ChinAI #246: The State of Large Model Governance in China
A Blue Paper Report on how ideas and policies are being implemented in practice
Greetings from a world where…
Christmas markets in Berlin are just around the corner
…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: Blue Paper Report on Large Model Governance (2023): From Rules to Practice
Context: Thanks for voting for your favorite picks from last week’s Around the Horn issue (ChinAI #245). At this point, I should no longer be surprised when ChinAI readers pick the most technical and “boring” (at least, from a fast-news perspective) option. Still, it’s pretty cool that what won out was this blue paper report on the implementation details of China’s large language model governance, a joint publication (link to original) by the China Academy of Information and Communications Technology (CAICT) and the Institute of Computing, Chinese Academy of Sciences.
Key Takeaways: Let’s start with implementation details on the algorithm filing registry. To file with the algorithm registry, entities have to give detailed information about their algorithm and include other information such as an algorithm safety self-assessment report. For more background on this tool, see Matt Sheehan and Sharon Du’s essential piece.
As of November 2023, the Cyberspace Administration of China has issued a total of 151 algorithm registry numbers, 100 of which are for generative AI algorithms. Within the generative AI subcategory, 54 percent of the algorithms were text generators.
Blue paper report also adds important background context for the “AI safety assessments” in China’s large model regulations. These extend back to requirements in regulations published back in December 2017 directed at all internet information services with public opinion properties or social mobilization capacity. Some of China’s early generative AI regulations were interpreted as proactive measures targeted at a broad notion of AI safety, whereas they should probably have been seen as an extension of measures established way back in 2017 that were targeted at information content security (read: censorship).
Now, let’s turn to the implementation status of large model safety assessments. There has been a lot of momentum in this area, and the blue paper report lists a few key examples:
Tsinghua’s CoAI team launched a Chinese large model safety assessment platform (The blue paper links to their website, which I found pretty impressive). CAICT has its own set of benchmarks. And the Beijing Academy of Artificial Intelligence has also established an open platform to evaluate model safety.
Enduring issues in this space, according to the report: “There are problems such as too many evaluation standards, serious problems with “swiping scores”[刷分], and large differences in evaluation results.” The report gives a specific example for the variation in safety benchmarks: a certain large model ranked near the top in a capability assessment list by reputable consulting company IDC, but it ranked relatively low (10th) on the SuperCLUE list. See ChinAI #237 for more on SuperCLUE’s safety benchmarks.
Implementation details for origin tracing of large language models to identify misuses of AI-generated content.
Chinese companies have used digital watermark technology to protect their AI-generated content. The report gives details on Alibaba’s efforts in this space, which includes adding such watermarks to their Tongyi Wanxiang text-to-image generator and business services such as Tongyi QIanwen.
The report also collates a lot of research from Chinese institutions on tools and methods for detecting machine-generated content. I’ve included the footnotes to arxiv papers that describe this research, including work by Peking University, Huawei, Harbin Institute of Technology, etc.
Going forward, on the topic of improving China’s large model governance system, the report emphasizes the need for international cooperation.
One important issue it mentions is the interoperability of AI safety and evaluation standards at the international level.
The report concludes: “It is recommended to actively promote cooperation in AI research, widely bring together AI experts from various countries, and jointly explore testing and evaluation methods on the basis of respecting the cultural diversity, political security and other demands of all parties, and assist late-developing countries to jointly reduce the risks of large-scale model technology.”
A lot more details and annotations in CHAPTER 5 and 6 TRANSLATION: Blue Paper Report on Large Model Governance (2023): From Rules to Practice
ChinAI Links (Four to Forward)
Must-read: Horizon’s emergingtechpolicy.org
Very cool to see this new website launched by The Horizon Institute for Public Service, co-founded by Joan Gass and Remco Zwetsloot. Aimed at people interested in the intersection of public service and emerging tech issues, the site aims to make high-quality information and tactical advice about policy and public service careers more accessible and inclusive. I will definitely be sharing the original in-depth guides on policy internships with my students!
Should-read: Assessing China's AI Workforce: Regional, Military, and Surveillance Geographic Job Clusters
By Dahlia Peterson, Ngor Luong, and Jacob Feldgoise, this CSET analysis analyzes 58,229 job postings as a way to identify trends in China’s technical AI workforce. One finding that stood out to me: within this sample, there were only 79 AI job opening advertised across 9 defense-affiliated universities.
Should-read: ‘A mass assassination factory’: Inside Israel’s calculated bombing of Gaza
A +972 and Local Call investigation into the Israeli army’s expanded authorization for bombing non-military targets in Gaza and the role of AI in that decision calculus:
According to the investigation, another reason for the large number of targets, and the extensive harm to civilian life in Gaza, is the widespread use of a system called “Habsora” (“The Gospel”), which is largely built on artificial intelligence and can “generate” targets almost automatically at a rate that far exceeds what was previously possible. This AI system, as described by a former intelligence officer, essentially facilitates a “mass assassination factory.”
Should-read: GM’s Self-Driving Car Unit Skids Off Course
A WSJ deep dive into significant safety issues with GM’s self-driving car unit Cruise, reported by Ryan Felton, Meghan Bobrowsky, and Mike Colias. Some gripping details, including an anonymous letter sent to a California regulator from a Cruise employee claiming his submission of safety risks through an internal reporting systems hadn’t been addressed after six months.
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).
Also! Listen to narrations of the ChinAI Newsletter in podcast format here.
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99