Conceptual map for artificial intelligence policy discourse analysis

Ðàçìåùåíî íà http://www.allbest.ru/

FEDERAL STATE AUTONOMOUS EDUCATIONAL

INSTITUTION FOR HIGHER EDUCATION

NATIONAL RESEARCH UNIVERSITY

HIGHER SCHOOL OF ECONOMICS

Faculty of Social Sciences

Master Thesis

A Conceptual Map for AI Policy Discourse Analysis

Master's program `Political Analysis and Public Policy'

You Li

Moscow 2019

Table of Content

Introduction

1. AI Research and Development

2. Economic Impacts, Labor Shifts and unemployment, and Inequality

3. Accountability, Transparency, and Explainability

4. Surveillance, Privacy and Social Risks

5. Fairness, Ethics, and Human Right

6. Natural Language Processing (NPL)

7. Human Health

Bibliography

Introduction

Artificial intelligence (AI) will become one of the most important technologies in coming decades. From its traditional province of science fiction, AI is entering the reality Cath, C., Wachter, S., Mittelstadt, B., Taddeo, M. and Floridi, L., 2018. Artificial intelligence and the `good society': the US, EU, and UK approach. Science and engineering ethics, 24(2), pp.505-528.. Applications of AI in driverless car, online advertisement, search engine, health care, and crime prevention Boudreau, A., Francis, Enjoli. 2011. Santa Cruz Police Using Computer Program to Predict, Prevent Crime. ABC News. suggest that the use of AI is become increasingly ubiquitous. The rapid development of AI is brought by better algorithms, increasing computing power, and a larger quantity of available data Presser, Gero. 2017 What Everyone Should Know About Machine Learning. Dzone. , and the potentials they may unlock can yet be comprehensively assessed. Scholars like Dirican believes that AI economy is the first stage of “Space Economy” that may generate a new “economic cycle, Dirican, C., 2015. The impacts of robotics, artificial intelligence on business and economics. Procedia-Social and Behavioral Sciences, 195, pp.564-573.” and the applications of AI in economy can substantially increase economic prosperity. Meanwhile, other scholars are emphasizing on the risks related to AI such as massive unemployment due to expanding uses of automated robots “As technology improves, robotic automation projects are likely to lead to some job losses in the future, particularly in the offshore business-process outsourcing industry.”

Davenport, T.H. and Ronanki, R., 2018. Artificial intelligence for the real world. Harvard business review, 96(1), pp.108-116.. Moreover, politics, wealth, power and other elements of human society will be drastically changed with the emergence of General AI (GAI), or superintelligence Dafoe, A., 2018. AI Governance: A Research Agenda. Governance of AI Program, Future of Humanity Institute, University of Oxford: Oxford, UK.. Allan Dafoe, a scholar at the Future of Humanity Institute of Oxford, believes that AI has the potential to become a general-purpose technology (GPT) similar to important discoveries in human history such as bronze, steam engine, and the internet Dafoe, A., 2018. AI Governance: A Research Agenda. Governance of AI Program, Future of Humanity Institute, University of Oxford: Oxford, UK..

These days, AI policy is receiving increasing attention from media, civil society, and decision makers. Apart from the rapid development of AI, studies on AI policy are far from sufficient. In August 27th 2018, Dafoe published a research agenda for AI governance in which he writes:

“Research is thus urgently needed on the AI governance problem: the problem of devising global norms, policies, and institutions to best ensure the beneficial development and use of advanced AI Dafoe, A., 2018. AI Governance: A Research Agenda. Governance of AI Program, Future of Humanity Institute, University of Oxford: Oxford, UK..”

As of current situation, AI policy consists of different theories taken from other areas - it is interdisciplinary by nature; however, the theoretical connection between different branches of AI policy - ranging from computer science and neuro network to economy - remains nearly absent. Thus, developing a theoretical framework is both urgent and necessary.

For addressing this situation, the main aim of this paper is to provide a “prototyped” theoretical framework - a conceptual map - for connecting these scattered “dots” in AI policy.

The paper consists of two parts. The first part features the research question, research aims, an introduction of AI and an overview of the recent history regarding the development of national AI policy. After that, a brief literature review on national AI policy is conducted, followed by introduction to methodologies used in this research. For developing the conceptual map, the paper uses grounded theory and key point coding method for drawing a conceptual map where branches of AI policy, or in this case “concepts” are situated on the map, and in relations with each other. Further details on are further elaborated and with a demonstration of the method used for collection. Furthermore, important concepts situated on the map are discussed in group highlighting their connectivity.

The second part of the paper emphasizes on using the conceptual map to encompass discourse analyses on three case-specific settings for highlighting the practicality of the conceptual map. In the first case, dominant discourses are identifiedFurther details canconceptual map validity of the conceptual map is tested with the use of discourse analysis and strengthened by existing theory. Sixth, two exercises of operationalizing the conceptual map is conducted with the help of discourse analysis. Finally, a summary of key findings is presented followed by a discussion.

Research Question

On 12th of February, 2019, a report titled “China Is Starting to Edge Out The US In AI Investment” caught the attention of the author. The report states that:

“The Chinese government is promoting a futuristic artificial intelligence plan that encompasses everything from smart agriculture and intelligent logistics to military applications and new employment opportunities growing out of AI.”

The author found it difficult, and asked himself a question that led to this paper - “how can we interpret a discourse about national AI policy?”

“AI is a broad field that involves extremely disparate disciplines. It's difficult to have state-of-the art expertise in all of those aspects in one place - Carnegie Mellon University”?

As the author began his attempt to research, a great deal of difficulties arose since similar discourse analyses have not been done. AI policy exhibits an interdisciplinary nature, and different branches of AI policy are often borrowed from a wide range of disciplines. Moreover, these branches of AI policy lack theoretical connectivity. Thus, amidst dampen research spirits, these difficulties led to the following research question “What are the theories can we use for drafting hypotheses suitable for discourse analyses of AI policy?”

Aims of the Research

There are four aims of the research:

1. Developing the “prototyped theoretical framework” that highlights the relationships between branches of AI policy for encompassing discourse analyses of national AI policy.

2. Explicating some important concepts of AI policy.

3. Operationalize the for developing hypotheses, and subsequent discourse analyses are conducted for primary testing of the hypotheses.

4. Additional insights are given in the discussion section.

Literature Review

Although literatures on discourse analysis of AI policy is absent, many comparative studies of national AI policy can be found. In academia, Corinne Cath and Sandra Wachter et al. conduct a comparative study of the national AI policies of US, EU, and UK centered around the concept of “good society Cath, C., Wachter, S., Mittelstadt, B., Taddeo, M. and Floridi, L., 2018. Artificial intelligence and the `good society': the US, EU, and UK approach. Science and engineering ethics, 24(2), pp.505-528.. Think tanks reports such as “Building an AI Word - Report on National and Regional AI Strategies” conducted by CIFAR highlights the priorities of 18 countries that have, or in the process implementing their AI policies.

Many academic articles also discuss the implications of AI in public policy. For example, Ajay Agrawal, Joshua Gans, and Avi Goldfarb consider advances in machine learning as quality-adjusted drop in the price of prediction. Regarding the price drop, they argue that the use of intellectual property policy can influence the diffusion of AI, and the use of labor and antitrust policies can influence employment, inequality, and competition in an AI era Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. 2019. Economic Policy for Artificial Intelligence. Innovation Policy and the Economy 19, no. 1 .. Regarding policy making process, Goolsbee raises the question of whether or not AI will eventually takeover policy making Goolsbee, Austan. 2018. Public policy in an AI economy. No. w24653. National Bureau of Economic Research, , to which he argues that AI may improve our ability to predict responses, but it does not contribute to balancing interests or engaging in politics Goolsbee, Austan. 2018. Public policy in an AI economy. No. w24653. National Bureau of Economic Research,. Moreover, regarding General Artificial Intelligence (GAI) or superintelligence, Nick Bostrom et al. propose several approaches that could be used to govern GAT.

In sum, the volume of existing literatures about AI policy remain low; however, there are more literatures on various branches of AI policy. As the second part of the literature review, these theories, concepts, and approaches will be addressed in the section titled “important concept.”

What is Artificial Intelligence?

The definition of AI evolves as it develops and expand. According to Stuart Russel and Peter Norvig, historically, scholars attempted to define AI based on four approaches - thinking humanly, acting humanly, thinking rationally and acting rationally Russell, Stuart J., and Peter Norvig. 2016. Artificial intelligence: a modern approach. Malaysia; Pearson Education Limited, . With the development of neural science and cognitive science, scientists are able to measure and model how humans think. Combined with AI, the implications and potentials of technological advances could be tremendous. Richard Bellman define `thinking humanly' as the automation of activities that people commonly associate with human thinking including problem solving, decision-making, and learning Bellman, Richard. 1978. An introduction to artificial intelligence: Can computers think? Thomson Course Technology.. Acting humanly means AI could act like human. A classic example of this is the Turing test developed by Alan Turing - it is a test for measuring the ability of machine to show intelligent behavior indistinguishable from a human Turing, Alan M. 2009. Computing machinery and intelligence. In Parsing the Turing Test, pp. 23-65. Springer, Dordrecht.. Kurzweil defines `acting humanly' as “the art of creating machines that perform functions that require intelligence when performed by people Kurzweil, Ray et al. 1990. The age of intelligent machines. Vol. 579. Cambridge, MA: MIT press, . Artificial intelligence is also unique due to its embedded rationality. `Thinking rationally' means that it uses logic for processing information, and for making conclusions and decisions. Winston states that AI science is the study of the computations that make it possible for AI to perceive, reason, and act Winston, P. H. 1992. Artificial Intelligence (Third edition). Addison-Wesley.. `Acting rationally' means that AI is capable of performing actions that create betterment of the state of the AI or the environment in which AI is performing these actions. For instance, Alpha Go is designed act rationally for a specific end goal - to win the game. Overtime, the four categories of definitions of AI became well-received in academic community, and it appears to be the most widely-used definition of AI nowadays Scherer, Matthew U. 2015, Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies. Harv. JL & Tech. 29.

Moreover, there also needs to be a distinction drawn between task-specific AI (weak AI), and general AI (strong AI) as they believe that weak AI means that the hypothesis that machines have the potential to behave intelligently, and strong AI means that machines would be counted as having consciousness, or `actual minds Russell, Stuart J., and Peter Norvig. 2016. Artificial intelligence: a modern approach. Malaysia; Pearson Education Limited, .' Drawing on their definition, this paper uses the term `AI' to refer to task-specific AI technologies and related technologies.

An Overview of National AI Policy

With tremendous opportunities and challenges that AI presents, Governments are preparing for a world where AI has the ability to transform the economy, security, society, and governance. Countries with advance research, investment, and policy formation capabilities like China, US, and the European Union scrambled to develop their AI policies. In October 10th, 2016, Obama administration published a government paper titled The National Artificial Intelligence Research and Development Strategic Plan that marked the first government-sponsored strategic plan regarding AI R&D National Science and Technology Council Committee on Technology., 2016. The National Artificial Intelligence Research and Development Strategic Plan. Washington D.C: Executive Office of the President.. Another report - Preparing for the Future of Artificial Intelligence - was a public policy paper that surveys the development of AI, discusses opportunities and policy challenges, and gives policy recommendations for further actions National Science and Technology Council Committee on Technology., 2016. Preparing for the Future of Artificial Intelligence. Washington D.C: Executive Office of the President.. In 2018, the Defense Advanced Research Projects Agency (DARPA) announced its investment plan in AI that exceeded 200 million dollars. It was followed by the AI Next campaign that focuses on the technical matters. The campaign includes automating Department of Defense (DoD) processes, improving the reliability and robustness of AI, improving “explainablity” of AI, pioneering AI algorithms, and reducing inefficienciesThe Defense Advanced Research Projects Agency. 2018. AI Next Campaign. Retrieved from https://www.darpa.mil/work-with-us/ai-next-campaign . Under Trump administration, the House Committee on Oversight and Government Reform held a series of hearings on AI, after which the subcommittee on Information Technology published a report with the collective efforts of members of the House, experts, scholars, and the industry. The report with the title Rise of the Machines: Artificial Intelligence and Its Growing Impact on U.S. Policy presents an important lesson from past lessons and brings forth updated policy guidance Hurd, W. and Kelly, R. 2018., Rise of the Machines: Artificial Intelligence and Its Growing Impact on U.S. Policy. United States. Committee on Oversight and Government Reform.

Whilst the US was the first among others to respond to the development of AI and develop their national AI policies, members in the EU, India and China were late in comparison. Since 2014, the Chinese government has announced a series of state-sponsored strategic plans concerning high-tech and new industry for establishing an AI market with a value over 14 billion $ by 2018 and securing China's position as a leading figure in AI by 2030. In 2015, the State Council of China published its Internet plus (»¥ÁªÍø+) strategy in which intelligent manufacture (ÖÇÄÜÖÆÔì), intelligent resource production (ÄÜÔ´Éú²úÖÇÄÜ»¯), intelligent logistics, delivery, and distribution system (ÖÇÄÜÎïÁ÷ÅäË͵÷ÅäÌåϵ), the development of a new AI industry (ÅàÓý·¢Õ¹È˹¤ÖÇÄÜÐÂÐ˲úÒµ), the innovation and development of AI products in important areas (ÍƽøÖصãÁìÓòÖÇÄܲúÆ·´´ÐÂ) and improving intelligent level of end-products (ÌáÉýÖն˲úÆ·ÖÇÄÜ»¯Ë®Æ½) are discussed The State Council of People's Republic of China. 2015. State Council's guiding opinion on developing actions on “Internet plus” (¹úÎñÔº¹ØÓÚ»ý¼«Íƽø“»¥ÁªÍø+”Ðж¯µÄÖ¸µ¼Òâ¼û) . None of the aforementioned documents are explicitly about AI, but it is apparent that the Chinese government had demonstrated incentive for the further advancement in articulating AI policy.

In 2017, two documents: “New Generation of Artificial Intelligence Development Plan (ÐÂÒ»´úÈ˹¤ÖÇÄÜ·¢Õ¹¹æ»®)” and “Three-Year Action Plan for Promoting Development of a New Generation Artificial Intelligence Industry (2018-2020) (´Ù½øÐÂÒ»´úÈ˹¤ÖÇÄܲúÒµ·¢Õ¹ÈýÄêÐж¯¼Æ»®(2018-2020Äê)” were published by the Chinese government. They are the first policy papers solely dedicated for the development of AI. They conclude that AI will become the new focal point of international competition because it is a technology that guides the future “The State Council of People's Republic of China. 2017. State Council's Announcement on publishing Development Plan for a New Generation of Artificial intelligence(¹úÎñÔº¹ØÓÚÓ¡·¢ÐÂÒ»´úÈ˹¤ÖÇÄÜ·¢Õ¹¹æ»®µÄ֪ͨ). Firstly, the papers summarize the current situation regarding the development of AI with assessments of challenges and opportunities it presents. Secondly, they provide fundamental principles regarding AI development in China and defines the roles that the Chinese Communist Party (CCP) and the government should play in the process. Thirdly, they emphasize on important tasks for AI development including AI innovation, facilitating AI industry and AI economy, healthcare, elderly care, security, and a series of state-funded programs labeled “1+N.”

Two years after the U.S published its government reports on AI, in 2018, member states of the European Union signed a 20 pages declaration called Communication Artificial Intelligence for Europe for promoting AI investment in the EU, preparing for societal transformation, and constructing ethical and legal framework Communication Artificial Intelligence for Europe., 2018, European Commission. Brussels. . In June 2018, the EU purposed the Digital Europe Program with a budget of 920 million euro for further developing AI and implement AI on a societal scale. Member states in the European Union have also developed and published their own AI policies. For instance, As the second largest economy, in 2018, France published a 147 pages “Villani Report” titled, “For a Meaningful Artificial Intelligence,” with 3 focuses. Firstly, it envisages a strategic plan for R&D, education, and innovation policy in the AI domain. Secondly, the reports also discuss regulations and implementation. Most importantly, the Mission Villani report stresses the importance of attracting AI experts and preventing brain-drain Villan, C. 2018 For a Meaningful Artificial Intelligence - Towards a French and European Strategy. French Parliament. Paris.. Soon after the report, “AI for Humanity - French Strategy for Artificial Intelligence” is published echoing most of the key points made by the report. Germany, as the largest economy in the EU, announced that it would devote 3 billion euro for becoming an AI powerhouse. In 2018, Germany published a policy paper titled “Key points for a Federal Government Strategy on Artificial Intelligence” in which goals, fields of actions, and policy recommendations are presented. Moreover, a more comprehensive report will be produced with the help of experts, associations, organizations, and institutions operating in Germany.

India is also amongst the late-comers in the game. In 2018, a working paper titled “National Strategy for Artificial Intelligence #AIforAll” is published by an Indian think tank defining the national policy on AI in India. It highlights growth and greater inclusions in areas such as healthcare, education, agriculture, and transportation. It also addresses five challenges facing India in the adoption of AI. They include absence of enabling data ecosystems, high resource cost, and lack of expertise and talents.

Methodologies

This paper uses two methodologies. Firstly, it uses grounded theory for developing an inductive theoretical framework. This approach is useful for achieving the purpose of the essay - to provide a conceptual map that can be used for conducting discourse analyses on national AI policy, which has been by far a new field in public policy. Secondly, the conceptual map is used for encompassing the subsequent discourse analyses are for testing conceptual validity, and for demonstrating operationalization.

For this research, the general corpus means the data used by either grounded theory analysis, or subsequent discourse analyses. They are policy papers published by China, India, US, Germany, and France. The details of these documents are presented by the chart below.

Grounded Theory

Grounded theory is a methodology of research for exploring theory in a substantive area Hernandez, Cheri Ann. 2009. Theoretical Coding in Grounded Theory Methodology. Grounded Theory Review 8, no. 3.. It is also a systematic methodology that involves theory construction by methodical gathering and analysis of data Martin P et al. 1986. Grounded theory and organizational research. The journal of applied behavioral science 22, no. 2 . A study based on grounded theory includes developing substantive codes, in either vivo codes or sociological constructs, that derive from data. They are repeated ideas, elements and concepts that become apparent during the line by line examination of the data Hernandez, Cheri Ann. 2009. Theoretical Coding in Grounded Theory Methodology. Grounded Theory Review 8, no. 3.. Through coding process, and integrating all substantive codes with core data, the final theoretical codes, or in this case, concepts are emerged. Substantive codes must fit the data, otherwise the theoretical codes related to them are empty abstractions Glaser, B. G. 1978. Theoretical sensitivity. Mill Valley, CA: Sociology Press The purpose of theoretical codes is to integrate the substantive theory, and it can be either implicit or explicit Glaser, B. G., and Holton J. 2005. Staying open: The use of theoretical codes in grounded theory. The Grounded Theory Review 5, no. 1. In sum, identifying these codes are critically important to explanatory substantive grounded theory Hernandez, Cheri Ann. 2009. Theoretical Coding in Grounded Theory Methodology. Grounded Theory Review 8, no. 3..

A benefit of using grounded theory in this research is that it shares tremendous similarities with discourse analysis, which will be used in further testing of the validity of the findings, and in operationalize the findings. The concept of “modes” of discourse, is a nearly outdated concept. Such a concept is being employed in discourse studies as early as in the 19th century. Samuel P Newman identified five types of modes in 1827 - didactic, persuasive, argumentative, descriptive and narrativeNewman, S. P. A. 1851. Practical System of Rhetoric. Newman and Ivis. that laid out the foundations for modes of discourse. In the 20th century, modes of discourse encountered major criticisms in academia, as many began to emphasize the relations of discourses rather than the natures of discourses in a given text; Smith, Carlota S. 2003, Modes of discourse: The local structure of texts. Vol. 103. Cambridge University Press however, in 21st century social science, “different disciplines have developed different modes of discourse analysis in dependently or through borrowing across disciplines Hewitt, Sally. 2009. Discourse analysis and public policy research. Centre for rural economy discussion paper series 24 . Similarly, grounded theory also treats the data as units, assign and develop various codes, either substantive or theoretical to these units.

Another major similarity between grounded theory is discourse analysis is that they can be used for dealing with the same type of data - written text; however, there is also incompatibility of these two methodologies. On one hand, the main feature of grounded-theory is its minimization of assumptions. It is an inductive methodology that requires mind-openness in coding. On the other hand, constructivist policy discourse analysis demands the interpretation of data based on a cultural context.

For addressing this incompatibility, the author makes three arguments. Firstly, AI policy is a new area in public policy; therefore, the relations between different branches of AI policy theories - ranging from machine ethics and accountable algorithms to AI in economy and superintelligence - are fragmented and disorderly. They are not unified by a general theory that seeks to comprehensively explain AI policy. The availability of a theoretical framework remains low, and thus, de facto theoretical assumption in academia remains minimal. Secondly, scholars such as Barney G. Glaser argue that a grounded theory research also needs to be theoretical sensitive - researchers are encouraged to build an understanding of possible theoretical codes Glaser, B. G., and Holton J. 2005. Staying open: The use of theoretical codes in grounded theory. The Grounded Theory Review 5, no. 1. Thus, extensive reviewing of established literature on AI policy and genealogy is necessary for developing the conceptual framework for sorting the relations between different components of AI policy throughout coding process.

Conducting a Grounded Theory Analysis

For this research, the data used for the coding process consists of 6 documents drawn from general corpus that are published by five countries - China, India, German, US, and France. The process of coding is demonstrated by the example below.

“Bet on French talent

France is already home to much talent. These experts are what make France a leader in artificial intelligence. It is up to us to support this promising ecosystem.

Tech giants including Facebook, Google, Samsung, DeepMind, Fujitsu and IBM have understood the potential of French talent.

They have chosen to establish their new artificial intelligence research and innovation centres in France.

To transform trials, support our ecosystem and ensure its success in global competition, here are three tangible measures:

· We will establish a national artificial intelligence programme, capable of training and attracting the best global researchers and coordinated by the French National Institute for Computer Science and Applied Mathematics (INRIA). It will include a network of four or five institutes across France.

· The number of students trained in artificial intelligence will be doubled.

· The industry and public research sectors are currently two different worlds. Public researchers will from now on be able to dedicate 50% of their time to private entities.

Speeding up the emergence of artificial intelligence also requires us to provide resources equal to our ambition: the government will dedicate ˆ1.5 billion to development of artificial intelligence by the end of the current five-year term, including ˆ700 million for research...”

The data is broken into key points and given identifiers. Where P indicates the key point in a continuous sequence, the general corpus is identified a suffix A, B, X, Y, Z where A indicates the document “AI for Humanity,” B indicates “National Strategy for Artificial Intelligence - #AIforAll,” X indicates “New Generation of Artificial Intelligence Development Plan,” Y indicates Three-Year Action Plan for Promoting Development of a New Generation Artificial Intelligence Industry (2018-2020),” and Z represents “Rise of the Machines Artificial Intelligence and its Growing Impact on U.S. Policy.” The example below demonstrates the coding process.

Concepts emerge through continuous and subsequent coding. It is important to keep in mind that these concepts are backed by the theoretical codes generated in the data. The map below represents the theoretical codes, or “concepts” that are processed with their underpinned conceptual frameworks and the relations between these concepts. Developing a conceptual map is not unique to this paper because the use of conceptual map in grounded theory enables readers to understand how an inductive framework is constructed by coding. It is recommended by scholars such as Glaser for helping to maintain a conceptual level when describing these concepts and their relationships Glaser, B. G., and Holton J. 2005. Staying open: The use of theoretical codes in grounded theory. The Grounded Theory Review 5, no. 1.

Important Concepts

In this section, the paper examines 8 groups of concepts that are situated across the conceptual map. A major limitation of the conceptual map presented by the research is that the connectivity between different concepts and the interconnectivity between various concepts are not elaborated. Thus, this section is dedicated for explaining some of these concepts. The selection of groups are not based on whether or not these concepts have a direct connection with a bigger concept (e.g.

For explicating some of these connections, the author identifies each concept from the conceptual map that can be considered “a point of convergence” where their subsequent concepts converge. Each individual concept is then clustered into eight groups of concepts based on their theoretical connectivity.

In order to understand the concepts behind benefits and risks associated with AI, the paper chooses a variety academic articles and conference reports presented by Future of Life Institute The paper also examines articles and documents from the digital library of China National Knowledge Infrastructure (CNKI, ÖйúÖªÍø) constructed by Tsinghua University and supported by the Ministry of Education, Ministry of Science, and General Administration of Press and Publication of the People's Republic of China. As a part of literature review, the paper seeks to identify and summarize major benefits and risks associated with AI and identified by scholars, researchers, and others in each group of concepts. The overview of benefits and risks established by academic community also provides theoretical validity for some of the emergences of codes. Nonetheless, it is important to keep in mind that we are discovering new benefits and risks associated with AI almost on daily bases; therefore, does not cover every existing document, literature, and report on AI, instead, it intends to pave the way for a more comprehensive overview of these concepts.

1. AI Research and Development

There are many industry reports, conferences, and scholarly articles that discuss the principles and goals of beneficial AI R&D and potential costs and risks. For instance, the 2017 Asilomar conference in which a group of AI researchers and thought leaders from both academia and industry were invited to share their opinions on beneficial AI envisions research goal, recommends research funding, and promote research culture regarding AI R&D. The Information Technology Industry Council (ITI) also released their AI Policy Principles that emphasize standardization, promote global collaboration, education, and public-private partnerships. In academia, recommendations regarding AI R&D are more specific. Seth D Baum at Global Catastrophic Risk Institute summarizes measures that could be employed by decision makers for regulating, promoting, and motivating AI researchers to choose beneficial AI designs and to consider their social impacts Baum, Seth D. 2016. On the Promotion of Safe and Socially Beneficial Artificial Intelligence.. Global Catastrophic Risk Institute.. Nicholas Chen, Lau Christensen et al. discuss the pros and cons of investments regarding AI R&D that include sector investments, private industry investment, and venture capital investment Chen, Nicolas et al. 2016. Global Economic Impacts Associated with Artificial Intelligence. Analysis Group..

The overall goal of AI R&D can be summarized as promoting beneficial research and development of artificial intelligence while minimizing, controlling, and banning unbeneficial, unethical, dangerous and hazardous research and development. These sets of principles, on one hand, might contradict to values such as academic freedom, and capital advances. On the other hand, scholars like Seth Baum believe “AI should be built so as to have net benefits for the whole of society--or, in the face of uncertainty, net expected benefits Baum, Seth D. 2016. On the Promotion of Safe and Socially Beneficial Artificial Intelligence.. Global Catastrophic Risk Institute..” The paper takes a similar position in this regard. Specifically, benefit maximization strategy in AI R&D involves general strategy development, securing funding, intersectional R&D strategy, data gathering and data security, and promoting a favorable environment, more cross-sectorial collaborations, and building social trust. Risk and cost minimization, on the other hand, involves constructing legal frameworks, disseminating legal information, standard building, and preventing an undesirable environment in which harmful researches are conducted.

2. Economic Impacts, Labor Shifts and unemployment, and Inequality

In comparison to AI R&D, scholars have reached a greater agreement on the economic impacts of AI, the subsequent potential labor shifts and unemployment. Nonetheless, it is necessary for policymakers to develop a framework for facilitating the changes brought by AI. On one hand, institutions such as McKinsey Global Institute predict that AI has benefits for the economy as it could become the doorway to many capital-saving and labor-saving technologies “A Future That Works: Automation, Employment, and Productivity.” 2017. McKinsey Global Institute.. Similar views are expressed by economists who generally enthusiastic about the scenarios where AI benefits the economies as they tend to link innovation to economic growth Furman, Jason, and Robert Seamans. 2019. AI and the Economy. Innovation Policy and the Economy 19.1. Empirical studies could help showing the relations between technologies and growth. For instance, an empirical study conducted by Graetz and Michaels on robotics concludes that it contributed 0.4 percent of annual growth of GDP on average for 17 sampled countries Graetz, Georg, and Guy Michaels. 2018. Robots at work. Review of Economics and Statistics 100, no. 5. On the other hand, most of the discussions on the economic risks of AI involve labor shifts and inequality. For the workforce, adopting AI and automation might jeopardize many people's livelihood since it is clear that AI has a potential disruptive nature to the labor market in crucial waysKorinek A, Stiglitz JE. 2017. Artificial intelligence and its implications for income distribution and unemployment. National Bureau of Economic Research - labor displacement and increase inequality are situations that policy makers ought to avoid. Similar views also are expressed by Pew Research Center Smith, Aaron. 2016. Public predictions for the future of workforce automation. Pew Research Center,, World Economic Forum The Future of Jobs Employment, Skills and Workforce Strategy for the Fourth Industrial Revolution. World Economic Forum, and OECD “The Risk of Automation for Jobs in OECD Countries A Comparative Analysis,” OECD Social (2016).

Maximizing the economic benefits of AI therefore involves the implementing various principles, strategies, and policies for encouraging economic growth, increasing profitability for businesses, and reforming education policy for preparing future workforce to meet their challenges. Policy makers are also encouraged to minimize economic risks that come with the development and implementation of AI by ensuring human control over the economy, developing retraining programs, enhancing safety net, and taking care of people who are left behind or left out.

3. Accountability, Transparency, and Explainability

The accountability, transparency, and the explainability of AI Explanable AI (XAI) is defined by three features: An opaque systems that offer no insight into its algorithmic mechanisms; interpretable systems where users can mathematically analyze its algorithmic mechanisms; and comprehensible systems that emit symbols enabling user-driven explanations of how a conclusion is reached.

See Doran, D., Schulz, S. and Besold, T.R., 2017. What does explainable AI really mean? A new conceptualization of perspectives. arXiv preprint arXiv:1710.00794. are principles where tremendous risks are involved. Sacrificing any of these principles in the research, development, and use of AI and its related technologies could lead to great ramifications in society, economies, cybersecurity, and even national security. AI has the potential to complicate present legal structure since we could not define who should be held accountable Scherer, Matthew U. 2015, Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies. Harv. JL & Tech. 29 when AI causes harms. The absent of transparency in coding, data gathering, decision-making of AI could also cause negative results. Asilomar AI Conference recommends that it is important to obtain judiciary transparency - any involvement by AI in judicial decision-making should provide a satisfactory explanation auditable by a competent human authority, and failure transparency - the ability to explain when AI causes harms.The 2017 AI Now report states that the unconstraint use of “black box Bathaee, Yavar. 2018 The artificial intelligence black box and the failure of intent and causation. Harvard Journal of Law & Technology 31, no. 2” - the barriers in regulation and implementation of AI technology caused by its unexplainable nature. In other words, regulating and implementing AI decisions that we could not comprehend would come with great risks and the consequences of unexplainable AI could be disastrous. Even when AI that acts “rationally “acting rationally” is a key concept defining artificial intelligence developed by Russell and Norvig in their book - Artificial Intelligence: A Modern Approach.

See Russell, Stuart J., and Peter Norvig. 2016. Artificial intelligence: a modern approach. Malaysia; Pearson Education Limited,,.,” it could pose great public risks Scherer, Matthew U. "Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies." Harv. JL & Tech. 29 (2015): 353.. Thus, regulatory regimes might be pressed for reforms in the near future because of the regulatory purpose on AI is unclear.

The benefits that come with ensuring the accountability, transparency, and explanability of AI might be less clear; however, these qualities have the potential of improving AI end-products that are directly bought and sold in the market. The paper draws from the example of socially conscious products for demonstrating how consumer's changing expectation - from environmentally unconscious to enviromentally conscious - has an impact on the market. 67 percent of American consumers in 1989 said that they would be willing to pay 5-10 percent more for products that are ecologically compatibleCoddington, Walter. 1990. It's no fad: environmentalism is now a fact of corporate life. Marketing News 15, no. 7., but by the time of 1991, consumers that are conscious about the environment were willing to pay 15-20 percent more for green products Suchard, Hazel T., and M. J. Polonski. 1991. A theory of environmental buyer behaviour and its validity: the environmental action-behaviour model. In AMA summer educators' conference proceedings, vol. 2, pp. 187-201. Chicago, IL: American Marketing Association. Similarly, AI end products should be encouraged to resemble values that humanity embraces. Policy makers should motivate AI researchers, firms, and AI system designers to develop algorithms that could prevent causing public harms before unexpected and undesirable outcomes could erupt. American scholars like Joshua A. Kroll states that in the American legal system, uncertainties and ambiguities are often left undiscussed and unaddressed until a dispute arises and its resolution becomes necessary; therefore, computer scientists should try to create algorithms that are reviewable instead of simply complying with the requirements generated during the drafting process of new law and regulationKroll, Joshua A et al. 2016. Accountable algorithms. U. Pa. L. Rev. 165.. Designing better AI products that not only comply with legal requirements, but are also intrinsically socially conscious could be beneficial to both consumers and businesses.

There are two main approaches to what policymakers should do in benefit maximization and risk & cost minimization. The hardliner approach - standardization, regulations, and legal constraints - emphasizes on minimizing risks and costs whereas the alternative approach emphasizes on making AI systems and AI products accountable, transparent, and explainable by means that are outside the realm of traditionally regulatory and legal toolkits. In contrast, creative measures such as the tort system Scherer, Matthew U. 2015, Regulating artificial intelligence systems: Risks, challenges, competencies, and strategies. Harv. JL & Tech. 29 or a sliding scale Bathaee, Yavar. 2018 The artificial intelligence black box and the failure of intent and causation. Harvard Journal of Law & Technology 31, no. 2 seek to strike a balance between intellectual freedom and accountability.

4. Surveillance, Privacy and Social Risks

AI expands with the emergence of facial recognition and preventive policing Defined as Guidance on Where and When to Patrol provided by AI system and algorithms. On one hand, AI embedded surveillance technologies could be beneficial in response to challenges facing governance, law enforcement, public safety, and crime-prevention. For instance, Commercial Off the Shelf (COTS) systems provide functions like motion detection and movement recording in the field of view with the use of autonomous video analysis Clift, Louis G., Jason Lepley, Hani Hagras, and Adrian F. Clark. 2018. Autonomous computational intelligence-based behaviour recognition in security and surveillance. In Counterterrorism, Crime Fighting, Forensics, and Surveillance Technologies II, vol. 10802, p. 108020L. International Society for Optics and Photonics.. They could set-off alarms via text message, email and other alarm systems based on defined thresholds when event triggers are configured. Companies such as Sensentime (ÉÌÌÀ) and Intellifusion have implemented these technologies in their products, and governments have already been using them in day-to-day detections, tracking, monitoring, and other purposes. Other companies like PredPol pride themselves for their advance predictive algorithms that are used in crime prevention. They use machine-learning technology for calculating predictions based on historical dataset, real-time data collection, and other information such as weather, shop locations and traffic light signals. Moreover, the commercial values of such technologies are tremendous as they enable companies to gather consumer profiles and information, predict future business prospect, and to grant quicker access of payment to consumers with facial-based payment systems

On the other hand, the use of AI in surveillance and crime detection spurred controversies in both the US and China as many people fear their potentials of infringing privacy and civil liberties. Privacy International has identified facial recognition technology used by authorities in monitoring assemblies that requires particular attention. Although China has been the major target of criticism as the Chinese government deploys facial recognition systems in Xinjiang for monitoring its Uyghur population, the ubiquity of AI embedded surveillance technologies is also increasing elsewhere. In the UK, combined with a security camera, a trial facial recognition software used by the police department could identify suspects in merely seconds. In Russia, an app called Findface that allows users to photograph people within a crowd and figure out their identities with 70 percent reliability has signed a contract with Moscow city government to deploy their technologies in a network of 150,000 surveillance cameras. As a journalist describes, it might effectively bring an end to public anonymity. Nonetheless, it is important to understand that facial recognition technology also comes with great risks.

Machine bias is another aspect of the potential risks that come with the use of AI. A ProPublica investigation conducted in 2016 revealed a case where an AI software used across the US to predict criminals is biased against blacks. They discover that the formula used by the software is more likely to flag blacks as potential criminals, and mistakenly label them in this way by almost 100 percent in comparison to whites Angwin, Julia et al. 2016. Machine Bias - There's software used across the country to predict future criminals. And it's biased against blacks. ProPublica. Cathy O'Neil also points out that there are harmful consequences of reinforcing discrimination and inequality due to opaque and unregulated big data mathematical models O'Neil, C. 2016. Weapons of Math Destruction. New York, New York: Crown...

In sum, benefits related to AI-powered surveillance is centered around two themes. Firstly, it enhances authorities' ability to gather information, detect, monitor, track, and ultimately prevent events that could endanger public safety. Secondly, it improves quality of business and transaction for both business owners and consumers. Meanwhile, costs and risks are mostly related to its potential of jeopardizing privacy and civil liberties. It is worth noting that to some, the importance of privacy and civil liberties is a debatable matter. Paul Triolo and Kaifu Li believe state that “China possesses a vast amount of usable data, and to the Chinese people, privacy is a lesser concern (such as using facial recognition cameras in shopping malls for locking down costumers with the help of potential business information). Thus, China will become the leading figure in tangible AI technologies “ÖйúÓµÓдóÁ¿µÄ¿ÉÓÃÊý¾Ý£¬¶øÖйúÈ˶ÔÓÚÒþ˽Ïà¶Ô²»Ì«¾À½á£¨ÈçÉ̳¡ÄÚµÄÉãÏñÍ·¿É ÒÔʶ±ðÈËÁ³£¬²¢ÒÔDZÔÚÉÌÒµÐÅÏ¢Ëø¶¨Óû§£©£¬ ¹Ê¶øÖйú½«³ÉΪʵÌåÊÀ½çÖÇÄÜ»¯µÄÁìÍ·ÈË¡£”

Triolo, Paul and Li, Kaifu. 2018. China's AI revolution - understanding the structural advantages of China (ÖйúµÄÈ˹¤ÖÇÄܸïÃü£º Àí½âÖйúµÄ½á¹¹ÐÔÓÅÊÆ). China Academic Journal Electronic Publishing House. Tangible AI technology (ʵÌåÊÀ½çÖÇÄÜ»¯) here, can be interpreted as AI applications.

5. Fairness, Ethics, and Human Right

Fairness, ethics, and human rights are concepts closely related to the realm of public policy since they often invoke philosophical problems that emphasize on the relations and connections between rationality and morality Risse, Mathias. 2019. Human Rights and Artificial Intelligence: An Urgently Needed Agenda. Human Rights Quarterly 41, no. 1 . It is generally agreed that AI must consist rationality. For instance, Stuart Russell and Peter Norvig define AI in four different aspects - thinking humanly, acting humanly, thinking rationally, and acting rationally. They believe that the broadest sense of AI embodies the quality of a “rational agent” that “that acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome Russell, Stuart J., and Peter Norvig. 2016. Artificial intelligence: a modern approach. Malaysia; Pearson Education Limited, .” Mathias Risse et al. stress that the problem of value alignment problem Risse, Mathias. 2019. Human Rights and Artificial Intelligence: An Urgently Needed Agenda. Human Rights Quarterly 41, no. 1 requires particular ontological discussion, because we cannot fathom the morality the morality and ethics of AI although we must ensure that AI is designed to respect human rights. It raises the question of how could we ensure that the values that are encoded in the design of AI align with the values of human.

On one hand, David Hume, for instance, points out that rationality has nothing to do with predefined morality. In other words, AI with rationality might determine its own goals, the way to achieve the goals, as well as disregarding human values in its rationality. A classic example illustrated by Nick Bostrom, a philosopher at Oxford University hypothesize a `paperclip maximizer” that

“decides to amass as many paperclips as possible. It devotes all its energy to acquiring paperclips, and to improving itself so that it can get paperclips in new ways, while resisting any attempt to divert it from this goal. Eventually it “starts transforming first all of Earth and then increasing portions of space into paperclip manufacturing facilities.”