Conceptual map for artificial intelligence policy discourse analysis

The “paperclip maximizer” example constitutes a frame problem - the problem of defining explicit sets of axioms without having to defining large sets of obvious non-effects in a machine environment Shanahan, Murray, and Bernard Baars. 2005. Applying global workspace theory to the frame problem. Cognition 98, no. 2.

On the other hand, Kantian philosophers hold a view in contrast to that of David Hume. They often invoke the concept of Categorical Imperative (CI) - an objective that is necessary for rationality and unconditional principle that applies to all agents including AI. In this scenario, all specific moral requirements are justified by this principle, thus, it implies that all immoral actions such as lying, sealing and killing are irrational because they violate the CI, and therefore rendering actions such as “paperclip maximizer” impermissible. If Kent were right, AI would naturally commit to these sets of principles precisely because of its supreme rationality, and therefore would become the absolute role-model in human society Risse, Mathias. 2019. Human Rights and Artificial Intelligence: An Urgently Needed Agenda. Human Rights Quarterly 41, no. 1 .

This paper does not seek to participate in in-depth philosophical concepts although the author acknowledges that they constitute the foundations for understanding AI v. human values. In reality, since AI is already deployed in important areas such as criminal justice, access to financial system, healthcare, education, and online content moderation, it is necessary to discuss what policymakers ought to do in order to promote beneficial AI and to solve value alignment problem due to its potential to harm human rights, incite ethical hazard, and undermine fairness and equality.

Filippo A. Raso, et al suggest that both opportunities and risks exist in areas where AI impacts on human rights. For example, they reveal that AI systems may reproduce biases in criminal justice due to selected data that is used in machine learning, whereas other evidence indicates that racial disparity may be reduced in sentencing and bail Raso, Filippo A et al. 2018. Artificial Intelligence & Human Rights: Opportunities & Risks. Berkman Klein Center Research Publication . Ethically Aligned Design (EAD) purposed by the IEEE Global Initiative on Ethics of Autonomous recommends three pillars - universal human value, political self-determination and data agency, and technically dependability - upon which the solutions for value alignment problem may be built Ethically Aligned Design - A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems. 2017. IEEE Advance Technology for Humanity. 2ed Edition.. Furthermore, stakeholders including universities, firms, researchers, scholars and thought leaders have taken a substantial step in the making and signing of the Toronto Declaration: Protecting the right to equality and non-discrimination in machine learning system in which they suggest using framework of international human rights law, emphasizing states' duties and their obligations to human rights, and encouraging private sector actors to share responsibilities. It is apparent that the emergence of AI may lead to great disruptions for fairness, ethics, and human rights. These are the universal values that the United Nations - the most important institution upon which the current international system is built - also holds sacrosanct

6. Natural Language Processing (NPL)

Natural language generation technology uses structured data for automated generation natural language. The first use of such a technology in journalism can traced back to 1980s, but the term computational journalism was not used as a collective term in discussion until very recently. The term was coined in 2006 by Irfan Essa and subsequently by Diakopoulos blog post in 2017 Cohen, S., Hamilton, J. T., & Turner, F. 2011. Computational journalism. Communications of the ACM, 54(10), 66-71. Early discussions about computational journalism had been around the concept of watchdog journalism that seeks to “hold leaders accountable, unmask malfeasance, and make visible critical social trends,” and to provide citizens with “the information they need to make many important choices Thurman, Neil. 2018. Computational Journalism. Forthcoming, Thurman 2019. Scholars focused on the positive attributes, potentials, and new excitement that could be beneficial to producing better quality journalism. For example, Sarah Cohen believes in the possibility of increasing the public's ability to monitor power and envisions computational journalism as a tool to help leveling the playing field between the relatively powerful interests and the public Cohen, S., Hamilton, J. T., & Turner, F. 2011. Computational journalism. Communications of the ACM, 54(10), 66-71.

After the initial optimism, a wider variety of the use of computational journalism and practices began to appear. Although the number academic articles about the benefits of computational journalism in helping journalists to discover, extract, visualize and disseminate information continue to grow, more articles also began to pay more attention and emphasis on the risks, limitations, bad behaviors and negative impacts associated with computational journalism. Recent events demonstrate that highly personalized and targeted computational propaganda and disinformation are already being used by interest groups, institutions and individuals for personal and political gains. An investigation into bot propaganda concludes that bots have created measurable influence during the 2016 American election by controlling flow of information between users and garnering the most attention and influence amongst human users on Twitter Wooley, Samuel C and Guilbeault, Douglas R. 2017. Computational Propaganda in the United States of America: Manufacturing Consensus Online, Computational Propaganda Research Project, University of Oxford. Another working paper produced by scholars at Oxford University draws three more alarming conclusions. Firstly, the masterminds behind computational propaganda are aware of the increasing negative coverage of their doings in the news media and they are adopting new algorithms in response. Secondly, sleeper bots - accounts that only tweeted a couple of times and in scattered pattern - are prevalent in countries where Twitter is not a common social media platform. Thirdly, due to the differences of legal aspects in countries where bot-propaganda exists, providing public policy recommendations based on the findings is deemed difficult Woolley, Samuel C and Howard, Philip N. 2017. Computational Propaganda Worldwide: Executive Summary. Oxford University.

Moreover, communication technology also risks increasing manipulations. MADCOMs - the integration of AI systems into machine-driven communications tools for use in propaganda Chessen, Matt. 2018. The MADCOM Future: How Artificial Intelligence Will Enhance Computational Propaganda, Reprogram Human Culture, and Threaten Democracy… and What can be Done About It. In Artificial Intelligence Safety and Security, pp. 127-144. Chapman and Hall/CRC may lead to potential undesirable outcomes due to its enhanced capabilities. For instance, Persuasive and manipulative communications could be generated by AI infused machine systems without our awareness - it knows the news websites and social media an individual frequently visits and uses, and it generates dynamic and robust content that is tailored based on the information it gathered from an individual user; however, AI-enhanced communications tools may lead to positive consequences as well. For instance, AI chatbot such as Microsoft Xiaoice is now taking up the tasks traditionally burdened by mankind - it is “now designing images and patterns on fabrics for international fashion and garment producers. Apart from developing intelligence quotient (IQ) by deep learning technique, Xiaoice also tries to build up its emotional intelligence (EQ) with the help of interactions with human to acquire understandings of human social skills and knowledge Spencer, Geoff. 2018. Much more than a chatbot: China's Xiaoice mixes AI with emotions and wins over millions of fans. Microsoft. . AI breakthroughs and advances in the case of Xiaoice, may have enormous positive implications.

In addition, computational journalism also has impacts on firms, businesses, and the capital market. In 2014, the Associated Press (AP) started using algorithms to produce articles on firms' earning announcement. Elizabeth Blankespoor et al note that a significant increase in trading volume was observed minutes after the release of AP articles generated by automation during and after trading hour for both low and high visible firms Blankespoor, Elizabeth, and Christina Zhu. 2018 Capital market effects of media synthesis and dissemination: Evidence from robo-journalism. Review of Accounting Studies 23, no. 1.

Guiding policy principles and policy recommendations thus focus on four major aspects - first, using natural language generation technology and AI systems to improve the quality of journalism in exploring, extracting, articulating, and disseminating information; second, improving social and processes by using technologies that respect, embody, and share human values; third, utilizing technological advances for improving quality of services; and fourth, preventing political manipulation and computational propaganda. Moreover, there are areas such as national security and cybersecurity where AI-enhanced, machine-driven technologies may maximize their utilities.

7. Human Health

AI is used for analyzing a great amount of biomedical data and helping with the development of drugs, diagnosis and treatment. It also has the ability to recognize patterns in data and advancing further based on its mistakes for identifying new information Pauwels, Eleonore. 2017. How to Optimize Human Biology: Where Genome Editing and Artificial Intelligence Collide. Wilson Center. The potential of AI in medical fields is apparent in predicting RNA splicing patterns in mouse cells Pauwels, Eleonore. 2017. How to Optimize Human Biology: Where Genome Editing and Artificial Intelligence Collide. Wilson Center, optimizing human biology combining with CRISPR technology, and predicting and diagnosing cardiovascular disease Krittanawong, Chayakrit, HongJu Zhang, Zhen Wang, Mehmet Aydar, and Takeshi Kitai. 2017. Artificial intelligence in precision cardiovascular medicine. Journal of the American College of Cardiology 69, no. 21. Eleonore Pauwels at Wilson Center identifies three potential benefits of AI in the medical field. Advances in machine deep learning envisage are enabling a shift in many areas. Firstly, AI has the power to provide additional utilize optimization tools and lead to the developments of new optimization methods. Secondly, AI is capable of providing in-depth analyses of available scientific information for modeling the molecular mechanics in discovering the relations between genetic variation and disease. Thirdly, technological developments may lead new breakthroughs in the application of AI in genome editing Pauwels, Eleonore. 2017. How to Optimize Human Biology: Where Genome Editing and Artificial Intelligence Collide. Wilson Center. Simultaneously, there are limitations in AI as well. Firstly, there is the problem of `black box'. Improving the expandability of AI in the medical field is critical to building trust and the expansion of AI applications especially in the medical field and genome editing. Secondly, the accuracy of deep learning projects for modeling human biology and making predictions needs to be improved. Since high-quality and high-availability of data and scientific information are critical to deep learning, they must be improved.

Moreover, AI-enhanced brain-computer interface (BCI) technology is making the province of science fiction into reality. Scientists are now able to use chips implanted against the brain to read the firings of hundreds of neurons in the brain. BCI is already utilized for provides new functionality to people suffering from paralysisJavaid, Muhammad. A. 2014. Brain-Computer Interface. SSRN. It is evident that continuous research for advancing in BCI may lead to vast amounts of benefits we are pursuing on a path to a reality where decoding people's mental processes and manipulating the brain mechanism that configure intentions, emotions and decisions will be possible Wolpaw, Jonathan R., Niels Birbaumer, Dennis J. McFarland, Gert Pfurtscheller, and Theresa M. Vaughan. 2012. Brain-computer interfaces for communication and control. Clinical neurophysiology 113, no. 6. With BCI technologies, we may be able to discover better treatment for brain injury, paralysis, epilepsy and schizophrenia. We may also adopt BCI for improving human experience Yuste, Rafael, Sara Goering, Guoqiang Bi, Jose M. Carmena, Adrian Carter, Joseph J. Fins, Phoebe Friesen et al. 2017. Four ethical priorities for neurotechnologies and AI. Nature News551, no. 7679.

Meanwhile, scientists are addressing the ethical issues for the research and development of BCI technologies. Yuste et al point out that in neurotechnologies and AI there are four ethnical concerns. Firstly, BCI may lead to violations of privacy and consent; therefore, commercial data sharing and transfer and centralized processing of neural data must be regulated and restricted. Secondly, agency and identity may be disrupted by neurotechnologies that have the capability to influence people's perceptions, behaviors and actions. They suggest that adding and incorporating `neurorights' into international treaties such as the Universal Declaration of Human Rights (UDHR) may help developing a mechanism for sufficient protections against the misuse of neurotechnologies. Thirdly, they believe that an augmentation arms race may lead to devastating consequences. Thus, augmentation should be heavily regulated especially in the field of national security. Lastly, preventing bias and making sure everyone can be benefited by the development of AI-enhanced neurotechnologies are important for the development, and such a concern can be addressed by including the data of problem user groups into the design of algorithms and AI systems Yuste, Rafael, Sara Goering, Guoqiang Bi, Jose M. Carmena, Adrian Carter, Joseph J. Fins, Phoebe Friesen et al. 2017. Four ethical priorities for neurotechnologies and AI. Nature News551, no. 7679.

Tremendous opportunities and benefits that come with the use of AI inhuman health are reflected in the general corpus. They all stress the importance of investments in AI systems, AI applications, and algorithms for improving public health. The possibilities of using AI-enhanced robots for improving diagnosis, treatment, and operation of surgery are discussed in the Chinese, French, German, and the American documents, whereas the communication issued by the EU focuses on facilitating research and infrastructures, and financial services for supporting AI development in health. Notably, the Chinese documents do not specifically touch upon the risk minimization aspect of AI for human health, whereas supporting the growth of AI in health industry is the primary focus for the Chinese government.

AI Safety

AI safety is a term used to for referring to multiple meanings. It can refer to efforts to prioritize research of robust and beneficial AIRussell, Stuart J, Daniel Dewey, and Max Tegmark. 2015. Research priorities for robust and beneficial artificial intelligence. Ai Magazine 36, no. 4., and it can also refer to efforts to design core safety mechanism that tackles the “control problem,” and other issues related to predictability and undesirable side-effects Amodei, Dario, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. Concrete problems in AI safety. arXiv preprint arXiv:1606.06565 . Additional tests and experiments must be conducted for ensuring AI safety before using an AI system in applications for avoiding unwanted consequences. It is sometimes referred to the value alignment problem that may lead to mismatches of values between human and AI. Moreover, there are social and political factors and conditions that can promote AI safety by stressing the importance of equality, fairness, and equity.

AI may have the possibility to improve AI safety by itself. Such duties rely on governments, researchers and engineers. For instance, a report titled “The Malicious Use of Artificial Intelligence Forecasting, Prevention, and Mitigation,” suggest that they “should take the dual-use (beneficial use and malicious use) nature of their work seriously,” by always considering the impacts of misuse when deciding research priorities and norms. Moreover, they should also collaborate with other stakeholders. Dario Amodei et al. identify that many problems of accidents in AI systems and harmful behavior may emerge from poor design of real-world AI systems, and thus, the solution could be tackled with better-designed algorithms and AI systems. Moreover, in terms of technical solutions, they presented the idea that increasing trend towards end-to-end (as oppose to case-by-case rules or hoc fixes for fixing the issues) points towards a necessary unified approach to prevent AI systems from causing unintended harms Amodei, Dario, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. Concrete problems in AI safety. arXiv preprint arXiv:1606.06565 .

Nonetheless, a vast amount of issues, questions, and concerns related to AI safety exist. Yampolskiy and Spellchecker M. S. conclude that “full autonomous machines can never assume to be safe Yampolskiy, Roman V., and M. S. Spellchecker. 2016. Artificial intelligence safety and cybersecurity: A timeline of AI failures. arXiv preprint arXiv:1610.07997 .” Although we are still decades behind fully embracing autonomous machines, fully autonomous and semi-autonomous systems and robots are already operating in fields such as medicine, business management, and transportation.

1. Summary

The table below summarizes some of the risks related to these groups of concepts. The author acknowledges that more concepts that can be considerer “points of convergence” need to be identified. Thus, it requires further exploration in academic literatures.

Nonetheless, elaborating these concepts and their connectivity along does not prove the validity of the map in which they are situated. Testing the validity of the conceptual map is as important as explaining these concepts.

Part II

For the purpose of understanding whether or not AI-related benefits and AI-related risks are at the center of AI policy, whether or not the general corpus exhibits the features of these two theoretical codes is examined through a discourse analysis. For further operationalizing the conceptual map, the author seeks begins with drawing hypotheses based on the conceptual map and real-life evidence. A following discourse analysis is conducted for testing these conceptually generated hypotheses. Nonetheless, before proceeding the analysis, existing definitions of AI policy need to be studied.

What is AI Policy?

AI bring about benefits ranging from promoting economic growth and strengthening public safety to enhancing human health and national defense. Simultaneously, it also creates issues like unemployment and violation of privacy, and increases risks in areas such as ethical liability, and automated weapons. To Tim Dutton, a senior policy advisor for the Office of the Premier of Ontario and a prominent AI researcher, AI policy is the set of public policies that aim at maximizing the benefits of AI while minimizing its potential costs and risks Dutton, Tim. 2018. AI Policy 101: An Introduction to the 10 Key Aspects of AI Policy. Medium. Similar definitions are developed by other scholars as well. For instance, Miles Brundage and Joanna Bryson states that national governments can “help society to reap the benefits while reducing the downsides of artificial intelligence.

From a genealogical perspective, more evidence can be found in digital library of Future of Life Institue - a non-profit AI research organizations with members and advisors coming from world-renown universities such as Harvard University and Oxford University, Cambridge University and MIT, and from tech-giants such as Skype and Tesla See Who We Are? Future of Life Institute. Retrieved from https://futureoflife.org/team/. In their genealogy, there are 14 areas of AI policy Future of Life Institute. Retrieved from https://futureoflife.org/team/. These areas - ranging from Global Governance, Race Conditions, to Existential Risk are mostly situated in the conceptual map and in relations with one another with the exception of Catastrophic Risk, Human Dignity, and psychological impact.

In any policy, states are the primary actors in the agenda-setting, articulation, implementation, and monitoring - these stages of policy process. It is evident in the increasing number of countries with national AI policies and strategies. From 2017 to 2018, Canada, China, the EU Commission, Denmark, France, Finland, Italy, India, Japan, Singapore, South Korea, Taiwan and Sweden have all published their strategies to encourage and promote AI Dutton, Tim. 2018 An Overview of National AI Strategies. Medium. . Thus, it is important to understand whether a similar narrative- benefits v. risks - can be reflected in discourses at a national level. If so, what are the evidences? For addressing these questions, a comprehensive discourse analysis of the general corpus is required before coming to a conclusion.

Benefits and Risks - Are they Dominant Discourses?

Returning to the general corpus, a discourse analysis is conducted by the author for operationalize the conceptual map. Analyzing the use of language in the general corpus is important for understanding the construction of reality - what the text really means? In this research, the author takes a social constructivist approach that treats written, vocal, or sign language as social constructs that are reflected in the centrality of everyday activities including the making of AI policy Potter, Jonathan. 1996. Discourse analysis and constructionist approaches: Theoretical background. British Psychological Society. The underlying assumption here is that AI's benefits, and its risks are a set of competing ideas. Moreover, they are situated as the dominant discourses for most discussions on AI policy. In most cases, emphasis on AI's benefits reflects to a techno-optimist position whereas focusing on the risks related to AI reflects to a techno-pessimist position. Related stakeholders' onions can be either general or specific. For instance, when many are enticed by the opportunities provided by AI, in 2017, Elon Musk baldly proclaimed on twitter that the development of AI will become an arms race that will lead the World War III. Whilst scholars in public policy generally believes that a certain level of regulations Scholars also share different opinions regarding regulations on AI. For instance, Seth D. Baum identifies both extrinsic (e. g regulatory policy) and intrinsic measures (e. g. constructing social norms by encouraging certain behaviors and values) for promoting beneficial AI, whereas the Malicious Use of Artificial Intelligence report recommends and emphasizes more collaborations between different stakeholders (e.g. a “red team” strategy and conferences). are required to ensure AI safety, prominent futurists like Michael Spencer believe that AI regulation may be simply impossible because it will eventually outsmart human. The apparent of this set of discourses is reinforced by the conceptual map grounded on data.

Discourses are used in different contexts. Norman Fairclough identifies three contexts in which discourses are used - 1. The making of meaning as an element of the social process; 2. The language associated with a specific social perspective; and 3. A way of interpreting aspects of the world related to a specific social perspective. It is important, for any discourse analysis, to define the social process and the particular social perspective for effective research. Firstly, the social process here is defined as the making, implementation, monitoring, evaluation, and the evaluation or reevaluation of AI policy. It is based on a simplified version of policy cycle theories Jann, Werner, and Kai Wegrich. 2007.Theories of the policy cycle. Handbook of public policy analysis: Theory, politics, and methods 125 in which the process of public policy is categorized into various stages. Secondly, it should have at least one actor in policy process involved, and is directed at one or more actors in policy process. Thirdly, whether or not actors in policy process are involved also defines premises for discourses on policy recommendations. intelligence unemployment inequality justice

From a linguistic perspective, discourse studies are defined by Jan Renkema as “the discipline devoted to the investigation of the relationship between form and function in verbal communicationRenkema, Jan. 2004. Introduction to discourse studies. John Benjamins., and it can be used in both board contexts and narrow contexts. When discourse refers to particular contexts, it becomes similar to concepts such as text typeBaker, Paul, and Sibonile Ellece. 2011 Key terms in discourse analysis. A&C Black..

Borrowing from the concept developed by Maarten A. Hajer, a discourse here is defined as “an ensemble of ideas, concepts, and categories through which meaning is given to social and physical phenomena, and which is produced and reproduced through an identifiable set of practices van den Brink, M. A., and Tamara Metze. 2006. Words matter in policy and planning. Cp. 4 Doing Discourse analysis: coalitions, practices, meaning.” Specifically, for instance, in his study, the use of term “acid rain” in text would also mean the traditions of dealing with environmental issues in the UK.

Combining with the aforementioned definition of a discourse, for this research, the author uses a constructivist methodology adopted by scholars such as Potter and Wetherell Potter, Jonathan, and Margaret Wetherell. 1987. Discourse and social psychology: Beyond attitudes and behaviour. Sage and Latour and Woolgar Latour, Bruno. 1986. Visualization and cognition. Knowledge and society 6, no. 6. The underlying unity of their approaches, according to Jonathan Potter, is that in their assumptions, they tend to emphasize that the mind and the actions contingent on specific cultural forms - in this case, the actions are debates and discussions on AI-related benefits and AI-related risks, as they can suggest other competing ideologies such as techno-optimism and techno-pessimism. They can also be considered the storylines, the narrative, the main theme, and the focus for AI policy debate.

From a comparative approach, the documents used for conducting the analysis represent four distinctively different countries with their unique strengths, weaknesses, socio-economic and cultural contexts. For instance, in an economic context, these countries can be sorted into two clusters. Based on GDP (current prices), China and the US make up a cluster with two largest economy, whereas Germany and India have similar sizes of GDP on a relative term. Similarly, based on GDP per capita, the US and Germany are in a similar position, whereas China and India significantly lag behind. In a cultural context, China and India are unique in their own ways, while the US and Germany belong to the Western hemisphere.

GDP, current prices

GDP per capita, current prices

If the presence of these discourses can be revealed in the general corpus that come from four vastly different countries, the hypothesis of their centrality can also be argued with a greater level of confidence.

Benefits and Risks - the Divergence of Discourses

In the general corpus, different documents exhibit various structures, styles of language, and messages. At a first glance, these documents have different volumes, writing styles and explicitness. For instance, India's “National Strategy for Artificial Intelligence - #AIforALL” has the largest volume amongst these documents with 115 pages, whereas “Rise of the Machines

Artificial Intelligence and its Growing Impact on U.S. Policy” contains only 16 pages; however, a larger volume does not necessarily indicate that more policy recommendations are given in the document. The Chinese and German documents demonstrate a greater level of structural and textual explicitness. In “New Generation of Artificial Intelligence Development Plan,”, language is used in a short, concise manner that tend to envisage multiple policies in a sentence. Documents such as “The National Artificial Intelligence Research and Development Strategic Plan,” and “National Strategy for Artificial Intelligence - #AIforALL” spend greater lengths on assessments of current situation regarding AI, and their strenghths and weaknesses regarding the development of AI.

Amidst these differences, the author reveals that there are four similar focal points in which the text flows are diverged, and eventually centered around the dominant discourses - benefits and risks.

First and foremost, it is apparent that these documents acknowledge the transformative nature of AI. A pressing concern for many is that the general population tends to be oblivious about the emergence of AI. Sometimes, people who question the AI's potential impacts the society and their way of life are in a state of denial due to the lack of understanding. National governments, in contrast, firmly believe in the transformative power of AI. They also acknowledge that the development of AI has entered a new phase where its potentials are rapidly being unlocked. The chart below demonstrates government's acknowledgements of the importance of AI.

Secondly, each document approaches AI in a similar fashion with the focus of its task-solving functionality in particular areas. For instance. Indian government defines AI in its policy paper as

“the ability of machines to perform cognitive tasks like thinking, perceiving, learning, problem solving and decision making.”

For India, AI embodies the functionalities of performing cognitive tasks that are previously only performed by humans. Similarly, we can find similar definitions of AI in “Rise of the Machines.”

“AI generally falls into two categories: “narrow AI” and “general AI.” Narrow AI addresses or solves specific tasks, “such as playing strategic games, language translation, self-driving vehicles, and image recognition…The examples of AI that are referred to in this paper concern the field of narrow AI.”

In the German and the Chinese documents, such definitions can seem less visible; however, insights can be gained from the structuring of texts. The Chinese documents use abstract concepts for defining the role of AI (functionality) in various areas; however, direct connections between AI and areas of its functionalities are explicitly drawn by defining areas for AI functionality. Two examples are illustrated in the chart below.

In the German document “Key points for a Federal Government Strategy on Artificial Intelligence,” although the definitions of AI remain absent, the document defines explicit areas of AI functionality. Two examples are illustrated below.

Thirdly, the author identifies several evidences that suggest that the divergence of discourses is centered around two themes - AI-related benefits and AI-related benefits. In the documents released by China, US, and Germany, these themes are very apparent and self-explanatory. For example, in the first section of “A New Generation of Artificial Intelligence Development Plan,” for providing an assessment of current situation, a statement is extracted by the author.

"While vigorously developing artificial intelligence, we must attach great importance to the possible safety risk [of AI]"

In this statement, developing artificial intelligence and attaching great importance to the possible risk is presented as a set of contrasting concepts. More examples of this set of concepts are illustrated by the chart below.

In the case of India, these two concepts become less apparent, as a similar expression cannot be located in a sentence in the document; however, the structure of the document provides an evidence for the centrality of this set of concepts. The structure of content in “National Strategy for Artificial Intelligence - #AIforALL” consists of 7 parts with the exclusion of introduction and executive summary, and the function (s) of each chapters is documented by the chart below.

Part 4 and the part 5 of the document share a significant commonality as they both discuss the usage of AI. In part 4, five areas that require AI intervention are identified due to their significance for the future of India, and part 5 focuses on the challenges in these identified. It suggests these of contrasting discourses in a less direct and obvious way.

Combining these commonalities that are identified, a model for the divergence of discourses into a set of discourses - AI-related benefits, and AI-related risks - can be drafted. The chart below illustrates the divergence of discourses along with the text flow.

In the beginning of the analysis, the author finds that these documents on AI policy follow a trajectory of story telling as illustrated above; however, after the divergence of discourse, these documents become vastly different. Nonetheless, AI-related benefits and AI-related risks remain the center of their focus. They serve as the key set of concepts that pull these documents together.

In conclusion, AI-related risks and AI-related benefits are the dominant discourses in the general corpus. In Hajer's term, they are a coalition of discourse Hajer, Maarten A. "Acid rain in Great Britain: environmental discourse and the hidden politics of institutional practice." In Greening Environmental Policy, pp. 145-164. Palgrave Macmillan, New York, 1995. that dominate the discussions on how to make sense of AI policy.

What is the dominant discourse amongst these two concepts in different countries? How to interpret the priorities of governments and their AI policies with distinctive feature? In the next section, the author operationalizes the conceptual map in two different ways - situating key points, and

Operationalizing the Conceptual Map

Considered a coded key point in the data-base.

Firstly, the ID Pz8a indicates that it is a key point extracted from “Rise of Machine,” and the codes associated with it can be situated on the conceptual map under the code of “AI Safety” which has three subsequent concepts - Accountability of AI systems, Physical Safety, and Assessing AI Safety. Combined with other codes, it belongs to the code “Risks Related to Politics and Governance” which is under AI-related risks.

Secondly, the author clusters these key points with identifiers indicating different sources of data and in connection with AI-related risks and AI-related benefits based on the origins of these key points. In other words, the author tries to map the distribution of key points under the cluster of AI-related risks and AI-related benefits. By doing so, a conceptual representation of their distribution is reflected in the graphs below.

It is important to point out that these conceptual distributions of key points do cannot accurately mirror the reality because the selection of key points that are used for developing the conceptual map is based on the judgement of the author for preventing the entanglement of illustrations. More importantly, these graphs can only be utilized in a comparative manner. Nonetheless, two hypotheses can be drawn from these graphs.

1. In the case of China and India, the main focus of the development of AI is around the benefits of AI

2. In the case of Germany and US, the main focus is harder to identify.

For confirming the conceptual validity of these graphs, an exercise is further conducted. The testing of these hypotheses consist of two parts. Firstly, the author conducts a discourse analysis for understanding the different priorities on a relative term. Secondly, the author attempts to draw a connection with the reality by comparing with data provided by World Values Survey.

For China and India, the national priorities regarding the development of AI are straight-forward. In “A New Generation of Artificial Intelligence Development Plan,” the first paragraph explains the transformative nature of AI, Chinese aspirations, strategic objectives, and sets the main theme for the document.

“The rapid development of artificial intelligence will profoundly change human social life and the world. To seize the major strategic opportunities for the development of artificial intelligence, build China's first-mover advantage in artificial intelligence development, accelerate the construction of innovative countries and the world's science and technology power, this plan is enacted in accordance with the requirements of the CPC Central Committee and State Council.”

There are three linguistic expressions here indicating that China seeks to prioritize the benefits of AI: 1. Seizing opportunities for the development of artificial intelligence; 2. To build first-mover advantage in AI development; and 3. To accelerate the construction of innovative countries and the world's science and technology power. There is an absence of mentioning AI-related risks is apparent in perhaps the most important paragraph.

Similarly, a paragraph with a similar structural position in the Indian document states that:

“AI refers to the ability of machines to perform cognitive tasks like thinking, perceiving, learning, problem solving and decision making. Initially conceived as a technology that could mimic human intelligence, AI has evolved in ways that far exceed its original conception. With incredible advances made in data collection, processing and computation power, intelligent systems can now be deployed to take over a variety of tasks, enable connectivity and enhance productivity. As AI's capabilities have dramatically expanded, so have its utility in a growing number of fields.”

The paragraph highlights the transformative nature of AI with particular emphasis on its advances, task-solving capabilities, potentials for enhancing connectivity and productivity, and its expanding utility in its development. Positive words such as “exceed,” “incredible,” “enable,” and “enhance” are used continuous in the paragraph. Negative words are in absent along with mentions of AI-related risks.

Turning to the German and US documents, paragraphs positioned at similar places are different from that of China and India. In “Preparing for the Future of Artificial Intelligence,” the very first paragraph of the introduction states that

“Artificial Intelligence (AI) has the potential to help address some of the biggest challenges that society faces. Smart vehicles may save hundreds of thousands of lives every year worldwide, and increase mobility for the elderly and those with disabilities. Smart buildings may save energy and reduce carbon emissions. Precision medicine may extend life and increase quality of life. Smarter government may serve citizens more quickly and precisely, better protect those at risk, and save money. AI-enhanced education may help teachers give every child an education that opens doors to a secure and fulfilling life. These are just a few of the potential benefits if the technology is developed with an eye to its benefits and with careful consideration of its risks and challenges.”

The text reveals that, although it spends a great length portraying “the potential” of “Artificial Intelligence (AI),” the paragraph ends with a twist reminding the reader that there are risks that come with AI.

Similarly, in the German document, the first paragraph under the section “2. Current Situation” states that:

“Over the past few years, artificial intelligence has matured considerably and is becoming the driver of digitalisation and autonomous systems in all areas of life. The public sector, society, business, administration and science are all called upon to embrace the opportunities it provides and face up to the risks it poses. The Federal Government is seeking to actively shape AI in all areas of policy. The current progress in AI, particularly in the field of machine learning, is the result of exponential growth in hardware capabilities and the use of these capabilities to process large volumes of data. German research institutes have long been among the best in the world.”

Risks are also mentioned in here, but with a different focus. The message that the German document wants to be delivered to an audience- various agents that should be a part of the national AI strategy.

In sum, the author observes that differences in priorities are apparent in the data, which further strengthens the validity of the hypotheses.

Last but not least, the author tries to understand if the hypotheses are also valid in a real-life setting.

AI policy, similar to other areas of public policy, involves the process of policy making. In public policy and democratic theory, a vast amount of studies is conducted for theorizing the relationships between public opinion and government policies. Public opinion has impacts on government policies. In Robert Y. Shaprio's study examining public opinion and policy data for the US from 1935 to 1979 reveals that opinion changes are important causes of policy change Page, Benjamin I., and Robert Y. Shapiro. "Effects of public opinion on policy." American political science review 77, no. 1 (1983): 175-190.. A similar concept to public opinion is public perceptions. Sometimes what the public perceive may not always a reflection of reality. In a digital era, the boundaries between truth and lies become increasingly blur. For example, Blinder and Scott find that “public perceptions of immigration in Britain diverge dramatically from statistical immigration Blinder, Scott. "Imagined immigration: the impact of different meanings of `immigrants' in public opinion and policy debates in Britain." Political Studies 63, no. 1 (2015): 80-100.. For policy maker, it can indicate both greater or lesser constraint. On one hand, the erosion of public's capacity to impose democratic constraint Blinder, Scott. "Imagined immigration: the impact of different meanings of `immigrants' in public opinion and policy debates in Britain." Political Studies 63, no. 1 (2015): 80-100. in the digital era may lead to irresponsible decision making. On the other hand, in the aforementioned case of UK, “the government has explicitly claimed that its drive to reduce the number of immigrants coming to Britain is a response to public opinion Blinder, Scott. "Imagined immigration: the impact of different meanings of `immigrants' in public opinion and policy debates in Britain." Political Studies 63, no. 1 (2015): 80-100..

In discourse analysis, language is used to signal the relationship with the audience of the data. Recalling the German message “The public sector, society, business, administration and science are all called upon to embrace the opportunities it provides and face up to the risks it poses.”, it is safe to say that the civil society, and the public are included in the audience to which the general corpus seeks to address. Thus, how relevant is public opinion and public perceptions in AI policy? In other words, are public opinion and public perceptions reflected in the national AI policies? For understanding these concepts, the author examines the dataset provided by World Values Survey, and selects two questions used in its survey.

Question 1:

I'm going to read out a list of various changes in our way of life that might take place in the near future. Please tell me for each one, if it were to happen, whether you think it would be a good thing, a bad thing, or don't you mind?

"More emphasis on the development of technology"

And question 2:

Now, I would like to read some statements and ask how much you agree or disagree with each of these statements. For these questions, a 1 means that you “completely disagree” and a 10 means that you “completely agree”:

"Science and technology are making our lives healthier, easier, and more comfortable"

The graphs below demonstrate the results from China, US, Germany, and India for both questions

Regarding the first question - "More emphasis on the development of technology" - the majority of respondents from China, Germany, and India believe that it is a good thing. In the US, although the majority of the respondents believe that it is a good thing, people who simply do not care make up a close second. For the second question “Science and technology are making our lives healthier, easier, and more comfortable,” China stands out amongst the sampled question with the highest mean value. That indicates that respondents from China demonstrate a strong belief in technological developments for improving their quality of life.

In comparison to the conceptual graphs presented in the previous section, the author identifies many similarities. For instance, people in India and China show a stronger tendency towards techno-optimism, whereas people in Germany, although optimistic, demonstrates a greater level of distrust towards technology. Moreover, the respondents are the most skeptical towards technology. These tendencies suggest that another hypotheis can be conceived - Germany focuses cultivating AI-related benefits to a greater extend in comparison to the US.

So far, the first operationalizing the conceptual map can be divided into three parts. Firstly, key points are situated based on the connections they have with AI-related risks and AI-related benefits with their origins indicated resulting in comparatively conceptualized graphs. Form the graphs, two hypotheses are made in regard to AI policy:

1. In the case of China and India, the main focus of the development of AI is around the benefits of AI

2. In the case of Germany and US, the main focus is harder to identify.

After that, a discourse analysis is conducted, and further findings in the general corpus for gaining confidence. Lastly, the author tries to strengthens the hypotheses by comparing the them with real-life indicators. As a result, the validity of them is further enhanced with adding a new hypothesis that German AI policy focuses on AI-benefits more than that of the US.

#2 - The Case of UK

The second operationalization involves the used of conceptual map in a discourse analysis of the UK of which the documents are a part of the data used for grounded theory analysis and the development of subsequent conceptual map. In this operationalization demonstration, “Policy paper - AI Sector Deal,” is used as the data for discourse analysis.

A section under the title “People” in “AI Sector Deal” states that:

“People

The Industrial Strategy has people at its core: it is focused on creating good jobs and greater earning power for all people in the UK. To do this, we must equip citizens for jobs shaped by next generation technology. Growing the AI industry in the UK outlined the fast-growing demand for expertise to develop and apply AI technologies, and proposed ways to increase the supply of skills at different levels. Building on these recommendations and the commitments in the Industrial Strategy and Digital Strategy to grow science, technology, engineering and maths (STEM) and digital skills training, this Sector Deal sets out how the government, universities and industry will work together to greatly improve the supply of skills. It also sets out how we will attract the best, and most diverse, global AI talent to the UK.”