Case Illustration: Computational Analysis
Understanding Policies (10700193-90)
Tsinghua University
Overview
- Big Models, Big Data
- From the 6 Vs to LLMs
- Computational analysis: description, relationship, pattern
- Policy on data — amplified by AI/LLMs
- Who owns the data (and the model)
- Who accesses it (and the inference pipeline)
- Who uses it to do — and to generate — what
Big Models, Big Data
The political meaning of the “Big-Data” (Computational) approach
大数据发展日新月异，我们应该审时度势、精心谋划、超前布局、力争主动，深入了解大数据发展现状和趋势及其对经济社会发展的影响，分析我国大数据发展取得的成绩和存在的问题，推动实施国家大数据战略，加快完善数字基础设施，推进数据资源整合和开放共享，保障数据安全，加快建设数字中国，更好服务我国经济社会发展和人民生活改善 (习近平 2017)。
What’s Big Data?
Big data usually includes data sets with sizes beyond the ability of commonly used software tools to capture, curate, manage, and process data within a tolerable elapsed time.
From the 6 Vs to LLMs
Big data’s 6 Vs
- Volume — petabytes of records
- Variety — text, image, audio, video
- Velocity — streaming, real-time
- Veracity — noise, bias
- Value — signal for decisions
- Variability — drift over time
What LLMs do with them
- Volume → trillions of tokens for pre-training
- Variety → multimodal models (text + audio + vision)
- Velocity → inference at conversational speed
- Veracity → alignment / RLHF to curb bias
- Value → emergent capabilities at scale
- Variability → continual fine-tuning, retrieval augmentation
Big data is the fuel; the architecture (e.g., transformer) is the engine.
Quiz: Data-based advantages?
“I established a CPC elite database containing extensive biographical and career information for over 20,000 positions with a unique number of 4,700 cadres, including all the members of the Central Committee and Provincial Standing Committee of the CPC from 1982 to 2020……”
“To provide a further test of the classic arguments on democracy and public support, we generated estimates of democratic support……assembling as much survey data on democratic support as possible. We employed 4,905 national opinions on democracy from 1,889 national surveys, representing a 32.0% and 37.3% increase respectively over the 3,716 opinions and 1,376 national surveys used in Claassen (2020a; 2020b).”
“We prompted an open-source LLM to read every provincial Government Work Report from 2000 to 2024 and return structured labels on policy priorities — producing a 300,000-paragraph, human-validated corpus that would have taken a research team of ten over a year to code by hand.”
Policy Analysis w. big data and big models
Sources
- Data from e-government, civil service hotlines, government-citizen interactions
- Official data from bureaus and departments
- Speeches, statements
- Traditional and new media
- Official document archives
- Judicial documents
- Multimodal policy data — audio & video transcribed at scale
- LLM-processed and LLM-generated corpora ……
Goals
- Initiation of the agenda
- Profile of policy makers
- Policy making process
- Logic of policy decisions and implementation
- Prediction of policy outcomes
- Policy evaluation ……
Methods
- Description (夯)
- Relationship analysis (拉)
- Pattern learning
Description
Description → relationship (Hu et al. 2022)
Description → patterns (Hu 2020)
Beyond words
Policy on data
AI/LLMs do not replace the three classical questions
—they amplify each of them.
Issues in the AI/LLM Era
- Who own the data (and the model weights)
- Who can access the data (and the inference pipeline)
- Who can use the data to do and generate what
Who owns the data: Data Responsibility
Who owns the training data?
The LLM era extends “whose data” from stored records to model memory:
- NYT v. OpenAI (2023–): did OpenAI lawfully ingest copyrighted news to train GPT? (Pope 2024)
- China’s “奥特曼案”: the 2024 Guangzhou Internet Court ruling held that an AI image service infringed when its outputs reproduced a licensed character (初萌 2021)
- 生成式AI训练语料合规:《生成式人工智能服务管理暂行办法》requires providers to use “data with lawful source” and to respect personal-information and IP rights (生成式人工智能服务管理暂行办法 2023)
Once data are absorbed into model weights, ownership becomes provenance.
Who can access the data: information management
LLM-era access: new gates, new leaks
New “data” to access
- Model weights (open vs. closed)
- Training corpora
- Inference APIs & rate limits
- Embedding / vector stores
New ways access can fail
- Prompt injection → exfiltration of system prompts or tools
- Membership inference → recovering training records
- Jailbreak → bypassing safety and policy controls
- Shadow use → staff pasting confidential docs into public chatbots
Access control used to end at the database. Now it extends to every prompt and every response.
Who can use the data to generate what
Who can use the data to generate — and impersonate — whom
Laws and policies of information management
2017.6.1实施
2021.9.1实施
2021.11.1实施
2013.9.1实施
2021.9.1实施
2025征求意见稿
2025.5.1实施
2023.8.15实施
2023.1.10实施
2025.9.1实施
Policies of information protection
Take-Home Points
- Big Models, Big Data
- 6 Vs → the fuel of LLMs
- Computational analysis: description (夯), relationship (拉), pattern
- Policy on data — amplified in the AI/LLM era
- Whose data (and whose model weights / training corpora)
- Who accesses data (and the inference pipeline)
- Who uses data to do and generate what
Big data feeds big models; the three classical questions now cover training, deployment, and generation.