Reading objectives and introduction
This is an English translation of the published reading. Based on the principles of proportionality, privacy, fairness, and transparency, we continue to understand trustworthy AI, compare two governance frameworks, and consider benefits, risks, and responsibilities.
Trustworthiness is not a marketing label. The original source uses three aspects from the EU guidelines to illustrate: legality, ethical compliance, and robustness. Robustness encompasses both technical and the social environment in which the system operates.
1. Behavioral Guidelines for the Hiroshima AI Process
The original document describes the Hiroshima AI initiative led by the G7, aiming to promote the safety, assurance, and trustworthy use of advanced AI systems. The G7 consists of Canada, France, Germany, Italy, Japan, the United Kingdom, and the United States.
Related international behavioral guidelines for organizations developing advanced AI systems emphasize eleven actions and a risk-based approach. While reading, you may consider answering:
- What safety, assurance, or societal risks are explicitly identified in the document? Please identify three.
- What measures does it recommend for developers or enterprises? Identify at least two and explain which risk each corresponds to.
This page retains the reading guide and accompanying transcript. When you need to review the reading guidelines, you may proceed directly. The complete English translation of the Eleven Actions , which respectively outlines risk, testing, reporting, governance, and data controls.
II. EU Trustworthy AI Ethics Guidelines
The original material introduces basic rights, ethical principles, seven key requirements, and issues regarding the handling of conflicts between different rights or principles.
The core knowledge content of the guide has been divided into five sections as originally structured: Abstract, Introduction, and Framework 、 Rights and Four Principles 、 Seven Requirements and Methods 、 Comprehensive Evaluation Checklist 、 Opportunities, Concerns, and Terminology .
Think while reading: What are the three fundamental components of trustworthy AI? What do the seven requirements individually signify? How does the document recommend handling conflicts among the principles?
It differs from the readership of the Hiroshima Criteria in that it is not limited to developers, but also addresses deployers, decision-makers, and other stakeholders. Its broader scope enables cross-role discussions, but practical implementation still requires translation into specific actions.
Three, maintain caution regarding autonomous capabilities
The original page also mentioned a public research interview discussing why stronger autonomous systems should be advanced with caution. The existing archives do not contain the full text of this video for verification, and therefore the spoken content is not included here, nor are specific arguments inferred from the summary.
Can be combined Agent's Tools and Oversight Boundary After thinking ability improves, why must control, error correction, and responsibility also keep pace?
Video script: Principle review and two frameworks
Previously, ethical issues were divided into four categories. First, review the principle of proportionality: when costs and benefits are considered together, is there a sufficient justification for using AI? Prompting people to seriously consider this issue forms the foundation of many frameworks.
The relevant content also includes respect for fundamental rights, such as freedom of expression, protection from arbitrary detention, protection from deception, or harm based on unfounded allegations. These are not issues that can simply be offset by efficiency.
There must also be a remedy mechanism. If an AI makes a lending or other opportunity-related decision, and a party believes it is unfair, can they raise an objection? Who would independently handle such objections? The fact that the decision is technically generated does not justify the elimination of inquiry and correction.
System shutdown and product discontinuation are also within consideration. What if the service provider disappears? How should the data and related equipment be handled? These should have plans in place, not just plans for initial deployment.
Privacy and Data Protection
Privacy and data protection are of long-term importance and were previously discussed in the context of large-scale data collection. The applicable rules may vary by country and region, so it is first essential to understand the specific conditions under which you operate.
Generally, data minimization should be considered: only collect the information necessary to achieve the stated purpose. The sharing and access of data by third parties should be clearly explained; if data is sold, used for other revenue-generating purposes, or incorporated into products, users should not be left unaware.
At the same time, protecting data from unauthorized manipulation and modification is also part of this scope. It is not sufficient to only concern oneself with whether initial consent was obtained and to ignore subsequent accesses and uses.
Fairness and Expected Users
Fairness requires equal opportunity and also requires defining the product's target audience. Requiring every technology service to reach every person worldwide is not necessarily feasible; however, within a reasonably defined user group, it is essential to ensure that people of different races, religions, genders, and abilities are treated fairly.
There may be variations in the regulations regarding protected features across different regions, therefore it is necessary to consider applicable conditions. During development and testing, efforts should be made to carefully review data and results; achieving absolute consistency is unlikely, but differences should be proactively identified, documented, and explained, along with recommendations for improvement.
Some restrictions can have legitimate justifications. For example, a product designed exclusively for adults does not necessarily need to be accessible to children. The key issue is whether the restriction is justified, not whether it uses a vague "intended user" to exclude those who are not served.
Transparency, Explanations, and Human Control
Why use AI, who is responsible for oversight and control, and do users know that AI is involved? The latter point is particularly important, because some interfaces are not easily discernible.
From the perspectives of public governance and independent oversight, appropriate third-party audits will also be considered. Companies may have concerns regarding intellectual property and commercial confidentiality, but at a minimum, they should provide necessary human controls, accountability, and remedial mechanisms.
The original explanation then compares two international guidance documents. Its description of regulation still in development pertains to the context at the time and should not be interpreted as indicating that any jurisdiction today is unconstrained or exempt from existing laws.
Hiroshima AI Process
The G7 includes Canada, the United States, the United Kingdom, France, Germany, Italy, and Japan. It is a forum for government-level policy coordination, different from international organizations that have different formal structures.
In 2023, the Hiroshima AI Initiative responded to advances in generative capabilities and foundational models by proposing international behavioral guidelines for organizations developing advanced AI systems. It is built upon principles such as those from the OECD AI Principles, and primarily addresses organizational development, rather than serving as a unified operational manual for all individual user behaviors.
The OECD divides principles into five groups, though each group may include multiple aspects. The first group links well-being, sustainable development, and inclusive growth, covering both economic, environmental, and social dimensions. The second group emphasizes human rights and democratic values, including privacy and fairness.
Group three focuses on transparency and explainability. The original explanation notes that these terms overlap, and different authors use them in varying ways. Transparency can address whether users know they are interacting with AI and whether they have sufficient information to understand the interaction and ask questions; explainability can further examine how data, algorithms, and specific outcomes are connected.
When reading, use the definition provided by the document itself, and do not assume the terms need to be distinguished simply because they are often used together.
Group Four: Resilience, Safety, and Protection: Does the system function, is it safe, and is it protected? Group Five: Responsibility and Accountability. These are interrelated requirements, not optional elements that can be selected individually.
EU Trustworthy AI Guidelines
Another document is the EU Trustworthy AI Ethics Guidelines from 2019. Developed by a relevant expert group, it is intended for developers, deployers, policymakers, and other stakeholders, offering a broader readership than the narrower focus of the Hiroshima Criteria, which primarily targets developers.
A broad scope has advantages, but it also requires further translation into roles and responsibilities during actual implementation; related supporting materials can help facilitate on-the-ground execution.
The original explanation lists seven requirements: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination, and fairness; social and environmental well-being; and accountability.
You'll notice that these themes share many commonalities with the OECD, merely differing in ordering and grouping. Social and environmental impacts appear later in one document and possibly at the forefront in another; accountability appears at the end, which does not imply it is unimportant.
Next, by reading two materials, compare their similarities and differences, and consider how the target audience and recommendations influence actual adoption. Only through this can you gradually understand what "trustworthy AI" specifically means in different contexts.
Related Reading
Specific examples of the four principles are seen in Four Principles of AI Ethics ; Different definitions of action and responsibility can be further explored Intelligence and agency .
Reference and Extended Reading
The following is the official framework's query entry point; the translation scope of this page remains the current introductory text and accompanying script.
- EU: Ethics guidelines for trustworthy AI : English, query three components, ethical principles, and seven requirements.
- European Commission: Hiroshima AI Process Behavior Guidelines : English, inquiry about code of conduct and related reporting mechanisms.
- OECD AI Principles : English, for comparing the common themes of the framework.