advances’ might always help labor -- what we call a productivity Bandwagon.” (Acemoglu and Johnson, 2022). The deceleration of labor demand growth over the last 30 years is due to a combination of anemic productivity growth and low reinstatement effect Acemoglu et al. (2022)  Automation reduces the labor share and may also reduce the (average) wage and/or employment, especially when productivity gains from automation are small. In brief…. Automation may also be excessive from a welfare point of view due to distributional concerns. If AI is used predominantly for automation, it will have similar effects to other automation technologies, and depending on its productivity effects and relevant welfare criteria, it may have a negative impact on social welfare. Too much Automation? The direction of technology is partly shaped by the business models of leading firms and the aspirations of researchers who favor automation (Acemoglu, 2021) A related argument is that US corporations may have become too focused on cost-cutting, which might also encourage excessive automation. Tax codes impose a much higher marginal tax rate on labor than on equipment and software capital, thus favoring automation (Acemoglu et al. (2020)  Job loss and gained per sector Estimation of job creation AI and employment In detail Estimation of job loss Source: PwC analysis of OECD PIAAC and ONS APS data  Job loss and gained per annual revenue AI and employment Effects of automation on employment (Estimation) Effect of automation on net job creation/destruction Médiane Median desrevenue of annual gains (£) annuels (£) Source: PwC analysis of OECD PIAAC and ONS APS data AI and wages In consequence of better machinery, of greater dexterity, and of a more proper division and distribution of work, all of which are the natural effects of improvement, a much smaller quantity of labour becomes requisite for executing any particular piece of work, and though in consequence of the ourishing circumstances of the society, the real price of labor should rise very considerably... " Adam Smith Polarization of wages The richest get richer Increasing disparity Polarization of wages Education and increasing disparity Daron Acemoglu & Simon Johnson Annual WIDER Lecture, October 2022  OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic Time Polarization Billy Perrigo January 18, 2023 of wages Reinstatement…? AI and ethics Adverse examples  “The instructions given to the AI and the fine-tuning of its parameters could transform an apparently harmless system AI and ethics into one focused on malicious intent, as well as lower the level of technical skills required to do so. Moreover, such a transformation could be carried out at low cost and with minimal data.” “it is quite challenging to digest such reflections and make the necessary mindset shift they imply. It’s a lot to take in, especially when considering the potential implications for the future” “We do not yet know how to make an AI agent controllable and thus ensure the safety of humanity! And yet, we are making great strides – myself included until now – towards building such systems”. Yoshua Bengio, 2023  What is morally good and bad and morally right and wrong (Cambridge dictionary). AI and ethics Ethics consists of the fundamental issues of practical decision making Definition Its major concerns include the nature of and the standards by How should we live? Shall we aim at which human actions can be judged right or wrong. happiness or at knowledge, virtue, or the creation of beautiful objects? If we choose Ethics is based on ultimate value and well-founded standards happiness, will it be our own or the of right and wrong that prescribe what humans ought to do happiness of all? (Britannica) (Velasquez et al., 2010). Ethics is not related to a given Not a field of study, or a branch of inquiry anymore…? morality but take morality as a Guidelines build on pre-defined values. subject of inquiry Are value homogenous across cultures/time? AI and ethics Two levels of ethical concerns Safety Laws, rules, principles recommendations Level 1: Justice – Fairness – Human rights Security Privacy Level 2: Values and existential issues Unconscious evolution of values Public debate and political decisions AI and ethics Data Main source of biases: biases already in dataset Discrimination (biases): if society starts trusting AI algorithms, their discriminatory choices may come to be accepted as more justifiable than when they were made by individual decision-makers  Privacy: when an individual shares her data, she indirectly reveals information about others. AI and ethics Data Data prices will be depressed and will not reffect users’ true values of data and privacy. Excessive data shift surplus from users to platforms and companies. Competition distortion Better collection and processing of data for prediction by one firm provide a competitive advantage (more data => better prediction => more users => more data) Behavioral manipulation AI captures users’ preferences and characteristics and hence shapes their goals and nudge their behavior, decisions and identity to an extent that may undermine their self-determination. => shift surplus from consumers to platforms and distort the composition of consumption AI and ethics AI reduces learning capacities: substituting judgment by AI in automated tasks may limit judgement capacities of human for other tasks. Excessive monitoring Reduces workers’ freedom in companies and threat their bargaining power May affect freedom and may weaken democracy Community effect: AI-powered social media favor community effect (I only talk with people like me) AI and ethics Correlation is not causation! Limits? AI and ethics Correlation is not causation! Limits? AI and ethics Correlation is not causation! Consequences? AI and ethics The importance of choices, both on the use of existent AI technologies and on the direction of AI research The inadequacy of market solutions that mainly rely on increasing competition The need for regulation. AI and ethics Interesting questions Can ethical AI be translated into rules, standards and guidelines? The definition of AI and related ethical concerns are contingent on the “level” of autonomy (what is automated?) ✓ Autonomous action (given predefined goals and deliberations: i.e. how should the system respond to different contingencies) ✓ Autonomous deliberation/decision (given predefined goals) ✓ Autonomous goal-making (given predefined values/rules/rules ) ✓ Autonomous identity (definition of values given specific practices and contexts) Think of AI-powered weapons: who should decide a lethal shot? AI and ethics Ethical governance ✓ Distributed agency (delegation and responsibility) AI distributes moral responsibility among designers, regulators, and users. Delegation may lead to unintended consequences, and while excluding or precluding Human supervision. ✓ Many principles, guidelines, standards, regulations produced by national and international bodies. There are many confusing or incompatible prescriptions that makes their application difficult. ✓ The evolution of AI systems and their interaction with other technologies as well as economic and social systems make the application of rule-based ethics difficult if not counter-productive. How can we design an ethical governance system that is not only a constraint, but also a guide that is part of the innovation process, the support for a discussion between values, principles and technical or strategic effectiveness in the context of the application? AI Act Human should keep control over AI systems General Data Protection Regulation (GDPR) Digital Services Act (DSA) Digital Markets Act (DMA) Data Governance Act (DGA Data act (DA) AI Act (2025) Three Guidelines Lawful – comply with all applicable laws and regulations Ethical - respecting ethical principles and values Robust - both from technical and social perspectives since AI systems can cause unintentional harm Seven key requirements Human agency and oversight (including fundamental rights, human agency and human oversight) Technical robustness and safety (including resilience to attack and security, accuracy, reliability and reproducibility) Privacy and data governance (including respect for privacy, quality and integrity of data, and access to data) Transparency (including traceability, explainability and communication) Diversity, non-discrimination and fairness (including the avoidance of unfair bias, accessibility and universal design, and stakeholder participation) Societal and environmental wellbeing (including sustainability, social impact, and democracy) Accountability (including auditability, minimisation and reporting of negative impact, trade-offs and redress). AI act Source : https://www.civilsdaily.com/news/eus-artificial-intelligence-ai-act/  Explainable AI Explainable AI is essential to simplify the algorithm and reduce the number of factors. AI and ethics However, how open-source/low code AI can be explainable? Ethics Guidelines for trustworthy Artificial Intelligence (Kaur, 2022) Different levels of human involvement in making AI systems trustworthy  Data Visualization Data Standardization Feature Importance Interpretable Model Example based Rule based Vizualization based Interpretable Model Evaluation Methods Different levels of human involvement and different control points (Kaur, 2022)  If Blaise Pascal was asked to tell something about AI, « he would say that what matters is not technology, all that matters is the heart » Cédric Villani, Sept. 20, 2023 interview, Les Grands Themas de CIO, Le Monde Informatique, (https://themas.lemondeinformatique.fr)