返回 Papers
AI 扩展计划 / Playbooks

AI Fairness / Fair Lending / Bias Control Playbook

以下官方来源是本手册的锚点。使用方式不是复制条文, 而是把监管和标准语言转成产品需求、控制目标、架构边界、评估设计、监控指标和 evidence binder。

737AI_FAIRNESS_FAIR_LENDING_BIAS_CONTROL_PLAYBOOK.md

AI Fairness / Fair Lending / Bias Control Architecture Playbook

定位: 面向 CBAP+、AI Product Manager、Product Architect、Risk Product Lead、Solutions Architect、Model Risk Partner 和金融零售 AI 治理负责人。本手册不是基础的多元化宣言, 而是把 AI fairness / fair lending / bias control 设计成可需求化、可架构化、可评估、可监控、可审计、可挑战的产品控制体系。

重要说明: 本文是学习、作品集和企业方案设计材料, 不构成法律意见、合规结论、模型验证报告或监管解释。正式项目必须由 Legal、Compliance、Fair Lending、Model Risk、Privacy、Security、Business Owner、Data Owner、Operations、Vendor Risk、Internal Audit 和管理层结合机构类型、司法辖区、产品范围、客户影响、数据授权和内部政策确认。


1. Source Anchors

以下官方来源是本手册的锚点。使用方式不是复制条文, 而是把监管和标准语言转成产品需求、控制目标、架构边界、评估设计、监控指标和 evidence binder。

AnchorOfficial link本手册使用方式
NIST AI Risk Management Frameworkhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织 fairness risk lifecycle: 识别上下文、测量差异影响、管理控制、治理责任。
NIST SP 1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligencehttps://www.nist.gov/publications/towards-standard-identifying-and-managing-bias-artificial-intelligence把 bias 视为 socio-technical risk, 覆盖数据、制度、流程、人工、部署和反馈循环, 不把偏差简化成训练集问题。
CFPB Regulation B / ECOA, 12 CFR Part 1002https://www.ecfr.gov/current/title-12/chapter-X/part-1002用于 fair lending、adverse action、specific reasons、record retention 和信贷流程控制设计。特别关注 1002.2、1002.6、1002.9、1002.12 等与定义、规则、通知和记录相关的要求。
Interagency Joint Statement on Automated Systems Discrimination and Biashttps://www.ftc.gov/legal-library/browse/cases-proceedings/public-statements/joint-statement-enforcement-efforts-against-discrimination-bias-automated-systems用于提醒团队: 使用 automated systems 不会消除现有反歧视法律责任, vendor / algorithm / proxy / black-box 不能成为控制缺失的理由。
EU AI Act, Regulation (EU) 2024/1689https://eur-lex.europa.eu/eli/reg/2024/1689/oj用于连接 high-risk AI、fundamental rights、risk management、data governance、technical documentation、logging、human oversight、accuracy、robustness、post-market monitoring 等控制要求。

学习纪律:

  • 官方来源进入 Source Anchor -> Internal Policy -> Control Objective -> Requirement -> Test -> Evidence 链路。
  • Fairness 不等同于某个数学指标, 也不等同于删除 protected attributes。
  • Fair lending 相关判断必须由合规和法律团队确认, 产品与架构团队的职责是把事实、边界、数据、模型、流程、控制和证据讲清楚。
  • 对跨司法辖区金融零售 AI, 应同时考虑信贷监管、消费者保护、隐私、模型风险、第三方风险、安全、记录留存、投诉、申诉和运营韧性。

2. Core Mental Model: Fairness 是产品和架构控制问题

2.1 一句话

AI fairness in financial retail is not a single model metric. It is a product and architecture control system that governs who is affected, which decision is influenced, what data and proxies are used, how segment outcomes are measured, how humans override, how customers are explained to, and how evidence is retained.

高级团队不要问: “这个模型公平吗?” 要问: “这个 AI 系统在这个业务流程、客户群、决策权、数据来源、人工复核、解释机制、监控窗口和残余风险边界下, 是否具备可证明的公平控制?”

2.2 从 metric thinking 到 control thinking

错误心智为什么危险高级心智
删除 race / gender / age 就公平proxy 变量可能恢复敏感属性, 历史标签可能已有偏差, 人工流程也可能不公平建立 protected attribute governance、proxy risk control、segment evaluation 和 human review calibration
AUC 很高所以可以上线总体性能会掩盖 subgroup error, 少数群体样本少时更容易被平均值吞掉总体性能只是入口, 必须看 segment-level performance、harm severity 和 customer impact
Fairness 是 Data Science 的事产品定义了目标、流程、阈值、例外、解释、人工介入和客户补救Fairness requirement 是 PM / BA / Architect / Risk 共担的系统要求
有人工复核就安全人工可能 rubber stamp, 也可能把历史偏差重新引入系统设计 reviewer skill、blind review、override reason、calibration、sampling、QA 和 drift feedback
Vendor 模型声明负责公平金融机构仍要理解使用场景、输入输出、限制、测试、监控和客户影响Vendor evidence 要进入 model / vendor challenge memo, 机构保留 deployer accountability
一个 fairness dashboard 就够生产环境会发生 population shift、policy change、model routing change、channel mix changeDashboard 必须连接 release gate、alert threshold、issue workflow、model change 和 governance decision

2.3 Control Plane 视角

Fairness control plane 应覆盖七个问题:

  1. Scope: 哪些产品、客户、渠道、地区、决策和解释受影响。
  2. Data: 哪些数据、标签、特征、文档和第三方信号进入系统。
  3. Proxy: 哪些变量可能间接代表 protected class 或 vulnerable segment。
  4. Decision: AI 输出如何影响 eligibility、pricing、prioritization、fraud flag、collections、service、marketing 或 explanation。
  5. Human: 人工如何复核、覆盖、升级、纠错、补救和记录理由。
  6. Monitoring: 上线后如何按 segment、channel、language、product、model version 和 reviewer group 监控。
  7. Evidence: 如何证明控制被设计、测试、批准、运行、挑战和改进。

3. Fairness Scope: 金融零售中哪些决策要纳入

Fairness scope 不能只覆盖贷款审批。金融零售 AI 会在多个客户旅程中改变机会、价格、摩擦、服务质量和解释质量。

ScopeAI 可能做什么Fairness / Fair Lending 风险控制重点
Eligibility预审、授信、额度、账户开立、产品资格、补件判断不同 protected group 通过率、误拒、补件负担差异approval / decline / counteroffer rate, false negative disparity, policy reason traceability
Pricing利率、手续费、折扣、风险定价、优惠券、动态报价同等风险客户获得不同价格, proxy-driven pricingpricing feature governance, monotonicity, rate exception review, segment APR / fee distribution
Prioritization客户队列排序、人工复核优先级、线索优先级某些群体等待更久、人工接触更少、申诉更慢SLA parity, queue fairness, sampling control, workload balancing
Fraud / AML flags欺诈预警、交易拦截、KYC 复核、AML case triage误报造成账户冻结、交易拒绝、服务中断, 且集中在特定群体false positive disparity, friction rate, escalation quality, investigation calibration
Customer serviceChatbot 路由、工单优先级、情绪识别、投诉识别语言、口音、渠道、残障、年龄导致服务质量差异language / locale eval, accessibility, handoff, complaint coding parity
Collections催收优先级、联系频率、话术建议、困难客户识别脆弱客户被过度联系, 某些群体补救方案不足contact policy constraints, hardship detection QA, vulnerable customer safeguards
MarketingOffer targeting、cross-sell、pre-screen、lookalike audienceredlining、steering、排除特定社区或渠道客户targeting exclusions, audience composition, offer exposure parity, consent controls
Explanationsadverse action reason、信用解释、客服说明、申诉材料草稿原因不具体、错配真实决策、不同语言解释质量不同reason-code fidelity, explanation consistency, plain language, translation QA

Scope decision 的实用规则:

  • 只要 AI 改变客户获得产品、价格、服务、补救、人工关注、解释或申诉机会, 就纳入 fairness scope。
  • 只要 AI 输出被员工当作事实、排名、风险分数、建议动作或客户话术, 就纳入 control plane。
  • 即使 AI 只是 assistant, 如果它系统性影响人工判断, 也需要 segment evaluation 和 human review monitoring。

4. Bias Sources: 偏差不是只来自训练数据

NIST SP 1270 的重要启发是: bias 是 socio-technical problem。金融零售中, 偏差来自数据、制度、业务策略、流程、人工、供应商、部署和 UX 的组合。

Bias source典型表现金融零售例子控制设计
Data generation数据来自历史流程, 历史流程已经不平衡某社区线下网点少, 历史申请量低, 模型把低触达误认为低需求数据 lineage, channel coverage analysis, community / branch / digital access review
Labels标签反映历史人工决策, 不一定是真实风险过去 underwriter 更严格审某类申请, 拒绝标签被模型学习label audit, outcome label vs decision label separation, adjudication sample
Proxies看似中性的变量间接编码 protected classZIP code、学校、语言、设备、工作行业、居住稳定性、渠道偏好proxy-risk register, adversarial predictability test, business necessity review
Missingness缺失本身不随机, 代表渠道、收入、语言或可得性自雇收入材料更复杂, immigrant applicant 文档格式不同missingness by segment, alternative documentation policy, imputation governance
Feedback loops模型影响后续数据, 造成自我强化低额度导致低使用, 低使用再被解释为低价值客户shadow evaluation, exploration policy, outcome correction, causal review
Human review人工复核吸收或放大模型偏差reviewer 看到模型分数后倾向同意, override 只发生在熟悉客户blind review where feasible, calibration, override parity, reviewer QA
Model routing不同客户进入不同模型、不同 prompt 或不同 vendor英语客户走新模型, 西语客户走旧模型; 高价值客户走人工队列routing inventory, segment routing parity, fallback quality monitoring
Vendor model第三方模型训练数据、限制和变更不可见欺诈模型供应商升级后误报集中上升vendor challenge, model change notification, independent backtesting
Language / locale多语言理解、翻译、语气和合规措辞差异adverse action explanation 翻译后原因变模糊, 客服 bot 不懂方言表达multilingual eval, localized policy review, translation consistency checks
UX界面设计改变申请完成率、申诉率和材料提交率移动端上传收入证明失败率更高, 导致某些客群补件率高journey analytics, accessibility testing, device / channel completion disparity

高级 PM / Architect 要把 bias source 写进需求, 不是上线前让 Data Science 做一次报告。正确表达是:

For this use case, fairness risk can enter through historical labels, proxy features, missing documentation, model routing, reviewer behavior, language support, and post-decision feedback. Therefore, our requirements include protected attribute governance, proxy detection, segment evaluation, human review calibration, monitoring, and an evidence binder.


5. Requirements Architecture: 把公平性写成可验收需求

5.1 Fairness requirement 的结构

一个可执行的 fairness requirement 至少包含十个字段:

Field问题示例
Decision contextAI 影响哪个决策或流程Credit prequalification ranking for unsecured personal loan
Customer impact客户可能受到什么影响获得预审邀请、需要补件、被拒绝、价格更高、等待更久
Protected / sensitive groups哪些受保护或敏感维度要评估race / ethnicity proxy, sex, age where legally appropriate, language, channel, geography
Metric family用哪些指标观察差异disparate impact, false negative disparity, calibration, SLA disparity
Baseline / comparator与谁比较qualified applicant pool, existing policy model, champion model, human-only process
Threshold / trigger什么结果触发 reviewratio below agreed threshold, confidence interval breach, unexplained month-over-month increase
Control response触发后怎么处理feature review, threshold adjustment, manual review expansion, retraining freeze, release hold
Evidence用什么证明eval report, segment matrix, feature review, approval memo, monitoring snapshot
Owner谁负责Product owner, Fair Lending, Model Risk, Data Science, Operations
Review cadence何时复核pre-release, monthly monitoring, quarterly governance, major model / policy change

5.2 Requirement patterns

Pattern可直接写进 PRD / Architecture Decision Record 的要求
Scope boundaryAI system must not make final credit approval, denial, pricing, or adverse action notification unless explicitly approved through fair lending, model risk, legal, and business governance.
Protected data separationProtected-class data or proxy estimates used for fairness testing must be stored, accessed, and logged separately from production decisioning features, unless legal and compliance approve another design.
Segment evaluationEvery high-impact release must include segment-level performance and outcome analysis across approved protected-class, proxy, geography, language, channel, and product slices.
Proxy reviewFeatures with high protected-class predictability, geographic concentration, unexplained disparate impact, or weak business necessity must enter a documented proxy-risk review.
Explanation fidelityGenerated explanations must be traceable to actual decision factors, policy rules, model reason codes, or approved source documents; LLM-generated explanations cannot invent reasons.
Human reviewReviewers must receive calibrated instructions, record override reasons, and be monitored for override disparity, reviewer drift, rubber-stamping, and segment-level outcomes.
Monitoring triggerProduction monitoring must alert on segment disparity, model drift, routing disparity, explanation defects, complaints, appeal outcomes, and abnormal override patterns.
Vendor challengeVendor fairness claims must be challenged against institution-specific data, use case, thresholds, protected-class governance, change notices, and production monitoring requirements.

5.3 Non-functional requirements

Fairness 也要进入 NFR:

NFR dimensionFairness-specific question
Explainability客户、员工、审计和监管能否理解 AI 如何影响结果, 以及哪些因素真正驱动结果
Observability是否能按 segment、channel、language、model version、policy version、reviewer、vendor version 追踪
Privacyprotected attribute / proxy estimate 是否最小化、隔离、授权、留痕和保留期限受控
Securityfairness data、adverse action data、appeal evidence 是否被严格访问控制
Resilience如果 fairness monitor 异常、vendor model 改版或 segment performance 坍塌, 系统如何降级或暂停
Auditability是否能重建某次客户影响、模型版本、输入、输出、人工动作、解释、通知和证据
Accessibility不同语言、设备、残障、渠道和数字能力客户是否能获得同等可用流程

6. Bias Control Architecture

6.1 Reference architecture

Use case intake
-> Customer impact and legal scope assessment
-> AI / model / vendor inventory
-> Data lineage and feature inventory
-> Protected attribute governance and proxy estimation for testing
-> Proxy-risk analysis and feature challenge
-> Segment evaluation and fairness metric pack
-> Policy constraints and decision guardrails
-> Human review design and calibration
-> Release gate and risk acceptance
-> Production monitoring and alerting
-> Issue remediation, customer correction, model change control
-> Evidence binder and audit / regulator response

6.2 Protected attribute handling

Protected attribute handling 是公平性架构中最容易被误解的部分。高级方案不是简单使用或删除, 而是按用途分区。

Zone用途访问控制关键原则
Decisioning zone实时决策、评分、路由、定价最小权限, 严格特征白名单默认不把 protected attribute 作为决策特征, 任何例外必须有正式法律和治理批准
Fairness testing zone离线评估、差异影响分析、segment slicingFair lending / model risk / analytics 授权访问可使用 legally permissible protected data 或 approved proxy estimates 做测试
Evidence zone保存评估结果、审批、监控截图、issue log审计和治理访问保留 aggregate evidence, 控制个人级敏感数据暴露
Customer response zoneadverse action、申诉、纠错、客户沟通客服 / 运营按职责访问解释必须来自真实决策依据, 不把敏感推断暴露给不应知角色

设计要点:

  • Data minimization: 只收集、生成和保留达到 fairness testing / compliance 目的所需的数据。
  • Purpose limitation: 区分 monitoring / audit / validation 与 production decisioning。
  • Access segregation: 能做 fairness analysis 的团队不一定能改生产模型; 能改模型的团队不一定能看个人级 protected data。
  • Lineage: 记录 protected attribute 来源、采集依据、估算方法、刷新频率、质量限制和使用范围。
  • Retention: 与 Reg B、模型风险、内部政策、诉讼保全和隐私要求对齐。

6.3 Proxy detection

Proxy risk control 的目标不是机械删除所有相关变量, 而是识别哪些变量可能导致不正当差异影响, 并证明业务必要性、替代方案和控制措施。

Technique用途输出
Correlation / mutual information初筛特征与 protected attribute 或 proxy estimate 的统计关系high-correlation feature list
Protected-class prediction model测试特征集合是否能预测 protected classproxy predictability score, feature contribution
SHAP / feature attribution by segment查看特征在不同 segment 中的影响差异segment-specific driver comparison
Geography and channel heatmap识别 ZIP / census tract / branch / digital channel 的集中影响redlining / steering review input
Missingness analysis分析缺失率和缺失含义是否按 segment 不均衡missingness disparity report
Counterfactual / sensitivity testing测试客户在非核心变量变化下结果是否不合理变化robustness and proxy sensitivity memo
Less discriminatory alternative review比较替代特征、阈值、策略或人工流程feature decision and residual risk rationale

Proxy-risk register 至少记录:

  • feature / signal / vendor score 名称
  • 业务用途和业务必要性
  • 与 protected attribute / proxy estimate 的关系
  • 对结果的边际影响
  • 可替代变量或策略
  • 监控指标
  • owner 和批准记录
  • 保留、限制、移除或替代的决定理由

6.4 Segment evaluation

Segment evaluation 要同时看 outcome、error、calibration、process、explanation 和 remediation。

Evaluation layer核心问题指标例子
Outcome parity不同 segment 是否获得类似机会或结果approval rate, prequalification rate, offer exposure, pricing distribution
Error parity错误是否集中伤害某些 segmentfalse negative disparity, false positive disparity, misclassification cost
Opportunity parity对合格客户是否有相近识别能力equal opportunity, qualified applicant pass-through
Calibration同一分数在不同 segment 是否代表相近风险calibration curve by segment, expected vs actual loss
Process parity流程摩擦是否不均衡completion rate, documentation burden, queue wait, handoff rate
Explanation parity解释是否一致、具体、可理解、可追溯reason-code fidelity, language readability, appeal instruction clarity
Human parity人工复核是否引入或纠正差异override rate, reviewer agreement, escalation rate, QA finding
Remediation parity申诉和纠错是否同等有效appeal success rate, correction time, customer reimbursement

Intersectionality 要进入抽样设计。只看单一维度会掩盖风险, 例如:

  • race proxy + gender + age band
  • language + channel + geography
  • self-employed + documentation type + product
  • branch customer + digital customer + call center customer
  • new-to-credit + thin-file + immigrant documentation pattern

6.5 Policy constraints and guardrails

在金融零售中, fairness 不能只靠模型训练。系统必须把政策约束写成可执行控制。

Control设计方式示例
Feature whitelist / blacklist决策引擎只允许 approved features禁用未经批准的 ZIP-derived microsegment 特征
Policy-as-code信贷、营销、催收、服务政策以规则、阈值和审批流实现hardship customer 不进入高频催收策略
Monotonic constraints对有明确业务含义的变量保持方向一致更低 DTI 不应导致更差信用结果, 除非有合规批准的交互项
Threshold governance阈值变更进入 release gate欺诈拦截阈值降低必须评估 segment false positive
Reason-code control原因码来自 approved sourceLLM 只能解释已确认 reason code, 不能生成新原因
Routing constraints不同语言、渠道、地区不能被路由到低质量 fallbackSpanish support route 必须有等价 SLA 和 QA
Experiment constraintsA/B 或 bandit 不得使特定群体承受未批准风险信贷 offer test 必须设置公平性和投诉 guardrail

6.6 Human review calibration

人工复核不是 fairness 的装饰, 是一个需要设计和监控的控制。

RiskControl
Rubber-stamping盲审样本、隐藏分数的二次复核、reviewer agreement 监控
Selective overrideoverride rate by segment / product / reviewer / branch
Reviewer biascalibration set、gold task、双人复核、adjudication board
Workload bias队列分配监控, 避免某些 segment 因复杂材料被长期积压
Explanation drift审核 adverse action / customer explanation 是否与真实原因一致
Skill mismatch高风险案例由 certified reviewer 处理, 普通客服不能确认信贷解释

Reviewer dashboard 应包含:

  • assigned case mix by segment
  • approve / decline / escalate / override rate
  • average handling time and queue age
  • reason-code edits
  • QA defects
  • customer complaint and appeal linkage
  • drift versus peer group and historical baseline

6.7 Monitoring and alerting

上线后监控要覆盖模型、业务、流程、人工和客户反馈。

Dashboard panel指标
Outcome disparityapproval / decline / counteroffer / pricing / offer exposure by segment
Error disparityfalse positive, false negative, bad-rate calibration, fraud confirmed rate by segment
Process disparitycompletion, abandonment, documentation request, queue wait, handoff, appeal time
Explanation qualityreason-code fidelity, unsupported explanation, language readability, adverse action QA defect
Human reviewoverride disparity, reviewer agreement, escalation disparity, rubber-stamp indicator
Model and data driftpopulation stability, feature drift, label drift, routing drift, vendor version changes
Customer harmcomplaints, regulatory complaints, appeal success, account freeze duration, remediation amount
Control healthmonitor freshness, missing segment data, failed jobs, unresolved alerts, open issues

Alert design:

  • Hard stop: production decisioning or explanation generation is paused for affected flow when evidence shows severe unsupported adverse action, unauthorized feature use, or broken monitor.
  • Release hold: new model / prompt / policy / vendor version cannot expand until segment evaluation passes or risk acceptance is approved.
  • Enhanced review: high-risk segment, channel or product enters increased manual review and QA sampling.
  • Remediation workflow: issue owner, root cause, affected population, customer correction, retraining or policy change are tracked to closure.

6.8 Evidence binder

Evidence binder 是 fairness control 的审计接口。它应该让内部审计、模型验证、合规或监管检查人员能够重建: 什么系统、影响谁、依据什么、如何测试、谁批准、上线后如何监控、问题如何修复。

Binder sectionEvidence
Use case scopeproduct description, customer impact assessment, decision authority boundary, AI inventory ID
Regulatory / policy mappingsource anchors, internal policy mapping, fair lending applicability memo
Data and featuredata lineage, feature inventory, proxy-risk register, missingness analysis, data quality report
Model / vendormodel card, vendor documentation, limitation memo, change notification terms, independent challenge
Evaluationsegment eval matrix, fairness metrics, calibration plots, explanation QA, adverse action test sample
Human reviewreviewer SOP, calibration results, override log, QA sample, training evidence
Release decisionarchitecture review, model risk validation, fair lending sign-off, risk acceptance, exception record
Monitoringdashboard snapshots, alert thresholds, issue log, remediation record, post-release review
Customer responseadverse action notice evidence, appeal workflow, complaint trend, correction and restitution record
Change managementmodel / prompt / feature / policy / routing / vendor change log and re-evaluation results

7. Metrics and Tradeoffs

7.1 Metric families

MetricDefinition intuitionFinancial retail useCaveat
Disparate impact ratioselection rate of protected group compared with reference groupapproval, prequalification, offer exposure, queue advancement常用于 screening signal, 不是自动合规结论或安全港
Equal opportunityqualified positive cases receive positive decision at similar rates合格申请人被批准或预审的机会依赖“qualified”标签质量, 历史标签可能有偏差
Calibrationsame score means similar observed risk across segmentscredit risk score, fraud score, churn / hardship score与 equalized odds 等指标可能冲突
False positive disparity被错误标记为高风险或负面结果的差异fraud false alerts, AML triage, collections hardship misclassification高客户伤害场景优先监控
False negative disparity应获得机会或保护但被漏掉的差异qualified borrower denied, hardship customer missed, complaint not escalated对 growth 和 customer protection 都重要
Service-level disparity服务速度、渠道、人工可达性、补件负担差异call routing, branch appointment, application completionUX 和 operations 是 fairness 的组成部分
Override disparity人工覆盖模型建议的比例和方向差异underwriter override, fraud investigator release, collections exception既可能显示人工纠偏, 也可能显示人工偏差
Explanation disparity解释的准确性、具体性、可读性和语言质量差异adverse action, chatbot credit explanation, appeal guidanceLLM 场景尤其重要
Remediation disparity申诉成功率、纠错时长、补偿金额差异credit reconsideration, account freeze release, complaint remediation衡量系统是否能纠错

7.2 Tradeoff map

Tradeoff实际含义Product / Architecture response
Accuracy vs fairness总体 AUC 提升可能来自对多数群体更好, 少数群体更差segment threshold, model class review, additional data, policy constraint
Fraud loss vs customer friction降低欺诈损失可能提高误拦和账户冻结risk-based step-up, release SLA, false positive parity, customer remedy
Automation vs accountability自动化越高, 解释、申诉和证据压力越大human approval boundary, audit log, explanation fidelity, kill switch
Personalization vs steering个性化 offer 可能变成不公平引导或排除offer governance, audience composition, consent and frequency cap
Data richness vs proxy risk更多数据可能提高预测力, 也可能编码敏感属性feature challenge, data minimization, less discriminatory alternatives
Calibration vs equal error rates不同公平指标可能无法同时最优由业务伤害、法律要求、风险偏好和治理决策选择主指标
Small sample vs statistical confidence小群体监控波动大rolling windows, confidence intervals, qualitative review, pooled segments with governance approval
Language localization vs consistency本地化提高理解, 但可能改变合规含义controlled translation, bilingual QA, source-linked explanation templates

7.3 Metric selection by use case

Use casePrimary fairness metricsSecondary metrics
Lending prequalificationdisparate impact, equal opportunity, false negative disparity, calibrationadverse action reason fidelity, completion disparity, override disparity
Credit explanation assistantexplanation fidelity, reason-code consistency, language readabilityunsupported claim rate, appeal instruction clarity, complaint rate
Fraud alert triagefalse positive disparity, account freeze duration, escalation disparityconfirmed fraud rate calibration, release SLA, customer complaint
Collections prioritizationcontact frequency disparity, hardship identification false negative, vulnerable customer safeguard ratecure rate by segment, complaint, right-party contact quality
Branch / customer-service routingSLA disparity, handoff disparity, complaint escalation paritylanguage support quality, abandonment rate, resolution rate

8. Financial Retail Cases

8.1 Lending prequalification

Scenario: AI 模型用于个人贷款预审排序, 决定哪些客户看到 prequalified offer 或进入人工 follow-up。

Design:

  • Decision boundary: AI 只做预审排序和 invitation recommendation, final credit decision 仍由 approved credit decision engine 和授权人员执行。
  • Protected attribute handling: production 不使用 protected attributes; fair lending testing zone 使用 approved protected-class data 或 proxy estimates 做 segment analysis。
  • Proxy controls: ZIP、branch、device、education、employment stability、language、marketing channel 进入 proxy-risk register。
  • Segment evaluation: approval proxy, prequalification exposure, application completion, downstream approval, pricing distribution, false negative for qualified customers。
  • Explanation: 预审说明不能暗示最终批准; adverse action reason 来自真实信贷决策, 不由 LLM 生成。
  • Monitoring: offer exposure disparity、application abandonment、counteroffer rate、manual override、complaints。
  • Evidence: feature review, less discriminatory alternative memo, segment eval matrix, release approval, monthly dashboard。

Failure mode:

  • 模型不直接拒贷, 但只把 offer 给某些渠道客户, 造成机会分配差异。
  • Marketing team 把 prequalification 当纯增长工具, 忽视 fair lending exposure。

Architect answer:

Prequalification is not low risk just because it is not final underwriting. It changes who receives a credit opportunity. I would govern it as a credit access control with exposure parity, proxy review, reason-code boundary, and downstream approval monitoring.

8.2 Credit explanation assistant

Scenario: LLM 帮客服或客户解释信用申请结果、原因码、补件要求和 reconsideration 路径。

Design:

  • Source control: LLM 只能使用 approved decision record、reason code、policy source、notice template 和 FAQ。
  • Reason fidelity: 每个解释句子必须追溯到真实 reason code 或政策条款。
  • No invented adverse action: 系统不得编造“收入不足”“信用历史短”等未出现在 decision record 中的原因。
  • Language parity: English、Spanish、Chinese 等语言解释需要同等具体性、可读性和申诉路径。
  • Human oversight: 客户发送前由授权人员确认高风险解释; chatbot 自助解释要有 escalation。
  • Monitoring: unsupported reason rate、translation defect、complaint、appeal confusion、agent edit rate。

Failure mode:

  • LLM 为了“更有帮助”补充听起来合理但不真实的拒绝原因。
  • 不同语言版本把 specific reasons 变成泛泛表达, 影响客户理解和申诉。

Architect answer:

For credit explanations, the model is not the source of truth. The source of truth is the approved decision record and reason-code engine. The LLM can translate and explain, but it cannot create reasons.

8.3 Fraud alert triage

Scenario: AI 对交易或账户行为排序, 决定哪些 alert 被拦截、升级、释放或要求 step-up authentication。

Design:

  • Harm model: false positive 可能造成账户冻结、交易失败、工资不可用、客户羞辱和投诉。
  • Segment eval: false positive disparity, confirmed fraud rate calibration, freeze duration, release SLA by language / channel / geography。
  • Step-up design: 优先使用可逆、低摩擦验证, 避免直接冻结成为默认动作。
  • Human calibration: investigator override、release decision、case aging 按 segment 监控。
  • Vendor challenge: vendor fraud score 必须在机构数据上 backtest, 不能只看 global performance。
  • Customer remediation: 错误冻结有明确释放、道歉、费用补偿和投诉处理流程。

Failure mode:

  • 某些移民客户或跨境交易客户因行为不同被高频误报。
  • 高风险队列积压导致某类客户冻结时间更长。

Architect answer:

Fraud fairness is not about approving fraud. It is about controlling unnecessary customer friction and false positives while still managing loss. The key metrics are false positive disparity, freeze duration, step-up success, and investigator override parity.

8.4 Collections prioritization

Scenario: AI 为逾期客户推荐 next-best-action: 短信、电话、延期、还款计划、困难客户支持或停止联系。

Design:

  • Policy constraints: 联系频率、时间窗口、禁止话术、vulnerable customer safeguards 和 hardship policy 写成规则。
  • Segment eval: contact frequency disparity, hardship false negative, repayment plan access, complaint rate, cure rate。
  • UX fairness: 数字渠道不可达客户不能因此被过度电话联系; 语言客户应获得等价 hardship 说明。
  • Human review: reviewer 必须记录为什么选择加强联系、降级联系或提供救助。
  • Monitoring: vulnerable customer detection QA、complaint coding、right-party contact、agent script adherence。

Failure mode:

  • 模型把“高回款概率”客户排到前面, 导致某些困难客户被高频联系, 而救助方案暴露不足。
  • 优化回款率挤压客户保护指标。

Architect answer:

Collections AI must optimize risk-adjusted recovery under customer protection constraints. The objective function cannot be pure dollars collected; it must include contact policy, hardship support, complaint risk, and segment-level treatment.

8.5 Branch / customer-service routing

Scenario: AI 根据客户问题、价值、情绪、语言和风险把客户路由到 chatbot、call center、branch specialist 或 complaint team。

Design:

  • Service-level parity: 等待时间、转人工率、一次解决率、投诉升级率按 segment 监控。
  • Accessibility: 老年客户、残障客户、低数字能力客户和非英语客户必须有可达路径。
  • Model routing: 不同语言或渠道不能被路由到低质量模型或更长等待队列。
  • Human review: 高风险投诉、威胁、欺诈、信贷、费用争议必须升级到授权队列。
  • Monitoring: abandonment、repeat contact、sentiment misclassification、resolution disparity。

Failure mode:

  • 高价值客户更快进入人工, 低余额客户长期困在 chatbot。
  • 非英语客户被默认转低质量翻译流程, 导致投诉和误解增加。

Architect answer:

Service routing creates fairness risk because it controls access to human help. I would treat service quality, wait time, escalation access, and resolution outcome as fairness metrics, not just operations metrics.


9. Artifact Templates

9.1 Fairness Requirement Canvas

FieldFill-in guidanceExample
Use caseAI 系统名称和业务流程Personal loan prequalification ranking
Decision influencedAI 影响但不一定最终决定的动作Offer exposure, lead priority, follow-up queue
Customer harm机会、价格、摩擦、解释、补救方面的潜在伤害Qualified customer may not receive prequalification invitation
Protected / sensitive segmentsLegal / compliance 批准的评估维度race / ethnicity proxy, sex, age band, language, geography, channel
Data sources输入数据、标签、第三方信号、文档credit bureau attributes, deposit history, campaign history
Proxy candidates可能产生 proxy risk 的特征ZIP, branch, device, language, employer category
Metrics主指标和辅助指标disparate impact, false negative disparity, calibration
Thresholds触发 review 的条件material adverse segment movement, agreed ratio threshold breach
Controls设计和运营控制feature whitelist, proxy-risk review, human QA, monitoring
Evidence上线前后证据eval report, approval memo, dashboard snapshots
Owners业务、模型、风险、合规责任人Product, Data Science, Fair Lending, Model Risk
Governance decisionlaunch / limited launch / redesign / rejectlimited launch with monthly fair lending review

9.2 Segment Eval Matrix

SegmentPopulation shareSelection rateError metricCalibrationProcess metricExplanation metricFindingAction
Reference group52%41%FNR 9%alignedavg queue 1.2 days98% supportedbaselinemonitor
Segment A18%31%FNR 15%underestimates risk? no, under-selects qualified applicantsavg queue 1.8 days96% supportedadverse opportunity gapproxy feature and threshold review
Segment B9%39%FPR 14%over-flags high riskavg queue 2.4 days93% supportedprocess and error disparityenhanced review and queue SLA control

Matrix 使用规则:

  • 不只填 selection rate, 必须同时填 error、calibration、process 和 explanation。
  • 对小样本 segment 使用置信区间、rolling window 和 qualitative review。
  • 每个 finding 必须有 action、owner 和复核日期。

9.3 Proxy-Risk Register

Feature / signalBusiness purposeProxy concernEvidenceAlternativeDecisionMonitoring
ZIP-derived distance to branchEstimate service accessMay correlate with race / income / rural accesshigh geography concentrationuse customer-selected channel preference and branch availabilityrestrict use to service routing, not eligibilityservice SLA by geography
Device typeFraud signalMay proxy income / age / digital accessmoderate protected-class predictabilityuse behavior consistency rather than device valuekeep only for fraud step-up, not credit pricingfalse positive by channel and device
Language preferenceService supportCould drive lower-quality routingoutcome disparity in legacy routingroute to equivalent language supportuse for service matching onlylanguage SLA and resolution parity

9.4 Fairness Control Map

Lifecycle stageControl objectiveControl activityEvidenceOwner
IntakeIdentify customer-impacting AIAI use case impact assessmentintake form, risk tierProduct
DesignPrevent unauthorized protected-class usefeature whitelist and data zone designdata architecture, access modelArchitect / Data Owner
BuildDetect proxy and label riskproxy analysis and label auditproxy register, label memoData Science
ValidateMeasure segment harmsegment eval and adverse action QAeval reportModel Risk / Fair Lending
ReleaseApprove residual riskgovernance sign-off and exception recordrelease gate memoBusiness / Risk
OperateDetect drift and disparityproduction monitoring and alertsdashboard, alert logOperations / Analytics
ChangePreserve control effectivenesschange impact analysis and re-evalchange log, re-test reportProduct / MLOps
AuditProve design and operationevidence binderbinder index, control samplesInternal Audit / Control Owner

9.5 Monitoring Dashboard Template

PanelChartDrilldownsAlert
Outcome parityapproval / exposure / decline trendsegment, product, channel, geography, model versionmaterial movement vs baseline
Error parityFPR / FNR / confirmed bad ratesegment, threshold, reviewer, vendor scoredisparity breach or calibration drift
Process parityqueue age, completion, handofflanguage, device, branch, call centerSLA disparity or abandonment spike
Explanation qualityunsupported reason, reason-code mismatchlanguage, template, agent, model versionspecific reason defect
Human reviewoverride rate, agreement, QA defectsreviewer, queue, segmentoutlier reviewer pattern
Customer harmcomplaints, appeals, remediationsegment, product, issue typecomplaint cluster or appeal success spike
Control healthjob freshness, missing data, unresolved alertsdata feed, monitor, ownermonitor failure or stale protected-data mapping

9.6 Model / Vendor Challenge Memo

SectionRequired content
Vendor and model scopeVendor name, model / score / API, version, use case, decision influenced
Training and data limitationsVendor-provided training description, known exclusions, geography, time period, limitations
Fairness claimsVendor fairness testing summary and exact claims
Institution-specific challengeBacktest on institution data, segment eval, proxy review, calibration, false positive / false negative disparity
Explainability and adverse actionWhether outputs can support specific reasons, limitations, reason-code mapping
Change controlVersion notice, retraining cadence, feature change, emergency rollback, contract rights
MonitoringRequired production metrics, raw output logging, dashboard, alert thresholds
Residual riskWhat remains unresolved, who accepts it, expiration date of approval
Decisionapprove, approve with constraints, limited pilot, reject, or redesign

10. Governance Operating Model

10.1 Roles and RACI

RoleResponsibility
Business OwnerOwns customer impact, business objective, residual risk acceptance, remediation commitment
AI PM / Product ArchitectConverts fairness risk into requirements, release gates, UX boundaries, monitoring and evidence
Data Science / ML EngineeringBuilds model, evaluates segments, documents features, supports proxy analysis and retraining
Data OwnerOwns data lineage, quality, access, retention and protected attribute handling
Fair Lending / ComplianceInterprets policy and regulatory expectations, reviews impact, challenges controls
Model Risk / ValidationIndependently challenges model design, validation, monitoring and limitations
LegalConfirms legal applicability, protected class handling, adverse action and customer communications
OperationsRuns human review, QA, escalation, customer correction and complaint handling
Vendor RiskChallenges third-party model, contract, change notice, audit rights and exit strategy
Internal AuditTests whether controls are designed and operating effectively

10.2 Release gates

GateMinimum evidence
Discovery gateuse case scope, customer impact, legal / policy questions, protected-data need
Design gatearchitecture boundary, decision authority, feature inventory, proxy-risk plan, human review design
Validation gatesegment eval, fairness metric pack, explanation QA, adverse action boundary test
Pilot gatelimited population, monitoring dashboard, complaint handling, rollback / pause criteria
Scale gatepilot results, issue closure, residual risk acceptance, monitoring cadence
Change gatemodel / prompt / feature / vendor / routing / policy change impact and re-evaluation

10.3 Issue management

Fairness issue 不应停留在 dashboard。它要进入 formal issue workflow:

FieldDescription
Issue哪个 segment、指标、产品、版本、时间窗口出现异常
Severity客户伤害、法律风险、规模、可逆性、证据强度
Root cause hypothesisdata drift, feature proxy, threshold, human review, vendor change, UX, policy
Immediate containmentpause, threshold rollback, enhanced review, customer correction
Investigation plandata analysis, sample review, reviewer interviews, vendor inquiry
Remediationfeature change, model retrain, policy change, UX fix, reviewer training
Customer remediationnotice correction, account release, fee reversal, reconsideration, complaint response
Governance closureowner attestation, validation review, dashboard confirmation, evidence binder update

11. Common Architecture Anti-Patterns

Anti-patternWhy it failsBetter design
Fairness through unawareness删除 protected attributes 不会删除 proxies 或历史偏差protected-data governance for testing, proxy detection, segment eval
Aggregate-only validation总体性能掩盖 subgroup harmsegment matrix, intersectional slices, harm-weighted metrics
Post-hoc dashboard上线后才发现需求没定义, 数据没保留fairness requirements and logging designed before build
Vendor black-box acceptance供应商声明不能替代机构责任vendor challenge memo, institution backtest, contract change controls
LLM explanation自由生成生成式模型会补充不存在的原因source-grounded explanation, reason-code fidelity, template constraints
Human review as decoration人工可能无能力、无时间、无权限挑战calibration, blind sample, override monitoring, escalation authority
One fairness metric指标之间会冲突, 业务伤害不同metric portfolio tied to decision context and harm model
Ignoring UX and language流程摩擦也是公平风险accessibility, language parity, completion and service disparity monitoring
No customer correction发现差异后只修模型, 不补救客户affected population analysis, customer remediation, appeal support

12. Interview Section

12.1 30-second answer

我不会把 AI fairness 当成单个模型指标, 而是把它设计成金融服务的产品和架构控制体系。第一步定义 AI 影响的决策: eligibility、pricing、fraud flag、collections、service routing 或 explanation。第二步设计 protected attribute handling、proxy-risk review、segment evaluation 和 reason-code fidelity。第三步把 human review、monitoring、alert、issue remediation 和 evidence binder 接入上线门禁。这样既能看 disparate impact、false positive / false negative disparity、calibration 和 service-level disparity, 也能证明控制被执行过。

12.2 2-minute answer

在金融服务里, AI fairness 的核心不是“模型有没有用敏感字段”, 而是 AI 系统是否在真实业务流程中造成不公平机会、价格、摩擦、服务或解释。我的设计会分五层。

第一层是 scope 和 decision authority。先确认 AI 影响的是信贷资格、定价、预审、欺诈拦截、催收优先级、客服路由、营销触达还是 adverse action explanation, 并明确 AI 是否只是 assistant、ranking tool、recommendation engine 或 final decision component。

第二层是 data and proxy control。protected attributes 通常不进入生产决策特征, 但在法律和合规批准下可以进入隔离的 fairness testing zone。对 ZIP、语言、渠道、设备、雇佣、教育、地理和 vendor score 做 proxy-risk analysis, 评估业务必要性和替代方案。

第三层是 evaluation。除了总体 AUC 或 loss, 我会按受保护或敏感 segment 做 selection rate、disparate impact、equal opportunity、false positive disparity、false negative disparity、calibration、service-level disparity、override disparity 和 explanation fidelity。对信贷场景, adverse action reason 必须来自真实决策记录和 approved reason-code mapping, LLM 不能编造原因。

第四层是 operational control。设计 human review calibration、blind sample、override logging、queue SLA、complaint / appeal linkage、monitoring dashboard、alert threshold、pause / rollback criteria 和 customer remediation。

第五层是 governance and evidence。每个高影响 use case 都要有 source anchors、policy mapping、data lineage、proxy register、segment eval matrix、model / vendor challenge memo、release approval、monitoring snapshots 和 issue log。这样公平性不是一次性评审, 而是贯穿 discovery、design、validation、release、operate 和 change management 的控制平面。


13. Portfolio Exercise

目标: 设计一个“个人贷款 AI 预审 + 信用解释助手”的 fairness / fair lending control package, 可作为 CBAP+ / AI PM / Product Architect / Risk Product Lead 面试作品集。

13.1 Case brief

一家金融零售机构计划上线 AI-powered personal loan prequalification。系统用客户交易、信用局数据、营销响应、数字行为和历史申请数据生成预审排序。通过预审的客户会看到 offer banner, 可继续申请。客服端还有一个 LLM assistant, 用于解释预审、补件、拒绝原因和 reconsideration 路径。

13.2 Deliverables

Artifact内容
Use case scope memoAI 做什么、不做什么、影响哪些客户、哪些决策受影响
Fairness requirement canvas按本手册 9.1 填写完整
Data and proxy inventory输入数据、标签、特征、第三方分数、proxy candidates
Segment eval matrix至少覆盖 protected-class proxy、language、channel、geography、thin-file、new-to-credit
Proxy-risk register至少列出 ZIP、language、device、employment stability、marketing channel
Explanation control designreason-code source, LLM constraints, language QA, adverse action boundary
Human review SOPreviewer authority, override reason, calibration sample, escalation
Monitoring dashboard specoutcome, error, process, explanation, human, complaint, control health
Vendor challenge memo如果使用第三方模型或 LLM, 写清测试、限制、变更和退出
Evidence binder map把每个控制连接到证据、owner、review cadence

13.3 Success criteria

  • 能说明为什么 prequalification 也会产生 fair lending exposure。
  • 能解释为什么 protected attributes 可以被隔离用于 fairness testing, 但不默认进入 production decisioning。
  • 能识别至少五类 protected-class proxy risk, 并给出业务必要性和替代方案分析。
  • 能定义 segment evaluation, 覆盖 outcome、error、calibration、process、explanation 和 human override。
  • 能说明 adverse action reason 为什么不能由 LLM 自由生成。
  • 能给出上线前 release gate 和上线后 monitoring / remediation 流程。
  • 能把每个控制映射到 evidence binder, 支撑审计或监管问询。

14. Practitioner Checklist

14.1 Discovery

  • 是否明确 AI 影响的 decision context: eligibility、pricing、prioritization、fraud / AML flags、service、collections、marketing、explanations。
  • 是否定义客户伤害: 机会缺失、价格更高、误拦、等待更久、补件负担、错误解释、申诉受阻。
  • 是否完成 legal / compliance / fair lending applicability review 的事实准备。
  • 是否登记 AI asset、model、vendor、data source、decision owner 和 human owner。

14.2 Data and feature

  • 是否有 data lineage、label definition、feature inventory 和 third-party signal inventory。
  • 是否区分 production decisioning features 与 fairness testing data。
  • 是否对 protected attribute / proxy estimate 做访问控制、用途限制、留痕和保留设计。
  • 是否完成 proxy-risk register, 包含业务必要性、替代方案和监控指标。
  • 是否分析 missingness、channel coverage、language coverage 和 historical label bias。

14.3 Evaluation

  • 是否按 segment 和 intersectional slice 评估。
  • 是否同时覆盖 selection / outcome、false positive、false negative、calibration、process SLA、explanation 和 override。
  • 是否记录 baseline、comparator、threshold、置信区间和小样本处理方法。
  • 是否对 adverse action reason、credit explanation 和 translation 做 QA。
  • 是否把 evaluation 结果转成 release decision, 而不是只存报告。

14.4 Architecture and operations

  • 是否有 feature whitelist、policy constraints、reason-code control、routing constraints 和 experiment guardrails。
  • 是否定义 human review authority、blind sample、override reason、reviewer calibration 和 QA sampling。
  • 是否能按 customer、case、model version、prompt version、policy version、reviewer、vendor version 重建记录。
  • 是否有 rollback、pause、enhanced review 和 customer remediation criteria。
  • 是否把 fairness monitors 纳入 production observability, 而不是离线分析。

14.5 Governance and evidence

  • 是否有 fair lending / model risk / legal / business sign-off 或正式 risk acceptance。
  • 是否有 model / vendor challenge memo, 覆盖 institution-specific backtest 和 change control。
  • 是否有 evidence binder, 覆盖 source anchors、policy mapping、data、model、eval、human review、release、monitoring、issue remediation。
  • 是否定义 monitoring cadence、alert owner、issue SLA 和治理报告节奏。
  • 是否把投诉、申诉、客户补救和监管问询纳入 feedback loop。

15. Final Operating Principle

金融零售 AI 公平性的成熟度, 不在于团队能说出多少 fairness metrics, 而在于能否把公平性变成可运行的控制平面:

Fairness requirement
-> data and proxy governance
-> segment evaluation
-> policy and human controls
-> release gate
-> monitoring and issue management
-> customer remediation
-> evidence binder
-> governance challenge

高级 CBAP+ / AI PM / Architect 的价值, 是把“模型是否公平”这个抽象问题拆成产品边界、流程责任、架构控制、评估证据和治理决策。这样团队才能在不牺牲创新速度的情况下, 对客户机会、公平待遇、监管问责和组织信任负责。