Discover more from ChinAI Newsletter
ChinAI #193: All ChinAI Policy is Local
Emmie Hine breaks down local AI plans
Greetings from a world where…
this chronicling of a legendary soccer goal reminds me of the good old days of Grantland
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: Shenzhen New Generation Artificial Intelligence Development Action Plan (2019-2023)
*This week’s translation comes from Emmie Hine, a software engineer who received her MSc in Social Science of the Internet at Oxford in 2021 and will enter a PhD program in Law, Science, and Technology at the University of Bologna in November. What follows is Emmie’s analysis, drawn from her Master’s research.
Context: After China’s 2017 New Generation AI Development Plan (AIDP) and Three-Year Action Plan for Promoting the Development of a New Generation of Artificial Intelligence Industry (2018-2020) (Action Plan), provinces, districts, special economic zones, and even cities began issuing their own implementations of the Action Plan. All documents outline AI development goals for the specific area, but read very similarly to each other and to the Action Plan. Under China’s “federated authoritarianism” model of managing large projects, these local plans will provide a foundation for national AI initiatives as the best local approaches bubble to the top.
As part of my master’s thesis (a version of which is published open access here), I quantitatively and qualitatively analyzed 28 of these local documents, as well as the most prominent national documents. I used tf-idf analysis (which calculates the importance of a term based on the occurrence of a word in a text relative to its occurrence in the broader corpus) to identify the most significant terms of focus in these documents, then compared them to see what they can tell us about local AI development efforts.
This translation is of Shenzhen’s “New Generation AI Development Action Plan (2019-2023)”. The plan itself (link to original Chinese) is not especially remarkable, but is representative of the local plans in that it contains lofty goals, lots of discussion of AI applications, and boilerplate language.
Key takeaways: What are these plans focusing on?
This document outlines goals for Shenzhen in the AI industry, such as quantitative benchmarks for the industry by 2023. It outlines seven tasks, including strengthening basic research, expanding “smart application scenarios,” and analyzing risks while developing “forward-looking ethical regulations and standards.” Each of these goals contain subgoals, such as accelerating smart medicine as a “smart application scenario,” and the unit responsible for its development. These tasks overlap heavily with the Action Plan, but add some Shenzhen specificity, such as leveraging the National Supercomputing Shenzhen Center and local universities and industrial areas.
This plan is similar to many other local plans, which also outline somewhat lofty goals for local industry, areas of focus, and steps to achieve them.
When comparing top words by tf-idf score to the AIDP and Action Plan, local documents focus more heavily on the words “apply/application,” “innovation,” “smart/intelligent,” and “industry,” at a potential cost of “R&D” and basic research. According to Tse & Wang (2017), while focusing on application scenarios can lead to short-term successes, fundamental research is necessary for long-term breakthroughs. Under current guidance, though, success in specific applications is more heavily rewarded (despite advances in basic research being a goal in many of the documents and mentioned in the AIDP).
How serious are these plans?
It’s difficult to gauge the intentions behind the plans, but we can look at statement lengths, repetition, and plan feasibility to get some idea of whether the goals are genuine.
Document lengths (without punctuation) range from ~1600 characters to ~23000 characters; Shenzhen is above average at ~12300 characters (median ~8040 characters). Shorter documents may indicate that less effort was put into creating the documents and developing realistic and beneficial goals.
I noted above that the structure of the Shenzhen document (and many of the other documents) is similar to that of the Action Plan. This in itself doesn’t mean much—the Action Plan provides a good framework to work off of—but plagiarism might, as it shows that the officials responsible for writing the new document didn’t feel the need to elaborate or adapt the Action Plan to their local contexts. Using a 70% character match threshold, I identified 525 similarities between the local documents and the Action Plan, including several long passages; Shenzhen’s document (along with Nansha’s) copies wholesale a 32-character section on intelligent medical imaging systems (see Box 5 in the Shenzhen document and section 2D in the Action Plan).
Finally, there’s plan feasibility. AI industry targets in local plans sum to more than double the national industry target, so not every area will hit their targets. Furthermore, setting goals around the number of AI companies settling in an area may promote numbers fudging, or incentivize the creation of companies trading on the latest hot topic but lacking a serious business plan. In the 2019 “capital winter,” 336 start-ups shut down. As Jeff noted in ChinAI #191, AI company establishment peaked in 2015 with 1089 new companies and crashed to just 57 companies in 2021. Some of this is likely to do with the pandemic, but the decline began before that, with only 326 AI companies established in 2019. As the industry matures, more investment is going to B and C round financing (rather than seed/A round financing, which happens early in a company’s life). If the industry is solidifying around existing actors, local officials may have to re-evaluate their goals around entrepreneurism and attracting newly founded companies. This also indicates that it may be even harder for areas that aren’t thought of as tech hubs to achieve their goals.
Who are the winners and losers?
It’s difficult to find data on what areas are meeting (or not meeting) their goals, but Shenzhen appears to be making good progress, with a core industry value of 14.3 billion yuan in 2020 projected to increase to 20.3 billion yuan in 2021, and announced that Accenture will be building an “innovation hub” in the area.
Other wealthy areas are also doing well: Hunan seems to have achieved its goal of a 10 billion RMB industry in 2020, and Guangdong and Shenzhen have been successful in establishing industry partnerships. Less-wealthy Heilongjiang, on the other hand, is struggling—a report by the Jiusan Society indicated that lack of R&D capacity and infrastructure, coordination difficulties, and investment shortfalls were interfering with its goal of achieving a 5 billion RMB AI industry by 2020.
Other funding patterns reinforce this trend; 71% of Artificial Intelligence Industry Alliance members are in first-tier cities. Beijing, Shanghai, Jiangsu, and Guangdong received the vast majority of NSFC funding in 2019, indicating that it may be difficult for areas that aren’t established research hubs to access capital and industry resources.
Amidst a vast shortage of AI workers (over 5 million by some estimates), wealthier provinces are likely better equipped to attract workers and companies with financial and quality-of-life benefits.
What’s the benefit of this kind of analysis?
I want to remark a bit on the mix of methods used in this paper. Quantitative analysis can’t tell us everything about a document, but it can identify interesting themes and allow us to compare documents to each other. There are many other things that can be done with quantitative methods, including diachronic analysis (which I did on American documents), sentiment analysis, and more advanced NLP methods.
Though I’ve grown to love philosophy and the social sciences, the computer scientist in me still craves a more objective evidentiary basis for qualitative claims, which quantitative analysis provides. I’d love for more social scientists to incorporate this kind of analysis into their work, but more work needs to be done to teach critical skills (like coding) and break down the barriers between fields.
ChinAI Links (One to Open)
This fall, I’m teaching a course on “Emerging Technologies, AI, and International Politics.” As I prep the course, I wanted to share the reading list and open it up to suggestions and comments (I’m trying to keep the reading load relatively light, as the class is targeted at undergrads). Please follow along if you’re interested, and email me for a pdf if you can’t get access to any of the readings.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at email@example.com or on Twitter at @jjding99