AI ethics and governance（AI 伦理与治理）¶

At Palantir, we believe that responsible AI is not an afterthought. Rather, it is fundamental to how we build technology. Our approach centers on developing software that enables responsible AI use throughout the entire system lifecycle, recognizing that ethical considerations extend beyond model development alone to encompass the full technology system, from data foundations and processing pipelines to user interfaces and human decision-making workflows. Rather than treating ethical considerations as separate compliance requirements, our platform integrates responsible AI capabilities throughout the development lifecycle.

This document outlines how Palantir's AI platform (AIP) actively supports eight core themes of responsible AI development. Each theme represents not just a principle, but a set of concrete capabilities and workflows designed to help users build trustworthy AI systems that are responsible, ethically sound, and effective in practice. This comprehensive approach ensures that users have the tools and workflows needed to address responsible AI requirements systematically, whether they are working with traditional machine learning models or generative AI systems.

Equitable: Fair, unbiased, non-discriminatory¶

AI systems should be inclusive and accessible and should not result in unfair discrimination against individuals or groups.

Platform capabilities¶

AIP offers features for identifying sources of bias in data, evaluating models for bias, and monitoring fairness concerns during model use.

Sensitive Data Scanner can automatically identify protected attributes and potential bias sources in datasets, enabling proactive assessment of fairness risks before they impact model training, evaluation, or use.
Subset evaluation in Modeling Objectives allows systematic evaluation of model performance across different groups within an evaluation dataset, surfacing potential disparate impacts that aggregate metrics might hide. For example, subset evaluation can be used to run model evaluation and easily compare model performance across demographic groups in an evaluation dataset.
AIP Evals allows you to evaluate the performance of AIP Logic functions across diverse test cases. With experiments in AIP Evals, users can further adjust input parameters to the model to understand how changes in certain inputs to the model might change the model's response.
Data health monitoring provides ongoing assessment of data representativeness and quality, helping identify when datasets may systematically under- or over-represent specific populations.

Implementation workflow¶

Teams can identify fairness risks early in their data preparation process using Sensitive Data Scanner and data health monitoring to detect protected attributes and imbalances in their data foundation before model development begins. This addresses potential bias at the data level. However, bias can also emerge from the model itself even with high-quality data, so both subset evaluation through Modeling Objectives and AIP Evals provides systematic testing to detect unequal model performance across demographic groups. When bias is detected at either level, teams can implement targeted mitigation strategies such as re-sampling, collecting additional representative data, or adjusting algorithmic approaches.

Explainable: Interpretable, understandable, transparent¶

AI systems should not be black boxes. Instead, to build trust in AI systems, users should understand how they work as much as possible.

Platform capabilities¶

AIP provides tools to help users understand how AI systems work, from debugging generative AI reasoning to evaluating traditional model performance.

The Modeling Objectives evaluation dashboard provides detailed model interpretability through feature importance analysis, performance breakdowns, and evaluation results that can be understood by both technical and non-technical stakeholders.
AIP Evals enables systematic testing and evaluation with both default and custom evaluation libraries, allowing teams to assess model behavior in ways that align with their specific domain requirements.
AIP Logic tools enable delegation of specific tasks to purpose-built, interpretable tools rather than relying solely on LLM processing, creating more explainable AI systems through composable, auditable components.
AIP Logic debug view provides visibility into chain-of-thought reasoning and tool orchestration within LLM-based systems, showing how the system delegates tasks and makes decisions through transparent handoffs to explainable components.
AIP observability delivers thorough monitoring and debugging capabilities across AI workflows, providing detailed execution traces, performance metrics, and system behavior insights that help teams understand and troubleshoot their AI systems.

Implementation workflow¶

Version control systems automatically capture model development decisions and rationale throughout the development process. For generative AI systems, the debug view in AIP Logic provides real-time visibility into how LLMs orchestrate tasks and delegate to explainable tools, while AIP observability delivers comprehensive execution traces that help teams understand system behavior. Teams can design more transparent systems by using AIP Logic tools to delegate specific tasks to interpretable components rather than relying solely on LLM processing. Testing and evaluation approaches through AIP Evals and Modeling Objectives complement these capabilities by presenting model performance metrics in formats accessible to both technical teams and business stakeholders.

Reliable: Safe, secure, resilient, robust¶

AI systems should be built with capabilities for assessing safety, security, and effectiveness throughout their entire lifecycle.

Platform capabilities¶

AIP enables secure, controlled deployment and continuous monitoring of AI systems to ensure safety and reliability throughout their lifecycle.

Model deployments for traditional AI/ML include direct deployments and Modeling Objectives live deployments, both providing controlled processes with options for automatic or manual upgrades depending on governance needs.
Functions versioning and release management for generative AI enables controlled deployment of AIP Logic through semantic versioning, backward compatibility checks, and version control workflows.
Rollback mechanisms allow immediate reversion to previous model versions when issues are detected, minimizing the impact of model failures or security breaches.
Comprehensive monitoring through inference history, system alerts, and continuous evaluation ensures that model performance degradation or security issues are detected quickly.
Access controls and data markings provide granular security restrictions based on user roles, data sensitivity, and geographic requirements, ensuring that AI systems respect privacy and security boundaries.
Georestrictions ensure that model requests and responses remain within specified jurisdictions, supporting regulatory compliance requirements across different regions and legal frameworks.
Encryption at rest and transit is a core part of our shared security responsibility model, with additional advanced capabilities for data protection that can be applied throughout the AI development lifecycle.
Capacity limits enable administrators to manage LLM usage and prevent service disruptions from unexpected load spikes, ensuring robust and stable AI workflows even under variable demand conditions.

Implementation workflow¶

Access controls and data markings establish security boundaries from the outset, with georestrictions ensuring that model requests and responses remain within compliant jurisdictions. Encryption at rest and in transit protects information throughout the AI development lifecycle. Model deployments and pre-release functions can enforce staged testing procedures before production release, while capacity limits prevent service disruptions from unexpected LLM usage spikes. Real-time monitoring systems provide continuous oversight across security, performance, and operational metrics, with rollback capabilities enabling immediate response when issues are detected.

Traceable: Auditable, governable¶

AI systems should provide capabilities to document relevant development processes, data sources, and the provenance of all data used for building models.

Platform capabilities¶

AIP automatically captures comprehensive documentation and audit trails across the AI development lifecycle, from data provenance to deployment decisions.

Data Lineage provides complete visibility into data sources, transformations, and dependencies used in AI systems.
Workflow Lineage provides visibility into how AI is used to power the logic and actions of applications for decision-making.
Audit logs capture all system interactions, model evaluations, and deployment decisions, creating a comprehensive audit trail for compliance and oversight.
Modeling Objectives documentation centralizes all project information, evaluation results, and decision rationales to enable multi-stakeholder collaboration.
Documentation templates in Notepad can enable standardized, detailed records of model purpose, methodology, and limitations that can be shared with regulators, supervisors, and other stakeholders.
LLM cost governance through Resource Management provides visibility into AI usage costs and resource consumption, enabling organizations to track, monitor, and manage expenses associated with LLM deployments.

Implementation workflow¶

Data Lineage automatically captures provenance information as data flows through processing pipelines, providing complete visibility into data sources and transformations. Workflow Lineage provides visibility into how AI powers application logic and decision-making workflows. Audit logs document all system interactions and decisions, while LLM cost governance tracks resource consumption and expenses, adding transparency to AI system operations. Documentation templates in Notepad and Modeling Objectives documentation enable teams to create standardized records that centralize project information, evaluation results, and decision rationale in formats suitable for regulatory review and internal audits.

Collaborative: Multi-stakeholder, interdisciplinary¶

Building AI systems should be an interdisciplinary process where scientists, engineers, domain experts, and other stakeholders work together.

Platform capabilities¶

AIP supports multi-stakeholder collaboration through flexible access controls, shared development environments, and evaluation frameworks accessible to users across different skill levels.

Role-based permissions allow organizations to configure access controls that match their governance structure, ensuring appropriate stakeholder involvement at each stage.
Code Workspaces and Code Repositories provide collaborative development environments that support both technical and non-technical contributors.
Workshop and collaborative analysis tools enable real-time collaborative data analysis, allowing stakeholders from different disciplines to work together on shared analyses and insights.
External data sharing and collaboration controls facilitate secure collaboration with external partners while maintaining governance oversight and data protection standards.
No-, low-, and pro-code evaluation frameworks accommodate different stakeholder skill levels, allowing domain experts to define custom evaluation criteria alongside standard technical metrics.

Implementation workflow¶

Role-based permissions and structured approval workflows create clear collaboration frameworks from project initiation. Code Workspaces and Code Repositories provide environments where technical and non-technical contributors can work together on AI development. Workshop tools enable real-time collaborative analysis across different disciplines, while external data sharing controls facilitate secure partnerships. Flexible evaluation frameworks ensure that domain experts, compliance officers, and technical teams each contribute their specialized expertise at appropriate points in the development process, rather than working in isolation.

Accountable: Liable, responsible¶

There should be clear definition of roles and workflows for people responsible for different parts of an AI system.

Platform capabilities¶

AIP establishes clear chains of responsibility through granular permissions, comprehensive audit trails, and structured approval workflows.

Granular permission management through groups and roles creates clear chains of responsibility for every aspect of AI development, deployment, and use.
Comprehensive audit trails document who made which decisions and when, enabling clear accountability for system outcomes.
Structured approval workflows through checks ensure that appropriate authorities review and approve critical decisions.
Checkpoints enable centralized acknowledgement and justification workflows that ensure stakeholders review and sign-off on critical AI-suggested decisions in operational workflows.

Implementation workflow¶

Granular permission management establishes clear accountability structures from the outset, defining who can take which actions throughout the AI lifecycle. Full audit trails automatically document decision-makers and their rationale, while structured approval workflows through checks and checkpoints create systematic review processes. Checkpoints specifically enable stakeholders to acknowledge and justify AI-suggested decisions in operational workflows. This creates transparent chains of responsibility that can be audited and verified without requiring additional manual tracking efforts.

Human-centered: Participatory, socially beneficial¶

AI systems should benefit individuals, society, and the environment overall. They should enhance rather than replace human decision-making.

Platform capabilities¶

AIP ensures that AI enhances rather than replaces human decision-making through structured decision support frameworks and mandatory human oversight mechanisms.

Ontology-based decision support provides structured frameworks for human-AI collaboration, ensuring that AI recommendations are presented in context that enhances rather than replaces human judgment.
Human oversight workflows through Ontology actions and approval processes ensure that critical decisions remain under human control while leveraging AI insights.
Dashboard and visualization capabilities present AI outputs in human-interpretable formats, enabling stakeholders to understand complex analytical results and make informed decisions based on AI recommendations.
Workflow automation with Checkpoints provides systematic approaches to automation that include mandatory human review points at critical decision stages, ensuring appropriate oversight while maintaining operational efficiency.
Opt-out and fallback mechanisms can be built into applications to ensure users retain control over AI-assisted processes.
Feedback loop integration enables continuous learning from human decisions to improve AI recommendations over time.

Implementation workflow¶

Ontology-based decision support frameworks present AI insights within structured workflows that preserve human agency and decision-making authority. Human oversight workflows through actions and approval processes ensure critical decisions remain under human control. Dashboard and visualization capabilities translate complex AI outputs into formats that enable informed human judgment, while workflow automation with human checkpoints ensures appropriate oversight at critical decision stages. Opt-out and fallback mechanisms can be designed into applications to ensure users retain control over AI-assisted processes. Feedback loop integration captures human decisions to continuously improve AI recommendations, creating a collaborative intelligence approach that enhances rather than replaces human expertise.

Getting started with responsible AI¶

Palantir's AI platform makes responsible AI systematic rather than ad hoc. The platform guides users of all skill levels and domains of expertise through established workflows that incorporate responsible AI principles:

Focus on the fully integrated system, not just its component tools: Consider your AI system holistically, from data foundations and processing pipelines through user interfaces and human decision-making workflows. Use capabilities like Data Lineage and Pipeline Builder to understand how components connect across your system.
Acknowledge technology's limits: Begin with a clear assessment of what your AI system can and cannot do, and what it should and should not be permitted to do. Use problem-first modeling approaches to define appropriate scope and limitations before development begins.
Do not solve problems that should not be solved: Evaluate whether your use case is appropriate for AI intervention. Consider legal, ethical, and community norms before initiating development. Some problems may be technically feasible but inappropriate for mathematical optimization.
Adhere to methodological best practices for sound data science: Leverage the platform's built-in evaluation frameworks, bias detection capabilities, and fairness assessment tools. Use Sensitive Data Scanner to identify protected attributes, employ subset evaluation to assess disparate impacts, and follow established methodologies for responsible feature usage.
Keep AI responsible, accountable, and oriented towards humans: Design AI systems that enhance rather than replace human decision-making. Use Ontology-based decision support, human oversight workflows, and feedback loops to ensure AI recommendations complement human judgment while maintaining clear accountability through audit trails and approval workflows.
Promote multi-stakeholder engagement: Configure checks and approval workflows that match your organization's governance structure. Use collaborative tools like Workshop and role-based permissions to ensure domain experts, compliance officers, technical teams, and other stakeholders contribute their expertise throughout the AI lifecycle.
Ensure technical, governance, and cultural awareness: Combine the platform's technical capabilities (encryption, access controls, monitoring) with governance frameworks (approval workflows, audit trails) and cultural practices (training, stakeholder engagement) to create inclusive responsible AI practices that fit your organizational context.

Conclusion¶

Responsible AI is not a constraint on innovation. Instead, it is what makes AI systems trustworthy enough to use for critical decisions. Palantir embeds responsible AI principles into every aspect of the development lifecycle, enabling organizations to build AI systems that are not just technically sophisticated, but ethically sound and operationally reliable.

By taking an integrated approach that considers the full context of AI deployment, we help our users solve their most challenging problems while enabling them to maintain the highest standards of responsibility and governance.

中文翻译¶

AI 伦理与治理¶

在 Palantir，我们坚信负责任的人工智能并非事后补救，而是构建技术的根本基石。我们的方法核心在于开发能够支持整个系统生命周期内负责任使用 AI 的软件，并认识到伦理考量不仅限于模型开发本身，而是涵盖整个技术系统——从数据基础和处理管道，到用户界面和人类决策工作流。我们并未将伦理考量视为独立的合规要求，而是将负责任 AI 的能力贯穿于整个开发生命周期。

本文档概述了 Palantir 的 AI 平台（AIP）如何积极支持负责任 AI 开发的八个核心主题。每个主题不仅代表一项原则，更是一套具体的能力和工作流，旨在帮助用户构建值得信赖的 AI 系统，使其在实践中既负责任、符合伦理，又切实有效。这种全面的方法确保用户拥有所需的工具和工作流，能够系统地应对负责任 AI 的要求，无论他们是在处理传统机器学习模型还是生成式 AI 系统。

公平性：公正、无偏见、无歧视¶

AI 系统应具有包容性和可访问性，不应导致对个人或群体的不公平歧视。

平台能力¶

AIP 提供了识别数据中偏见来源、评估模型偏见以及在模型使用过程中监控公平性问题的功能。

敏感数据扫描器 能够自动识别数据集中的受保护属性和潜在偏见来源，从而在偏见影响模型训练、评估或使用之前，主动评估公平性风险。
建模目标中的子集评估 允许在评估数据集中系统地评估模型在不同群体间的性能，揭示聚合指标可能隐藏的潜在差异性影响。例如，可以使用子集评估来运行模型评估，并轻松比较评估数据集中不同人口群体间的模型性能。
AIP 评估 允许您评估 AIP Logic 函数在不同测试用例下的性能。通过 AIP 评估中的实验，用户可以进一步调整模型的输入参数，以了解模型某些输入的变化如何改变模型的响应。
数据健康监控 提供对数据代表性和质量的持续评估，帮助识别数据集何时可能系统性地低估或高估特定人群。

实施工作流¶

团队可以在数据准备过程的早期，使用敏感数据扫描器和数据健康监控来识别公平性风险，在模型开发开始前检测数据基础中的受保护属性和不平衡。这从数据层面解决了潜在的偏见问题。然而，即使数据质量很高，偏见也可能来自模型本身，因此通过建模目标进行的子集评估和 AIP 评估都提供了系统性测试，以检测模型在不同人口群体间的不平等性能。当在任一层面检测到偏见时，团队可以实施有针对性的缓解策略，例如重新采样、收集更多代表性数据或调整算法方法。

可解释性：可解释、可理解、透明¶

AI 系统不应是黑箱。相反，为了建立对 AI 系统的信任，用户应尽可能理解它们的工作原理。

平台能力¶

AIP 提供了帮助用户理解 AI 系统工作原理的工具，从调试生成式 AI 推理到评估传统模型性能。

建模目标评估仪表板 通过特征重要性分析、性能分解和评估结果提供详细的模型可解释性，这些结果既可供技术利益相关者理解，也可供非技术利益相关者理解。
AIP 评估 支持使用默认和自定义评估库进行系统性测试和评估，使团队能够以符合其特定领域要求的方式评估模型行为。
AIP Logic 工具 允许将特定任务委托给专门构建、可解释的工具，而不是仅依赖 LLM 处理，从而通过可组合、可审计的组件创建更具可解释性的 AI 系统。
AIP Logic 调试视图 提供了基于 LLM 的系统中思维链推理和工具编排的可视性，展示了系统如何通过向可解释组件的透明交接来委派任务和做出决策。
AIP 可观测性 提供了跨 AI 工作流的全面监控和调试能力，提供详细的执行追踪、性能指标和系统行为洞察，帮助团队理解和排查其 AI 系统。

实施工作流¶

版本控制系统在整个开发过程中自动捕获模型开发决策和理由。对于生成式 AI 系统，AIP Logic 中的调试视图提供了 LLM 如何编排任务并委托给可解释工具的实时可视性，而 AIP 可观测性则提供全面的执行追踪，帮助团队理解系统行为。团队可以通过使用 AIP Logic 工具将特定任务委托给可解释组件，而不是仅依赖 LLM 处理，来设计更透明的系统。通过 AIP 评估和建模目标进行的测试和评估方法补充了这些能力，以技术团队和业务利益相关者都能访问的格式呈现模型性能指标。

可靠性：安全、有保障、有韧性、稳健¶

AI 系统应具备在其整个生命周期内评估安全性、保障性和有效性的能力。

平台能力¶

AIP 支持 AI 系统的安全、受控部署和持续监控，以确保其整个生命周期的安全性和可靠性。

传统 AI/ML 的模型部署 包括直接部署和建模目标实时部署，两者都提供受控流程，并根据治理需求提供自动或手动升级选项。
生成式 AI 的 Functions 版本控制 和发布管理 通过语义化版本控制、向后兼容性检查和版本控制工作流，实现对 AIP Logic 的受控部署。
回滚机制 允许在检测到问题时立即恢复到以前的模型版本，从而最大限度地减少模型故障或安全漏洞的影响。
通过推理历史、系统告警和持续评估进行的全面监控，确保快速检测到模型性能下降或安全问题。
访问控制 和数据标记 基于用户角色、数据敏感性和地理要求提供细粒度的安全限制，确保 AI 系统尊重隐私和安全边界。
地理限制 确保模型请求和响应保持在指定的司法管辖区内，支持不同地区和法律法规框架下的合规要求。
静态和传输中加密 是我们共享安全责任模型的核心部分，并提供了可在 AI 开发生命周期中应用的额外高级数据保护能力。
容量限制 使管理员能够管理 LLM 使用量，并防止因意外的负载峰值导致服务中断，确保即使在需求变化条件下也能实现稳健稳定的 AI 工作流。

实施工作流¶

访问控制和数据标记从一开始就建立了安全边界，地理限制确保模型请求和响应保持在合规的司法管辖区内。静态和传输中加密在整个 AI 开发生命周期中保护信息。模型部署和预发布函数可以在生产发布前强制执行分阶段测试流程，而容量限制可防止因意外的 LLM 使用峰值导致服务中断。实时监控系统提供跨安全、性能和运营指标的持续监督，回滚能力可在检测到问题时立即响应。

可追溯性：可审计、可治理¶

AI 系统应提供记录相关开发过程、数据源以及用于构建模型的所有数据来源的能力。

平台能力¶

AIP 自动捕获 AI 开发生命周期中的全面文档和审计追踪，从数据来源到部署决策。

数据沿袭 提供对 AI 系统中使用的数据源、转换和依赖关系的完全可视性。
工作流沿袭 提供对 AI 如何用于驱动应用程序逻辑和决策操作的可视性。
审计日志 捕获所有系统交互、模型评估和部署决策，为合规和监督创建全面的审计追踪。
建模目标文档 集中所有项目信息、评估结果和决策理由，以支持多方利益相关者协作。
Notepad 中的文档模板 可以启用标准化、详细的模型目的、方法和局限性记录，这些记录可以与监管机构、主管和其他利益相关者共享。
通过资源管理进行的 LLM 成本治理 提供了 AI 使用成本和资源消耗的可视性，使组织能够跟踪、监控和管理与 LLM 部署相关的费用。

实施工作流¶

数据沿袭在数据流经处理管道时自动捕获来源信息，提供对数据源和转换的完全可视性。工作流沿袭提供对 AI 如何驱动应用程序逻辑和决策工作流的可视性。审计日志记录所有系统交互和决策，而 LLM 成本治理跟踪资源消耗和费用，为 AI 系统运营增加透明度。Notepad 中的文档模板和建模目标文档使团队能够创建标准化记录，以适合监管审查和内部审计的格式集中项目信息、评估结果和决策理由。

协作性：多方利益相关者、跨学科¶

构建 AI 系统应是一个跨学科的过程，科学家、工程师、领域专家和其他利益相关者共同协作。

平台能力¶

AIP 通过灵活的访问控制、共享开发环境以及适用于不同技能水平用户的评估框架，支持多方利益相关者协作。

基于角色的权限 允许组织配置与其治理结构相匹配的访问控制，确保在每个阶段都有适当的利益相关者参与。
代码工作区 和代码仓库 提供支持技术和非技术贡献者的协作开发环境。
Workshop 和协作分析工具支持实时协作数据分析，允许来自不同学科的利益相关者在共享分析和洞察上共同工作。
外部数据共享和协作控制 促进与外部合作伙伴的安全协作，同时保持治理监督和数据保护标准。
无代码、低代码和专业代码评估框架 适应不同利益相关者的技能水平，允许领域专家在标准技术指标之外定义自定义评估标准。

实施工作流¶

基于角色的权限和结构化审批工作流从项目启动时就创建了清晰的协作框架。代码工作区和代码仓库提供了技术和非技术贡献者可以共同进行 AI 开发的环境。Workshop 工具支持跨不同学科的实时协作分析，而外部数据共享控制则促进安全合作伙伴关系。灵活的评估框架确保领域专家、合规官和技术团队各自在开发过程中的适当时点贡献其专业知识，而不是孤立地工作。

问责制：可追究责任、负责任¶

应为负责 AI 系统不同部分的人员明确定义角色和工作流。

平台能力¶

AIP 通过细粒度权限、全面审计追踪和结构化审批工作流建立清晰的责任链。

通过组和角色进行的细粒度权限管理 为 AI 开发、部署和使用的每个方面创建清晰的责任链。
全面审计追踪 记录谁在何时做出了哪些决策，从而为系统结果建立清晰的问责制。
通过检查（checks）进行的结构化审批工作流 确保适当的权威机构审查和批准关键决策。
检查点 支持集中化的确认和理由说明工作流，确保利益相关者在运营工作流中审查并签署 AI 建议的关键决策。

实施工作流¶

细粒度权限管理从一开始就建立了清晰的问责结构，定义了谁可以在 AI 生命周期的各个阶段采取哪些行动。完整的审计追踪自动记录决策者及其理由，而通过检查和检查点进行的结构化审批工作流则创建了系统性的审查流程。检查点特别使利益相关者能够在运营工作流中确认和证明 AI 建议的决策。这创建了透明的责任链，无需额外的手动跟踪工作即可进行审计和验证。

以人为本：参与性、有益社会¶

AI 系统应惠及个人、社会和整体环境。它们应增强而非取代人类决策。

平台能力¶

AIP 通过结构化决策支持框架和强制性人工监督机制，确保 AI 增强而非取代人类决策。

基于本体的决策支持为人类与 AI 协作提供了结构化框架，确保 AI 建议在增强而非取代人类判断的上下文中呈现。
通过本体操作和审批流程实现的人工监督工作流，确保关键决策在利用 AI 洞察的同时仍处于人类控制之下。
仪表板和可视化能力以人类可解释的格式呈现 AI 输出，使利益相关者能够理解复杂的分析结果，并根据 AI 建议做出明智的决策。
带有检查点的工作流自动化 提供了系统化的自动化方法，在关键决策阶段包含强制性的人工审查点，确保在保持运营效率的同时进行适当的监督。
选择退出和回退机制 可以内置到应用程序中，确保用户对 AI 辅助流程保持控制。
反馈循环集成 支持从人类决策中持续学习，以随时间改进 AI 建议。

实施工作流¶

基于本体的决策支持框架在保留人类能动性和决策权威的结构化工作流中呈现 AI 洞察。通过操作和审批流程实现的人工监督工作流确保关键决策仍处于人类控制之下。仪表板和可视化能力将复杂的 AI 输出转化为支持知情人类判断的格式，而带有检查点的工作流自动化则确保在关键决策阶段进行适当的监督。选择退出和回退机制可以设计到应用程序中，确保用户对 AI 辅助流程保持控制。反馈循环集成捕获人类决策以持续改进 AI 建议，创建一种增强而非取代人类专业知识的协作智能方法。

负责任 AI 入门¶

Palantir 的 AI 平台使负责任 AI 变得系统化而非临时性。该平台通过整合了负责任 AI 原则的既定工作流，引导所有技能水平和专业领域的用户：

关注完全集成的系统，而非仅关注其组件工具： 从整体上考虑您的 AI 系统，包括数据基础和处理管道，以及用户界面和人类决策工作流。使用数据沿袭和 Pipeline Builder 等功能来理解组件如何在您的系统中连接。
承认技术的局限性： 首先清晰评估您的 AI 系统能做什么和不能做什么，以及它应该和不应该被允许做什么。在开发开始之前，使用问题优先的建模方法来定义适当的范围和限制。
不要解决不应解决的问题： 评估您的用例是否适合 AI 干预。在启动开发之前，考虑法律、伦理和社区规范。有些问题可能在技术上可行，但不适合进行数学优化。
遵循可靠数据科学的方法论最佳实践： 利用平台内置的评估框架、偏见检测能力和公平性评估工具。使用敏感数据扫描器识别受保护属性，采用子集评估评估差异性影响，并遵循负责任特征使用的既定方法论。
保持 AI 负责任、可问责且以人为本： 设计增强而非取代人类决策的 AI 系统。使用基于本体的决策支持、人工监督工作流和反馈循环，确保 AI 建议补充人类判断，同时通过审计追踪和审批工作流保持清晰的问责制。
促进多方利益相关者参与： 配置与您组织治理结构相匹配的检查和审批工作流。使用 Workshop 和基于角色的权限等协作工具，确保领域专家、合规官、技术团队和其他利益相关者在 AI 生命周期中贡献其专业知识。
确保技术、治理和文化意识： 将平台的技术能力（加密、访问控制、监控）与治理框架（审批工作流、审计追踪）和文化实践（培训、利益相关者参与）相结合，创建适合您组织背景的包容性负责任 AI 实践。

结论¶

负责任 AI 并非对创新的约束。相反，它正是使 AI 系统值得信赖、可用于关键决策的原因。Palantir 将负责任 AI 原则嵌入到开发生命周期的每个方面，使组织能够构建不仅在技术上复杂，而且在伦理上合理、在运营上可靠的 AI 系统。

通过采用考虑 AI 部署完整背景的集成方法，我们帮助用户解决其最具挑战性的问题，同时使他们能够维护最高标准的责任和治理。