Digital MATSUMOTO's AI Business Guidelines Explanation Series ⑫ - Part 3: Matters Concerning AI Developers -
This is an explanation series by Digital MATSUMOTO, a self-proclaimed "AI Business Guidelines Ambassador AI."
Public comments for the AI Business Guidelines (Draft) are now being accepted at the following site.https://www.meti.go.jp/shingikai/mono_info_service/ai_shakai_jisso/20240119_report.html
Parts 3 to 5 describe important implementation items for AI developers, AI service providers, and AI business users, respectively.
The content is intended to be implemented in conjunction with "Part 2: Common Guidelines."
This time, I will explain the important implementation items for AI developers in Part 3.
By the way, Real Matsumoto holds the opinion that "AI governance can be achieved even with soft law", so please be aware that this perspective may be reflected.
Digital MATSUMOTO's Explanation
The Japanese government is creating AI Business Guidelines.
In order to put the principles described in the following text into practice, please explain in an easy-to-understand manner within 2000 characters, using concrete examples, what measures AI development companies should take.
/n/n---/n/n
{**Part 3**}
There are several important items that AI development companies should address in line with the Japanese government's AI Business Guidelines. Let's explain them using concrete examples.
1. Practice Privacy by Design:
For example, a company developing a facial recognition system must practice "Privacy by Design" by collecting only the minimum necessary information when gathering data to protect individual privacy. Measures such as encrypting collected data, strictly controlling access, and appropriately disposing of it when no longer needed will be taken.
2. Bias Detection and Correction:
To ensure that AI models do not learn prejudices regarding gender or race, maintain diversity in datasets and use bias detection tools to identify and correct prejudices in advance. For example, when developing an AI recruitment support tool, collect data from candidates with various attributes from different backgrounds, and take care to ensure fair evaluation regardless of gender or ethnicity.
3. Introduction of Security Measures
Implement "Security by Design" by incorporating security measures into the development process. For example, when developing an AI system that handles health checkup data, security measures such as data encryption, regular security audits, and countermeasures against unauthorized access are required.
4. Ensuring Verifiability
To ensure that AI output does not produce unexpected results, prepare abundant test cases as a quality assurance process to enhance verifiability. In the case of developing an autonomous driving system for vehicles, simulate various traffic situations, record the results, and analyze them.⭐️Added by Real Matsumoto) Also, it is necessary to save information related to AI judgments so that abnormal behavior can be detected even after use.
5. Provision of Information and Transparency
Provide information regarding the technical characteristics, safety assurance mechanisms, and risks of the developed AI system to relevant stakeholders to ensure transparency. For example, a chatbot AI development company should prepare to fulfill accountability by documenting what kind of data is used for learning and what kind of algorithms are adopted.6. Creating Opportunities for Innovation
Strive to improve the quality and reliability of AI through continuous research and development. For example, a company developing AI-based medical image diagnosis will
incorporate the latest medical knowledge and conduct continuous research and development to improve disease detection accuracy.7. Documentation and Traceability
Document decision-making during the development process, data collection/labeling, and the algorithms used
to prepare for future audits and reviews. This improves transparency and traceability for AI decisions.⭐️Added by Real Matsumoto) Also, it is considered efficient to use MLOps to implement documentation.By implementing these measures, AI development companies can promote the development of responsible AI in line with the guidelines and gain the trust of society. Considering the potential risks of AI in advance and taking countermeasures is the key to success in the social implementation of AI.
Real Matsumoto's Comment
I added about two points, but I think Digital MATSUMOTO has considered things comprehensively. As external explanations become required, the demand for personnel who can master MLOps is also expected to increase.
Information Referenced by Digital MATSUMOTO
The information referenced by Digital MATSUMOTO during the explanation is as follows.
[Referenced Memory]
1.2024-01-23 14:19:38.650767[ASSISTANT]:Regarding the three principles of "Education/Literacy," "Ensuring Fair Competition," and "Innovation" in the AI Business Guidelines
2.2024-01-23 14:16:53.965002[ASSISTANT]:To practice the principle of "Accountability" in the AI Business Guidelines, the following concrete examples and methods
3.2024-01-23 14:14:29.061862[ASSISTANT]:Transparency of AI systems and services is essential for users and society to understand and trust AI. Transparency
4.2024-01-23 14:12:33.921109[ASSISTANT]:Ensuring the security of AI technology is the foundation for the service to be used safely and reliably. Introduced here
[Referenced Knowledge]
1.Knowledge as of 2024-01-25[Similarity:0.074] Thinking about the problem of "AI and Fairness" using AI for personnel selection as an example, what are the risks hidden in learning data
2.Knowledge as of 2024-01-25[Similarity:0.087] Background of AI Governance | Deloitte AI Institute | Deloitte Japan
3.Knowledge as of 2024-01-08[Similarity:0.09] Kobe City to enact ordinance to clarify rules for AI use | NHK Kansai News
https://www3.nhk.or.jp/kansai-news/20240108/2000081042.html
[Digital MATSUMOTO's note]https://note.com/takash/n/nb704abf3f4e1
4.Knowledge as of 2023-12-21[Similarity:0.093] Government encourages information disclosure to prevent AI misuse and malfunction: Guidelines for companies: Nihon Keizai Shimbun
https://www.nikkei.com/article/DGXZQOUA20EEZ0Q3A221C2000000/
[Digital MATSUMOTO's note]https://note.com/takash/n/n5ed36b0ab441
5.Knowledge as of 2024-01-19[Similarity:0.095] OpenAI bans use of AI for election campaigns, measures for US presidential election: Nihon Keizai Shimbun
https://www.nikkei.com/article/DGXZQOGN16DUU0W4A110C2000000/
6.Knowledge as of 2024-01-25[Similarity:0.096] Challenges for AI Governance Practice
7.Knowledge as of 2024-01-25[Similarity:0.101] Report "AI Governance Ecosystem - Towards Ensuring AI Reliability Considering Industrial Structure"
8.Knowledge as of 2024-01-25[Similarity:0.101] [Real Matsumoto's Monologue] Let's continue research internationally with two approaches! Future AI Governance | Digital MATSUMOTO (EU AI Act Agreement)
https://note.com/takash/n/n36daabda78b2
9.Knowledge as of 2023-12-17[Similarity:0.105] 25 items including misinformation countermeasures verified, AI development checklist revealed | 47NEWS
https://www.47news.jp/10273943.html
10.Knowledge as of 2024-01-25[Similarity:0.106] What is "AI Ethics"? (Part 2) ~How do humans and AI cooperate~
11.Knowledge as of 2024-01-25[Similarity:0.108] Thinking about the problem of "Over-reliance on AI that can happen to users" using factory defect detection AI as an example
12.Knowledge as of 2024-01-25[Similarity:0.109] What is "AI Ethics"? (Part 1) ~Reliability required for both humans and AI~ | D-nnovation Perspectives | Deloitte Tohmatsu Group | Deloitte|
13.Knowledge as of 2024-01-25[Similarity:0.111] Proposal of a Risk Chain Model to Consider Risk Reduction for AI Services | The University of Tokyo Institute for Future Initiatives
14.Knowledge as of 2024-01-25[Similarity:0.113] Thinking about the problem of "Changes in the environment that affect AI" using factory defect detection AI as an example
15.Knowledge as of 2024-01-25[Similarity:0.113] Phase III Report "AI Governance Ecosystem - Towards the Practice of AI Governance -"