道德操守
尽量减少AI治理的最佳办法吗？
在 Google 上的首选来源中添加 Unite.AI
随着各组织将人工智能纳入其业务，它们需要明确的方法来管理与人工智能的开发和使用相关的风险。 有效的治理提供了监督，但复杂的程序会消耗资源并放慢执行速度。 最大限度地减少影响，通过限制不必要的数据使用，同时保持问责制和保障措施，提供了使治理更加相称的潜在途径。
为什么AI治理事项
AI治理确立了指导各组织如何开发、部署和使用AI系统的政策、责任和监督。 它帮助各组织查明风险，分配决策权并确立监测业绩和解决问题的程序。
各组织在采用大赦国际时必须考虑若干风险，包括侵犯隐私、安全脆弱性、有偏见的结果、产出不准确和透明度有限。 如果没有适当的控制，AI生成的决定可能会以难以纠正的方式影响客户，雇员和其他利益攸关方。 治理还帮助各组织满足适用的法律要求，并使AI的使用符合业务目标。
国际隐私专业人员协会报告说，2025年，77%接受调查的组织从事AI治理工作。 调查结果表明，在组织上对这一问题给予了很大关注，尽管在治理方面开展工作并不一定意味着一个组织已经完成或充分执行了其框架。
明确的责任和持续监测有助于各组织查明新出现的问题，调查出乎意料的结果并随着系统的变化而调整控制。 有效的治理还可以加强利益攸关方的信任，表明一个组织考虑AI使用的影响，并对其决定负责。
AI治理方面的挑战
虽然治理有助于各组织管理AI风险，但建立和维持有效监督可能很困难。 确定监督范围要求各组织确定哪些AI系统、活动和依赖性需要治理。
为了确定AI治理的明确范围，各组织可以参考ISO/IEC 42001，该标准规定了建立和维持AI管理系统的要求。 合规可以改善模型性能，允许更快，更安全的AI. 它还帮助各组织确定其业务中哪些部分属于AI管理系统。 这些标准考虑到地理覆盖范围、与其他系统的联系及其AI活动的范围。
其他主要挑战包括：
- AI系统的复杂性：AI应用程序可能依赖于互联模型，数据集，第三方工具和自动化决定。 这些依赖性可能使得难以追踪产出、查明脆弱性并确定需要控制的地方。
- 数据隐私和安全：各组织必须在整个数据收集、示范培训、部署和持续使用过程中保护敏感信息。
- 偏执和有限透明度：无代表性的培训数据会造成不公平的结果，而不透明的模型行为则会使决定难以解释、评估或质疑。
- 不断演变的条例和标准：跨法域运作的组织可能面临不同的或不断变化的法律要求。 治理进程必须顾及相关义务，而不会造成不必要的重复。
- 资源和业务限制：专家技能、文件、审查和持续监测需要时间和资金。
- 平衡监督和创新：过多的审批要求可能拖延项目，而薄弱的控制可能使重大风险得不到解决。
这些挑战突出了减少不必要的复杂性的价值。 然而，尽量减少治理还必须包括建立保护人员和信息的控制措施，并确保组织问责制。
例如，纽约州主计长发现纽约市各机构对大赦国际的监督采取不一致的做法，有些机构缺乏评估偏见和准确性的政策。 调查结果突出表明，要减少复杂性，就必须有明确、一致的治理标准来维持问责制。
AI治理中的最小化意味着什么
尽量减少人工智能治理涉及减少不必要的数据使用和行政间接费用，同时保留管理风险所需要采取的措施。 它有两个主要层面，包括尽量减少数据和尽量减少治理。 两者都旨在使大赦国际的做法与其宗旨和潜在影响更加相称。
数据最小化
数据最小化将信息的收集、保留、处理和共享限制在AI系统为确定的目的所需要的范围内。 各组织可通过限制获取敏感信息，并酌情使用匿名或合成数据来适用这一原则。
减少不必要的数据可以限制隐私接触并减少各组织需要保护的信息量。 但是，要尽量减少数据，需要认真判断。 删除相关信息可能会削弱模型的准确性，减少不同群体的代表性，或更难识别出有偏见的结果。 各组织必须确定为实现预期目的需要哪些数据，并支持可靠、公平的结果。
治理尽量减少
尽量减少治理的重点是精简监督程序。 各组织可酌情合并重叠审查和再利用评估。 它们还可以澄清决策权，并去除不增加有意义风险防护的审批层。
这些改变有助于减少行政费用和延误，同时澄清小组的责任。 然而，尽量减少治理不应取消基本文件、独立审查、人的监督或强制性控制。 目的是提高监督的效率，同时保持问责制。
尽量减少AI治理战略
如果各组织将数据做法和监督要求与每个使用案例的风险相匹配，最小化方法可以提高AI治理的效率。 若干措施可以减少不必要的复杂性，同时保持问责制、可靠的业绩和适当的保障措施。
采用基于风险的办法
各组织可以根据AI系统的预期用途、潜在影响和数据敏感性进行分类。 国家标准和技术研究所的AI风险管理框架通过包括Govern、地图、计量和管理在内的四项职能，提供了管理AI风险的结构。 各组织可以利用这些职能查明风险，评估其重要性并作出适当反应。
定义清晰的数据要求
每个人工智能系统在数据收集开始之前应有明确的目的。 然后，各组织可以确定必要的数据类别，去除无关的字段，限制访问和设定适当的保留限制。 测试应确认，数据集的减少继续有助于取得准确和公平的结果。
优化模型和计算资源
模型优化，高效算法和节能硬件可以帮助减少不必要的计算需要。 公司可以探索联合学习，这减少了大量信息传输的需要。 这些做法可以降低计算成本和能耗，同时支持数据隐私。
简化治理工作流程
各组织可以合并重叠审查、涵盖相同风险的再利用评估、明确责任并按风险水平确定核准途径。 标准化程序有助于各小组高效地处理日常应用程序，同时将更复杂的案件交给适当的专家。
保护基本保障措施
最小化办法必须保留隐私和安全保护、偏见评估、适当的人员监督、审计线索和事件应对程序。 各组织还应确认，简化程序继续履行适用的法律和监管义务。
监测性能和重新评估控制
随着AI系统的发展，对新用户的接触或对额外任务的支持，治理要求可能会发生变化。 各组织应在部署后监测准确性、公正性和隐私风险。 它们还应随着条件的变化而重新审查数据要求和监督程序。
使 " 尽量减少 " 为AI治理服务
尽量减少风险作为基于风险的治理战略的一部分最为有效。 然而，随着人工智能系统的目的、复杂性和潜在影响，适当程度的最小化程度将有所不同。 通过兼顾效率与适当的保障措施、问责制和持续评价，各组织可以简化AI治理，同时保持负责任和可靠的AI使用。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Ethics
Is Minimization the Best Approach to AI Governance?
Add Unite.AI to your preferred sources on Google
As organizations integrate artificial intelligence (AI) into their business operations, they need clear ways to manage the risks associated with its development and use. Effective governance provides oversight, but complex procedures can consume resources and slow implementation. Minimization offers a potential way to make governance more proportionate by limiting unnecessary data use while preserving accountability and safeguards.
Why AI Governance Matters
AI governance establishes the policies, responsibilities and oversight that guide how organizations develop, deploy and use AI systems. It helps organizations identify risks, assign decision-making authority and establish processes for monitoring performance and addressing problems.
Organizations must consider several risks when adopting AI, including privacy breaches, security vulnerabilities, biased outcomes, inaccurate outputs and limited transparency. Without appropriate controls, AI-generated decisions can affect customers, employees and other stakeholders in ways that are difficult to correct. Governance also helps organizations meet applicable legal requirements and align AI use with business objectives.
The International Association of Privacy Professionals reported that 77% of surveyed organizations were working on AI governance in 2025. The finding indicates substantial organizational attention to the subject, although working on governance does not necessarily mean an organization has completed or fully implemented its framework.
Clear responsibilities and ongoing monitoring help organizations identify emerging issues, investigate unexpected outcomes and adjust controls as systems change. Effective governance can also strengthen stakeholder trust by demonstrating that an organization considers the implications of AI use and remains accountable for its decisions.
Challenges in AI Governance
Although governance helps organizations manage AI risks, establishing and maintaining effective oversight can be difficult. Defining the scope of oversight requires organizations to determine which AI systems, activities and dependencies need governance.
To establish a clear scope for AI governance, organizations can refer to ISO/IEC 42001, which provides requirements for establishing and maintaining an AI management system. Compliance can improve model performance, allowing for a faster, safer AI. It also helps organizations determine which parts of their operations fall within their AI management system. These standards take into account geographic coverage, connections to other systems and the scope of their AI activities.
Other key challenges include:
- Complexity of AI systems: AI applications may rely on interconnected models, datasets, third-party tools and automated decisions. These dependencies can make it difficult to trace outputs, identify vulnerabilities and determine where controls are necessary.
- Data privacy and security: Organizations must protect sensitive information throughout data collection, model training, deployment and ongoing use.
- Bias and limited transparency: Unrepresentative training data can contribute to unfair outcomes, while opaque model behavior can make decisions difficult to explain, assess or challenge.
- Evolving regulations and standards: Organizations operating across jurisdictions may face different or changing legal requirements. Governance processes must accommodate relevant obligations without creating unnecessary duplication.
- Resource and operational constraints: Specialist skills, documentation, reviews and continuous monitoring require time and funding.
- Balancing oversight and innovation: Excessive approval requirements can delay projects, while weak controls may leave significant risks unaddressed.
These challenges highlight the value of reducing unnecessary complexity. However, minimizing governance must also include establishing controls that protect people and information and ensure organizational accountability.
For example, the New York State Comptroller found that New York City agencies took inconsistent approaches to AI oversight, with some lacking policies to assess bias and accuracy. The findings highlight how reducing complexity requires clear, consistent governance standards to maintain accountability.
What Minimization Means in AI Governance
Minimization in AI governance involves reducing unnecessary data use and administrative overhead while retaining the measures needed to manage risks. It has two main dimensions, which include data minimization and governance minimization. Both aim to make AI practices more proportionate to their purpose and potential impact.
Data Minimization
Data minimization limits the collection, retention, processing and sharing of information to what an AI system needs for a defined purpose. Organizations can apply the principle by restricting access to sensitive information and using anonymized or synthetic data where appropriate.
Reducing unnecessary data can limit privacy exposure and reduce the amount of information that organizations need to protect. However, data minimization requires careful judgment. Removing relevant information can weaken model accuracy, reduce representation of different groups or make it harder to identify biased outcomes. Organizations must determine which data is necessary to achieve the intended purpose and support reliable, fair results.
Governance Minimization
Governance minimization focuses on streamlining oversight processes. Organizations can consolidate overlapping reviews and reuse assessments where appropriate. They can also clarify decision-making authority and remove approval layers that do not add meaningful risk protection.
These changes help reduce administrative costs and delays while clarifying teams’ responsibilities. However, governance minimization should not eliminate essential documentation, independent review, human oversight or mandatory controls. The objective is to make oversight more efficient while maintaining accountability.
Minimization Strategies for AI Governance
A minimization approach can make AI governance more efficient when organizations match data practices and oversight requirements to the risks of each use case. Several measures can reduce unnecessary complexity while preserving accountability, reliable performance and appropriate safeguards.
Adopt a Risk-Based Approach
Organizations can classify AI systems according to their intended use, potential impact and data sensitivity. The National Institute of Standards and Technology’s AI Risk Management Framework provides a structure for managing AI risks through four functions that include Govern, Map, Measure and Manage. Organizations can use these functions to identify risks, assess their significance and determine appropriate responses.
Define Clear Data Requirements
Each AI system should have a defined purpose before data collection begins. Organizations can then identify the necessary data categories, remove irrelevant fields, restrict access and set appropriate retention limits. Testing should confirm that reduced datasets continue to support accurate and fair outcomes.
Optimize Models and Computing Resources
Model optimization, efficient algorithms and energy-efficient hardware can help reduce unnecessary computing demands. Companies can explore federated learning, which reduces the need to transfer large amounts of information. These practices can lower computing costs and energy consumption while supporting data privacy.
Simplify Governance Workflows
Organizations can consolidate overlapping reviews, reuse assessments that cover the same risks, clarify responsibilities and establish approval pathways by risk level. Standardized processes help teams handle routine applications efficiently while directing more complex cases to the appropriate specialists.
Preserve Essential Safeguards
A minimization approach must retain privacy and security protections, bias assessments, appropriate human oversight, audit trails and incident response procedures. Organizations should also confirm that streamlined processes continue to meet applicable legal and regulatory obligations.
Monitor Performance and Reassess Controls
Governance requirements may change as an AI system develops, reaches new users or supports additional tasks. Organizations should monitor accuracy, fairness and privacy risks after deployment. They should also revisit data requirements and oversight procedures as conditions change.
Making Minimization Work for AI Governance
Minimization is most effective as part of a risk-based governance strategy. However, the appropriate level of minimization will vary with an AI system’s purpose, complexity and potential impact. By balancing efficiency with appropriate safeguards, accountability and continuous evaluation, organizations can simplify AI governance while maintaining responsible and reliable AI use.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.