Emmie Hine & Luciano Floridi
Introduction
Artificial intelligence (AI) has recently become a focus of governments worldwide. AI is a “growing resource of interactive, autonomous, and often self-learning agency” with many applications and the potential to reshape society (Floridi and Cowls 2019; Hagerty and Rubinov 2019). Globally, the United States of America (US) and China are two of the most prominent players in AI development (Ding 2018; Savage 2020). Their dynamic is often framed as a “race” for AI supremacy (Savage 2020), which is concerning because the US and China are geopolitical rivals and military superpowers.
Both countries only recently defined national AI strategies: China in 2017 and America in 2019. Some work has examined China’s AI strategy, including (Allen 2019; Ding 2018; Roberts et al. 2019). Rasser et al. (2019) has looked at America’s. There is a dearth of both comparative work focusing on the vision endorsed by the strategy and of work examining local plans in China (Roberts, et al. 2021a, b). In assessing the AI approaches of the US, UK, and EU (before new developments in American AI policy), Cath et al. (2018) used the term “Good AI Society” to analyse the visions of AI-enabled societies endorsed in policy documents, which informs this analysis. Recently, Roberts et al. (2021a) compared the strategies of China and the EU, and Roberts et al. (2021b) compared those of the EU and US. Nevertheless, a gap still exists for a comparison of the strategies of China and the US. This article seeks to fill it.
Given the competitive dynamic of the US and China, it is vital to understand not only their approaches but also how they might interact with each other. In this work, we use a philosophy of technology level of abstractionFootnote1 to analyse the relevant AI policies. Geopolitical competition and economic levels of abstraction can inform us about international dynamics in the present moment. However, a philosophy of technology level of abstraction provides a historical grounding encompassing political and economic developments and offers a framework with which to analyse possible future dynamics.
Although private actors are significant players in AI development, this article focuses on government strategies, incorporating the private sector insomuch as the government delegates implementing AI strategy to it. In addition, while our positionality informs our ethical stances and normative judgements, we intentionally adopt a framing of ethical pluralism, which holds some values to be desirable while allowing latitude for others to be interpreted by different cultures in different places (Ess 2020). Finally, we adopt a mixed-methods approach that combines quantitative analysis of textual documents with a philosophy- and digital-ethics-informed evaluation. Our quantitative methods are based in natural language processing (NLP), a branch of computer science that uses computers to “understand and manipulate natural language text or speech” (Chowdhury 2003). NLP methods allow us to analyse term frequency and importance across languages, providing a more evidence-based foundation for our qualitative analysis. This unusual combination of methods provides a more objective grounding for textual analysis and points at directions for future work.
The article is structured into four more sections: Sect. 2 outlines the quantitative analysis and its conclusions, Sect. 3 focuses on the US, Sect. 4 on China, Sect. 5 discusses our findings, and Sect. 6 concludes the article. Tables of translated Chinese terms and words deemed significant for quantitative analysis can be found in the appendix.
Quantitative analysis
To provide a more objective grounding to our later documentary analysis, in this section, we will quantitatively analyse American and Chinese policy documents to identify differences between them. We will focus on the following questions: (a) how American and Chinese national documents compare in terms of sentiment and focus; (b) how local Chinese AI policy documents compare to national policy documents; (c) how American documents have changed over time.
Conclusions
Our quantitative analysis utilises term frequency-inverse document frequency (tf-idf), sentiment, and frequency analyses of AI policy documents from the US and China to help reveal their priorities. Tf-idf analysis uses the frequency of individual terms in a document relative to their frequency in a larger corpus to calculate a statistic indicating the relative importance of a word in a document. By identifying the top 20 words by tf-idf score in each document, we reveal that American documents have a broader focus than the Chinese documents. With garbled textFootnote2 removed, there are 206 unique terms in the top 20 tf-idf lists of the 16 American documents, an average of 12.9 per document. This is a similar number to the national Chinese documents (53 unique terms, 13.25 per document). However, the Chinese documents are generally more consistent in focus, with 7 words appearing in at least 75% of the documents; no words are mentioned that consistently in the American documents.
Our diachronic analysis of significant words between administrations reveal significant differences in American documents over time and across government branches. When comparing an Obama-era AI R&D plan with two Trump-era documents based on that plan, there was a statistically significant increase in certain terms used in rhetorically bombastic ways, including “American”, “leadership”, and “partnerships” in both Trump documents. In one Trump document, we saw a significant increase of the rhetorical terms “federal”, “partnership”, and “economy”; in the other, we saw a significant increase in the terms “R&D” and “innovation”. We generally saw decreases in technological terms, with “AI”, “internet”, and “system” decreasing in one and “development”, “research”, and “technology” decreasing in the other (though this could perhaps partially be explained by the increase in “R&D”). When comparing a Trump executive order (EO 13859, which established the American AI Initiative) with two Congressional documents from the Trump administration, we saw a statistically significant increase in focus on ethics-related terms and a decrease in rhetorical flourishes in the Congressional documents, indicating a disconnect between the rhetoric of the executive branch and the actions of the legislative branch.
Sentiment analysis identifies the “prevailing emotional opinion” of a text; the Google Cloud NLP API assigns a score value to a document indicating its sentiment and a magnitude value indicating how much emotional content a document contains (Natural Language API Basics 2021). By looking at these scores, we see that American documents are more balanced in terms of emotional sentiment and also more densely emotional than the Chinese documents. This makes sense considering that many of the words of focus (determined by tf-idf analysis) in American documents, especially Trump-era executive documents, are rhetorical flourishes, while the Chinese documents focus on terms related to industry, technology, and innovative development.
Frequency analysis reveals that undergirding the American documents’ emphasis on American technology and development is a competitive dynamic with China. While competition-related terms do not appear in the overall top focal words, China is mentioned in several of the documents, including in Obama’s “AI, Automation, and the Economy” (3 times, 0.013% of all words), the National Security Commission on Artificial Intelligence (NSCAI) report executive summary (6; 0.190%), and its accompanying “Full Report” (205; 0.081%). To put those numbers in context, the Obama report discusses how students in China have math abilities exceeding their American peers (Executive Office of the President 2016b), while in the NSCAI report—which focuses on helping the government become “AI-ready” in security—the rhetoric is explicitly competitive, with sentences like “China possesses the might, talent, and ambition to surpass the United States as the world’s leader in AI in the next decade if current trends do not change” (NSCAI 2021). China’s documents, on the other hand, barely mention America. The only national document that mentions America is the White Paper on AI Standardization, which mentions America 8 times (0.025% of all words), and usually in concert with the EU or Japan, or else when discussing the National Institute of Standards and Technology (NIST) as one of several institutions working on developing AI standards (China Electronics Standardization Institute 2020). Only two local documents mention America, but in reference to the Cleveland Medical Center and Utah State University (General Office of the People’s Government of Heilongjiang Province 2018; Guangdong Provincial Department of Science and Technology 2018).
China’s national documents are largely positive in sentiment and development-focused, with a particular emphasis on “innovation” (创新, chuangxin) and words related to technology applications. Its local documents are distillations of its national documents. Our tf-idf analysis reveals that local documents prioritise application- and innovation-related words more intently than the national documents, focusing on applying technology in their local contexts. Both “artificial intelligence” and “innovation” have higher average magnitude in local documents that mention them than the national documents, indicating a heavier focus in the local documents. From a documentary analysis perspective, many of the documents read very similarly, with emphasis on building “innovation centres,” achieving breakthroughs in “key core technologies”, and identifying “application scenarios.” To quantitatively compare these files to the seminal 2017 Three-Year Action Plan laying out important steps to achieve goals in the 2017 New Generation AI Development Plan, we programmatically identified similar phrases between the local documents and the Action Plan. Overall, there were over 400 substantial similarities identified. This revealed a significant amount of boilerplate language (ranging from section headers to ideological statements) but also some larger passages that were copied wholesale. Shenzhen and Nansha’s documents copy a section about healthcare R&D wholesale, while Guangzhou and Hubei plagiarise a section about building new industrial ecosystems. These and other similarities raise the question of how truly committed these localities are to the centralised Party view of AI. The documents could to some extent be a box-ticking exercise to demonstrate local loyalty to the Party, with little genuine commitment. However, the success of these plans—and China’s national plans—depends on the broader development atmosphere and, indeed, interactions with the United States. In the next two sections, we will qualitatively analyse American and Chinese policy documents before comparing the two country’s approaches. In doing so, we aim to identify the vision of a “Good AI Society” endorsed by the various documents, its level of cohesion, and what it may mean for the competitive dynamic between the two countries.
No comments:
Post a Comment