FICO Uses Blockchain to Block Unaudited AI Models From Production

The Intersection of Blockchain and AI With FICO — Ai4

Ai4

Embedding development specifications and agentic workflows into private blockchains creates an immutable ledger that mathematically prevents unverified AI models from deploying in regulated industries.

AI-assisted summary of The Intersection of Blockchain and AI With FICO — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this presentation features Scott Zoldi, Chief Analytics Officer at FICO. According to the video description, the session explores how FICO embeds governance rules directly into private blockchain technology to reduce hallucinations, enforce Responsible AI standards, and ensure complete auditability throughout the model development lifecycle.

## TL;DR Drawing on decades of experience in high-stakes financial analytics, Scott Zoldi outlines how private blockchain ledgers solve artificial intelligence's deep trust problem by acting as immutable, cryptographic development contracts. The framework tracks every stage of model creation—from initial scoping and ethics testing to runtime monitoring with self-aware 'Humble AI' triggers. Zoldi demonstrates how this methodology scales to generative AI through curated Focused Language Models (FLMs) and autonomous multi-agent workflows via Agentic Task Blockchains. By capturing a verifiable 'decision DNA,' organizations achieve mathematical proof of compliance and post-hoc interpretability for regulators and end users.

## Atmosphere & style The presentation carries an authoritative, pragmatic, and enterprise-hardened tone, swapping breathless AI hype for the grounded realities of regulatory compliance and engineering rigor. Zoldi speaks with the measured confidence of a veteran practitioner, using vivid analogies—such as comparing immutable ledgers to a continuous 'movie' and contrasting financial models with 'chicken soup' trivia—to demystify complex governance concepts. The talk feels like a clear-eyed executive briefing, balancing acute urgency regarding autonomous agent risks with practical architectural solutions.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=4s) High-stakes AI applications in regulated industries require moving beyond vague responsibility pledges toward codified corporate development standards that govern algorithms, explainability, and ethics. - [03:30](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=210s) Private blockchain technology acts as an immutable, tamper-proof contract that mathematically enforces adherence to Responsible AI standards throughout the engineering lifecycle. - [05:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=310s) Capturing the full development journey on-chain creates an unalterable forensic audit trail across developers, testers, and approvers, automatically blocking models that lack necessary validations. - [07:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=439s) Embedding runtime operational rules and feature boundaries onto the blockchain enables 'Humble AI' systems that recognize low-confidence edge cases and gracefully defer decisions. - [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) Regulated industries are shifting toward domain-specific 'focused language models' (FLMs) to eliminate hallucinations and bias inherent in opaque, general-purpose commercial foundation models. - [10:48](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=648s) Logging domain data curation—from sampling methodologies and scoring to encoder filtering—onto the ledger enables verifiable pre-training audits for demographic balance and bias. - [13:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=790s) Hardening models on the blockchain with synthetic tokens derived from vetted seed data enforces statutory compliance far more reliably than brittle prompt engineering. - [15:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=919s) The emergence of autonomous agentic AI introduces severe operational risks, requiring an Agent Blockchain to restrict tools, execution paths, and trusted data sources. - [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) Hierarchical Agentic Task Blockchains consolidate the sub-decisions of collaborating agents into an immutable 'decision DNA,' providing full post-hoc explainability and remediation paths. - [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) Overcoming public skepticism of autonomous AI requires capturing complete operational provenance and providing verifiable 'proof of measure' to regulators and end users.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=4s) The Trust Crisis in High-Stakes AI and the Need for Standards

Having deployed hundreds of AI and machine learning models across highly regulated industries like financial services, telecommunications, and insurance over a 27-year career at FICO, Scott Zoldi opens by addressing the fundamental barrier facing artificial intelligence today: trust. While AI is advancing rapidly, organizations and consumers struggle with the transition from seeing AI as an experimental science project to adopting it as a dependable tool. The stakes are particularly high when AI algorithms make critical life-altering decisions, such as determining whether an applicant is granted a home loan or whether a student is admitted into their college of choice. In these high-impact contexts—classified alongside policing as high-risk AI applications—an inability to triage, understand, and audit decisions leaves people vulnerable to models that hallucinate, exhibit bias, or deliver flat-out erroneous outcomes.

To establish genuine trust, Zoldi argues that organizations must move beyond vague, aspirational "oaths" of responsibility and instead codify rigorous corporate AI development standards. In an organization staffed with hundreds of data scientists, leaving development unconstrained results in hundreds of disparate, idiosyncratic approaches, most of which will fail enterprise-level governance standards. A centralized standard establishes a uniform playbook: which specific machine learning algorithms are permitted, how explainability and interpretability are achieved, and how rigorous ethics and disparity testing are conducted. By focusing the organization's best minds on establishing and updating this standard as regulations evolve, companies prevent their data science teams from simply trying to "impress each other with different algorithms" and align them around building compliant, production-grade systems.

- **High-Risk Classifications:** Financial credit risk and fraud detection are categorized in the same high-risk tier as predictive policing due to their profound financial and societal consequences. - **Core Governance Pillars:** Model governance mandates strict requirements around ethics testing, model explainability, interpretability, and post-deployment auditability.

---

[03:30](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=210s) Enforcing Corporate Responsible AI Standards via Blockchain

Establishing an AI development standard is only half the battle; the harder problem is enforcement. Zoldi points to private blockchain technology as the ideal mechanism to operationalize and enforce compliance. Moving beyond the popular association of blockchain with cryptocurrencies, Zoldi explains that blockchain fundamentally acts as an unbreakable contract. In earlier days, compliance required data scientists to physically sign paper contracts pledging adherence to build criteria. Blockchain modernizes this by immutably recording every stage of development. Because each block cryptographically builds on all previous blocks, attempting to retroactively alter requirements or forge test results breaks the entire chain.

```mermaid flowchart LR accTitle: Blockchain-Enforced AI Development Lifecycle subgraph Roles["Stakeholder Roles"] DEV["Developer"] TST["Tester"] APP["Approver"] end

subgraph Chain["Private Distributed Blockchain Record"] B1["Model Requirements & Focus"] B2["Algorithm & Variable Selection"] B3["Disparity & Ethics Testing"] B4["Final Approvals & Sign-off"] B1 --> B2 --> B3 --> B4 end

subgraph Gate["Release Gate"] AUDIT{"All Criteria Met on Chain?"} PROD["Deploy to Production"] STOP["Block Release"] end

DEV --> B1 DEV --> B2 TST --> B3 APP --> B4 B4 --> AUDIT AUDIT -- "Yes (Immutable Proof)" --> PROD AUDIT -- "No / Incomplete" --> STOP ```

This immutability introduces total accountability into the engineering process. If development targets or fairness thresholds turn out to be difficult to satisfy mid-project, teams cannot quietly lower the bar or edit the criteria; they simply fail to meet the standard and cannot release the model. This guarantees that model governance is not treated as an afterthought or a rushed paperwork exercise before deployment, but as an embedded, tamper-proof tracking system.

- **First Patent Filing:** FICO filed its foundational patents applying blockchain architecture to AI model governance in 2018. - **Contract Enforcement:** Replacing manual paper sign-offs with cryptographic verification ensures mathematical certainty that development standards were followed.

---

[05:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=310s) Auditing the Model Lifecycle: The Blockchain as a "Movie"

Zoldi describes the governance blockchain as a continuous "movie" rather than a single snapshot in time. A lending model project begins by logging its explicit scope and business objectives onto the ledger. Specific ethical thresholds—such as quantitative measurements of demographic disparity—are locked in. Throughout the build, three distinct organizational personas interact with the chain: 1. **The Developer:** Documents algorithm selection, data lineage, and architecture decisions. 2. **The Tester:** Runs independent validation, bias evaluations, and stress testing against defined metrics. 3. **The Approver:** Evaluates the resulting evidence and formally signs off at prescribed gates.

Crucially, the ledger captures the entire developmental journey—highlighting where models struggled, how parameters were adjusted, why pivots occurred, and who authorized each change. By preloading governance and decision provenance directly into the ledger while the model is built, traditional retrospective governance hurdles disappear. When models enter production, the organization retains complete forensic clarity. If an audit is conducted or a model's behavior is challenged, compliance is proven through an unalterable audit trail. If any required validation check or approval signature is missing from the chain, the model deployment is automatically blocked.

---

[07:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=439s) Model Monitoring, Production Governance, and "Humble AI"

Building a responsible model is insufficient if its operational behavior degrades in production. Model monitoring remains one of the greatest operational challenges across enterprise AI, often plagued by naive reliance on generic data-drift metrics that provide little actionable insight. Zoldi emphasizes that the original model designer is uniquely qualified to define exact production monitoring specifications.

For complex models like neural networks, the designer specifies which latent features are expected to fire, their standard activation distributions, and the permissible tuples of co-occurring features. These operational rules are written directly into the blockchain alongside the model. The blockchain travels with the asset into runtime, establishing clear triggers for when a model must be recalibrated, temporarily overridden, or completely decommissioned. This mechanism forms the foundation of "Humble AI"—systems architected with built-in self-awareness to identify unfamiliar or edge-case transactions where their confidence is low, gracefully deferring decisions rather than guessing.

---

[08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) Applying Blockchain Governance to Generative AI and Focused Language Models

Turning to generative AI, Zoldi notes that while foundational concepts have been deployed across financial services for decades, recent advances in compute and massive commercial foundation models have sparked an arms race where tech companies constantly benchmark against one another. However, these massive off-the-shelf foundation models present significant compliance and safety risks for regulated scoring. Regulated entities are legally required to understand the underlying training data to guarantee that outputs are not discriminatory, ungrounded, or prone to hallucinations. Because commercial foundation models obscure their data pipelines, regulated industries are increasingly moving away from third-party general models in favor of purpose-built, domain-specific systems.

To solve this, FICO implements what Zoldi terms "focused language models" (FLMs). When constructing an enterprise language model, the initial step recorded to the blockchain is defining its explicit operational scope. If an FLM is built for financial services workflows, it must be restricted to relevant financial domain knowledge. As Zoldi illustrates, a credit evaluation assistant has no business knowing "how to make chicken soup," reciting the "history of Poland," or generating trivia about the "Spice Girls." Purging non-domain data prevents irrelevant associations that introduce bias and hallucinations, establishing an intentional, auditable scope from block zero.

[10:48](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=648s) Curating Focused Domain Data on the Blockchain

Determining what an AI model is intended to do is fundamentally a business decision rather than a raw data decision. Once that business objective is established, the development team extracts a sample from a massive public corpus such as Common Crawl and evaluates it with a research decoder model. Every critical artifact from this preliminary phase is committed directly to the private blockchain: the defined business focus, the sampling methodology, the exact decoder model selected, and the resulting scored sample dataset. Scoring filters signal from noise; high scores represent text closely aligned with the designated domain focus, while low scores isolate irrelevancies like "the Spice Girls or chicken soup."

```mermaid flowchart TD accTitle: Blockchain-Audited Focus Domain Dataset Pipeline subgraph Ingestion["1. Ingestion & Scoring"] A["Raw Corpus<br/>(Common Crawl)"] --> B["Sample Extraction"] B --> C["Decoder Scoring"] end subgraph Refinement["2. Distillation & Filtering"] C --> D["Train Encoder Model"] D --> E["Score Full Corpus"] E --> F["Focus Language<br/>Domain Dataset"] end subgraph Governance["3. Blockchain Ledger"] G["Immutable Blockchain Record<br/>• Business Focus & Specs<br/>• Model Checkpoints<br/>• Demographic Audits"] end B -.-> G C -.-> G D -.-> G F -.-> G ```

From these scored samples, developers construct a lightweight encoder model that codifies the decoder's filtering logic. This encoder model—also logged to the blockchain—is then unleashed across the full dataset to extract only the pertinent text, producing a tailored "focus language domain dataset." Because every step is immutably tracked, model builders retain end-to-end auditability and demographic control. Teams can empirically verify whether geographic regions like Wyoming have adequate representation or run tests to detect racial disparities before any foundational architecture is trained.

- The business focus, sampling strategy, decoder model, and scored calibration data are immutably logged to the ledger. - Distilling decoder evaluations into an encoder model allows high-throughput extraction of domain-specific data from massive uncurated corpora. - Full provenance allows pre-training audits for geographic balance, racial bias, and compliance with safety mandates in highly regulated industries.

---

[13:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=790s) Domain Model Training and Regulatory Hardening

Once the domain dataset is fully assembled, engineers can begin training the base domain model in PyTorch. At this stage, the base model merely mirrors the linguistic patterns of the curated text. To make it functionally useful, it must undergo strict task-specific training tailored to a business function, such as banking customer compliance. When addressing heavily regulated environments—such as Australia, where the Australian Securities and Investments Commission (ASIC) enforces nearly nine separate regulations governing customer dispute calls—those statutory requirements are encoded directly as domain tokens.

```mermaid flowchart LR accTitle: Task-Based Model Hardening vs Prompting subgraph Prompting["Conventional Prompting (Brittle)"] P1["Generic LLM"] --> P2["System Prompt:<br/>'Act as ASIC Agent'"] P2 --> P3["Unpredictable Output"] end subgraph Hardening["FICO Hardened Architecture"] H1["Seed Data<br/>(Golden Examples)"] --> H2["Billions of Synthetic<br/>Task Tokens"] H2 --> H3["Task-Trained Model<br/>(Hardened on Ledger)"] H3 --> H4["AI Trust Score Monitoring"] end ```

Rather than relying on generic prompt engineering, the business provides curated "seed data" containing prime examples of stellar compliance and historical failure modes. From this seed set, billions of synthetic tokens are generated to train a hardened, task-specific model. This deliberate architecture avoids the fragility of asking a generic chatbot to "pretend to be a customer service representative who's aware of ASIC." The entire lineage—from seed examples to synthetic token generation—is permanently preserved on the blockchain, and secondary monitoring models produce runtime "AI trust scores" to flag anomalous outputs during live deployment.

- Domain models are hardened with billions of synthetic tokens generated from vetted business seed data rather than generic prompt personas. - ASIC regulatory constraints are codified directly into the training corpus for customer-facing banking interactions. - Secondary runtime models evaluate outputs to assign objective AI trust scores before end users receive responses.

---

[15:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=919s) Post-Hoc Accountability and the Shift to Agentic AI

Embedding development artifacts into an immutable ledger fundamentally resolves post-deployment ambiguity. When regulators, internal model governance teams, or aggrieved customers challenge an outcome months later, organizations no longer need to speculate about developer intentions. The blockchain provides definitive, auditable proof of whether specific regional data like Kazakhstan was included, or whether disability bias testing was conducted prior to release. This rigorous accountability is mandatory whenever automated algorithms make decisions that alter human livelihoods.

```mermaid flowchart TD accTitle: Constraints on Autonomous Agent Operations subgraph Execution["Agent Execution Environment"] A["Agent Invocation"] --> B{"Blockchain Permissions Check"} B -->|"Authorized Path"| C["Verified Data Source<br/>(e.g., User's Bank)"] B -->|"Blocked Path"| D["Untrusted Source<br/>(e.g., Unverified Web)"] C --> E["State & Tool Recording"] E --> F["Audited Action"] end ```

However, the rapid industry transition toward autonomous agentic AI introduces severe operational hazards. As models gain agency—executing tools, browsing forums, interacting with external APIs, and adapting their behavior dynamically based on task feedback—the surface area for failure expands exponentially. To prevent agents from behaving like the "wild west," FICO introduced the concept of an Agent Blockchain. This framework strictly binds autonomous agents to pre-approved execution paths, verified tool sets, and curated data sources, while recording the exact environmental state and references used during runtime execution.

- Immutable logs eliminate guesswork during regulatory audits regarding variable choice, geographic coverage, or protected-class fairness testing. - Agentic systems introduce unique compliance risks due to dynamic tool usage, runtime environment volatility, and continuous self-adaptation. - The FICO Agent Blockchain enforces operational guardrails, ensuring agents only query trusted data sources rather than unvetted internet endpoints.

---

[18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) Multi-Agent Systems, Public Trust, and "Decision DNA"

Managing autonomous AI systems becomes even more intricate when multiple specialized agents collaborate to execute composite tasks. Drawing on FICO's experience building synthetic fraud detection models for global payment card transactions since 1992, multi-agent frameworks require rigorous oversight of every sub-decision. In a routine corporate travel booking scenario, disparate sub-agents might negotiate flight bookings, hotel reservations, and rideshare options; if a user is billed $24 for a rideshare alongside an $800 airline ticket, they require transparent, step-by-step triage explaining why those specific choices were selected.

```mermaid flowchart TD accTitle: Hierarchy of Multi-Agent Task Blockchain to Decision DNA subgraph Agents["Autonomous Sub-Agents"] A1["Agent 1 Ledger<br/>(Travel Search)"] A2["Agent 2 Ledger<br/>(Expense Policy)"] A3["Agent 3 Ledger<br/>(Booking & Payment)"] end subgraph Aggregation["Composite Task Ledger"] A1 --> T["Agentic Task Blockchain"] A2 --> T A3 --> T T --> D["'Decision DNA'<br/>(Fully Interpretable Human Audit Trail)"] end ```

Widespread public hesitation underscores the necessity of this auditability. Industry survey data from Deloitte indicates that only 3% of individuals are eager "zealots" willing to surrender complete autonomy to AI. Around 20% reject the technology outright, while the vast majority insist on human-in-the-loop validation for every material choice, bottlenecking operational speed.

To bridge this trust deficit, FICO is establishing hierarchical "Agentic Task Blockchains." In a workflow orchestrating 20 individual agents, each agent logs its state, tools, and execution path to an aggregated ledger that functions as the composite "decision DNA." Just as credit scores today provide consumers with adverse action reason codes and clear remediation paths, future autonomous decisions must be fully unwindable, interpretable, and cryptographically verifiable.

- According to Deloitte research, only 3% of users fully trust autonomous AI without oversight, 20% reject it entirely, and the remainder require continuous human-in-the-loop verification. - Hierarchical Agentic Task Blockchains consolidate the audit trails of dozens of collaborating agents into a single verifiable ledger. - This immutable "decision DNA" enables post-hoc explainability, allowing users and auditors to trace the exact chain of logic behind complex automated decisions.

[21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) Establishing Provenance and Accountability in Agentic AI with Blockchain

Scott Zoldi emphasizes that tracking AI systems requires capturing a comprehensive record of provenance—not merely tracking where the underlying data originated, but documenting the complete operational lineage of how autonomous agents interact with one another and how effectively they execute their assigned tasks. Acknowledging that the industry remains "far away" from resolving its persistent trust issues across standard AI and generative AI, he stresses that building robust governance frameworks around agentic AI is one of the single most urgent challenges organizations must address.

By uniting blockchain with AI architectures, organizations can establish deep-seated trust and verifiable accountability. This dual layer of accountability serves internal standards while offering external assurance to customers and regulators that models behave strictly according to specification. Embedding these audit trails provides the tangible "proof of measure" required to meet strict regulatory guidelines, verify proper model construction, and enforce responsible governance from end to end.

```mermaid flowchart LR accTitle: Blockchain Governance for Multi-Agent AI Systems subgraph Agentic_Layer["Agentic AI Execution"] A["Agent Task Execution"] --> B["Multi-Agent Interactions"] end subgraph Blockchain_Layer["Immutable Provenance Ledger"] B --> C["Interaction & Data Lineage"] C --> D["Proof of Measure & Compliance"] end subgraph Trust_Outcomes["Accountability & Trust"] D --> E["Internal Standards"] D --> F["Customer & Regulatory Trust"] end ```

- **Operational Agent Provenance:** Extends beyond source data history to capture the full step-by-step history of inter-agent handoffs, behavior, and successful task completion. - **Dual-Layer Accountability:** Guarantees that internal teams strictly adhere to responsible AI guidelines while providing external stakeholders and regulators mathematical verification that models function as declared.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=4s) impress each other with different algorithms — Scott Zoldi

> [05:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=310s) movie — Scott Zoldi

> [07:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=439s) Humble AI — Scott Zoldi

> [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) focused language models — Scott Zoldi

> [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) how to make chicken soup — Scott Zoldi

> [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) history of Poland — Scott Zoldi

> [10:48](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=648s) the Spice Girls or chicken soup — Scott Zoldi

> [13:10](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=790s) pretend to be a customer service representative who's aware of ASIC — Scott Zoldi

> [15:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=919s) wild west — Scott Zoldi

> [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) zealots — Scott Zoldi

> [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) Decision DNA — Scott Zoldi

> [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) proof of measure — Scott Zoldi

## Mentioned Scott Zoldi, FICO, Common Crawl, PyTorch, Australian Securities and Investments Commission (ASIC), Deloitte, Spice Girls, Poland, Wyoming, Kazakhstan

## Video description Scott Zoldi - Chief Analytics Officer, FICO

AI has a trust problem. Research shows that most Americans are more concerned than excited about AI in daily life, and that concern is well-founded: without proper governance, AI models can introduce bias, produce unexplainable outcomes, and be nearly impossible to audit after the fact. To build genuine trust in AI, organizations must ensure that their models are interpretable, auditable, and enforceable, and blockchain technology offers a practical, operational path to achieving all three. In this session, I will discuss how FICO uses blockchain-based AI models to help reduce hallucinations, while ensuring adherence to our Responsible AI standards. By embedding governance rules directly into a private blockchain, we have created an immutable record of every decision made throughout the model development lifecycle, from variable selection and algorithm approval to ethics testing and final sign-off.

[Source](https://www.youtube.com/watch?v=BqaH9gAhJ3g)

中文简报

FICO利用区块链阻断未经审计的AI模型投产 将开发规范与智能体工作流嵌入私有区块链，可创建不可篡改的账本，从而在数学层面防止未经核验的AI模型在受监管行业部署。 依托在高风险金融分析领域数十年的经验，Scott Zoldi 阐述了私有区块链账本如何作为不可篡改的密码学开发合约，化解人工智能深层次的信任难题。该框架追踪模型创建的每个阶段——从最初的范围界定与伦理测试，到带有自感知“谦逊AI”（Humble AI）触发器的运行时监控。Zoldi 展示了如何通过精选的聚焦语言模型（FLMs）以及基于智能体任务区块链（Agentic Task Blockchains）的自主多智能体工作流，将该方法推广至生成式 AI。通过捕获可验证的“决策DNA”，机构能够为监管机构和最终用户提供合规性的数学证明以及事后可解释性。 FICO 首席分析官 Scott Zoldi 指出，高风险 AI 亟需建立严格的治理标准；通过将“负责任 AI”（Responsible AI）规范直接写入私有区块链，可以实现以数学方式强制履约，并自动阻断未达标模型的上线发布。 (3:30) 为避免通用商业大模型的幻觉与不可解释性，受监管行业正转向“聚焦语言模型”（FLM），将数据采样、模型评分与编码器过滤全流程记入链上，从源头确保训练数据的审计合规与偏见控制。 (8:54) 相较于脆弱的提示词工程，FICO 采用经过验证的种子数据生成数十亿合成 Token，直接将澳大利亚 ASIC 等监管法规内化硬化进任务模型中，并结合运行时 AI 信任评分进行异常监控。 (13:10) 针对自主智能体（Agentic AI）带来的失控风险，FICO 提出“智能体区块链”与“决策 DNA”架构，严格约束智能体的工具调用与数据源，为复杂的多智能体协作提供完整可追溯的审计记录。 (18:08)

Reading summary

FICO Chief Analytics Officer Scott Zoldi argues that bridging the AI trust deficit in regulated, high-stakes decisions requires replacing voluntary corporate pledges with private blockchain ledgers that immutably enforce development standards, model lineage, and autonomous agent behaviors. Enforcing AI Governance via Immutable Private Blockchains (3:30) In high-stakes environments such as credit underwriting and fraud detection—which Zoldi notes share the same high-risk tier as predictive policing—unconstrained data science teams often rely on idiosyncratic methodologies that fail compliance testing. Instead of relying on manual paper contracts or aspirational ethical oaths, FICO filed patents in 2018 to operationalize model governance using private blockchains. The ledger acts as an unalterable contract tracking development milestones, including algorithm choices, disparity tests, and formal sign-offs across distinct developer, tester, and approver personas. Because each cryptographic block builds on earlier entries, teams cannot retroactively loosen fairness thresholds or alter model parameters if they encounter development difficulties mid-project. If required validation criteria or approver signatures are missing from the ledger, release gates automatically block the model from production deployment. Zoldi likens this ledger to a continuous "movie" of the model lifecycle that preserves forensic clarity for subsequent internal or regulatory audits. Runtime Operational Controls and 'Humble AI' (7:19) Model governance must extend beyond initial deployment into live operations, where Zoldi observes that generic data-drift metrics frequently fail to deliver actionable insight. To prevent performance degradation, the original model designer logs explicit runtime operational criteria directly onto the blockchain, including expected neural network latent feature firing rates, baseline activation distributions, and permissible feature co-occurrence tuples. This continuous audit trail travels alongside the deployed asset to establish deterministic triggers for model overrides, recalibration, or decommissioning. It also underpins what Zoldi terms "Humble AI"—architectures designed with programmatic self-awareness to identify out-of-distribution or edge-case transactions where model confidence is low, prompting the system to gracefully defer automated decisions rather than generate ungrounded estimates. Domain-Specific Data Curation for Generative Models (8:54) Broad commercial foundation models introduce severe legal and compliance hazards in regulated scoring because their training pipelines are opaque, increasing the likelihood of ungrounded hallucinations and protected-class bias. In response, FICO advocates for domain-restricted "focused language models" (FLMs). Scope definition begins on the blockchain by codifying business objectives and stripping out irrelevant subject matter, such as culinary recipes or pop culture trivia, which Zoldi identifies as sources of extraneous bias. To construct compliant training sets, development teams evaluate public web samples from sources like Common Crawl using a research decoder model, committing the evaluation scores and sampling logic to the blockchain. An encoder model then distills this logic to extract only high-scoring domain text from the broader corpus. This auditable lineage allows organizations to verify geographic representation, such as data coverage for Wyoming, and conduct demographic disparity testing before model training begins in PyTorch. For customer-facing banking tasks, such as compliance with nine distinct Australian Securities and Investments Commission (ASIC) customer dispute rules, models are hardened using billions of synthetic tokens derived from curated seed data rather than generic system prompts. Governing Agentic Workflows and Building 'Decision DNA' (15:19) As artificial intelligence transitions toward autonomous agentic architectures that dynamically execute external tools and browse APIs, operational failure surfaces expand. Zoldi cites Deloitte industry survey data indicating that only 3% of users willingly grant complete autonomy to AI systems, while 20% reject the technology outright and the remaining majority demand continuous human-in-the-loop validation. To address this hesitancy without causing operational bottlenecks, FICO implements Agent Blockchains that restrict autonomous agents to pre-approved data repositories, authorized tools, and deterministic execution paths. In composite, multi-agent environments—such as travel orchestration systems coordinating independent flight, hotel, and rideshare agents—each sub-agent records its states and actions to a hierarchical "Agentic Task Blockchain." This consolidated ledger forms an auditable "decision DNA" analogous to credit score adverse action reason codes, ensuring complex composite decisions can be audited, untangled, and verified. Zoldi emphasizes that the enterprise sector remains far from resolving persistent trust issues in AI, but insists that combining blockchain verification with agentic frameworks provides the necessary cryptographic provenance and dual-layer accountability to satisfy both internal governance and regulatory scrutiny.

精炼阅读摘要

FICO分析官Scott Zoldi指出，高风险AI决策的信任缺失无法靠宽泛的道德倡议化解，必须通过私有区块链将开发规范、数据血统与运行时监控固化为不可篡改的数学合约，为模型全生命周期建立确定性的审计追踪。 高风险场景与区块链强力治理 (3:30) Zoldi结合在受监管行业部署数百个AI模型的实践指出，信用风险评估与欺诈检测对公众生活具有决定性影响，其潜在危害等同于预测性警务。当企业拥有庞大的数据科学团队时，缺乏约束的研发往往演变为个人算法偏好的展示，难以符合统一的企业级治理标准。为此，FICO于2018年申请了利用区块链架构进行AI模型治理的专利，用密码学验证彻底取代容易事后补签的纸质合规流程。 在这一架构下，项目目标、算法选型、数据血统与伦理偏倚阈值从研发初始即被固化在私有链上，形成连续记录的生命周期“电影”。区块链设立了开发者、测试者与审批者三类角色，测试团队对照既定指标进行独立的偏倚与压力测试。唯有所有验证环节与审批签名在链上得到完整确认，发布门禁才会放行。研发团队若在项目中途遭遇合规瓶颈，无法私自下调公平性门槛，从而杜绝了上线前的应付式合规或指标篡改。 运行期监控规范与谦逊AI架构 (7:19) 模型治理不仅局限于离线训练，更取决于上线后的实际行为。Zoldi强调，行业目前依赖通用数据漂移指标的监控方法难以提供切实可行的操作指引，而原模型设计者才是定义生产监控规范的最佳人选。针对神经网络等复杂架构，设计者需明确潜在特征的预期触发阈值、标准激活分布以及允许共现的特征元组，并将这些规则作为不可分割的资产随模型一同写入区块链。 这些上链规则在生产环境中建立了清晰的运维触发机制，明确了模型何时必须重新校准、临时人工接管或彻底停用。这构成了“谦逊AI”（Humble AI）的核心底座：通过赋予模型内在的自我感知能力，当系统遇到置信度过低或特征组合异常的边缘交易时，能够主动识别自身认知的局限并优雅地放弃决策，将控制权转交人工，而非在不确定状态下盲目猜测。 领域聚焦语言模型与法规级硬化 (8:54) 面对生成式AI浪潮，Zoldi认为商业通用基础模型对其底层训练数据的遮蔽，使受监管机构面临重大的歧视与幻觉合规风险。为此，FICO主张构建“聚焦语言模型”（FLM）。从建链初始，系统即明确业务范围并剔除一切非领域关联知识。研发团队从Common Crawl等海量公开语料中提取样本，经研究级解码器评分后训练轻量编码器，进而从全量数据中高速蒸馏出受控的领域数据集。该过程的业务目标、采样方法和模型节点均全程上链，使团队能在架构训练前审计地理分布与潜在偏倚。 在模型训练环节，面对如澳大利亚证券与投资委员会（ASIC）针对客户争议电话施加的近九项严格法规，FICO摒弃了脆弱的通用提示词工程，直接将法定要求编码为领域Token。企业利用沉淀的高质量种子数据生成数十亿合成任务Token，训练出深度硬化的业务模型。整个训练血统与合成数据源永久保存在区块链上，并由次级运行时模型输出客观的“AI信任评分”，在内容触达终端用户前拦截异常响应。 智能体运行约束与决策DNA (15:19) 随着自主智能体（Agentic AI）开始具备动态调用工具、访问外部API与动态适应环境的能力，AI系统的故障面呈指数级扩大。Deloitte的行业调研表明，公众对全自主AI存在严重的信任赤字：仅3%的用户愿意完全交付自主权，20%的人断然拒绝，绝大多数用户仍要求在关键决策环节保留人工干预。这种谨慎态度直接制约了自动化系统的大规模部署与运转效率。 为化解信任危机并遏制无序调用，FICO提出了“智能体区块链”（Agent Blockchain）框架，利用链上权限机制将自主智能体严格锁定在预先核准的执行路径、工具集与可信数据端点内。在差旅预订等多智能体协同场景中，系统建立分层的智能体任务区块链，汇总各个子智能体的状态转移与决策依据，凝练为人类完全可审计的“决策DNA”（Decision DNA）。这使得复合决策如同信用评分的不良行为代码一样，具备事后完整回溯、逻辑解构与密码学验证的能力。 Zoldi指出AI领域的信任危机远未彻底解决，特别是在多智能体协作与复杂模型纵深发展的背景下。通过将区块链作为AI不可篡改的“度量证明”，企业得以在满足内部治理的同时为监管机构提供透明的数学确定性，为高风险自动化决策确立了从数据源头到运行时交互的端到端问责基准。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布。视频由 FICO 首席分析官 Scott Zoldi 主讲，根据页面简介，内容聚焦于如何利用私有区块链技术将治理规则内嵌于模型全生命周期中，以解决 AI 的信任危机、降低幻觉并确保严格符合负责任 AI 标准。 ## TL;DR FICO 首席分析官 Scott Zoldi 针对金融等强监管高风险行业中的 AI 信任赤字，提出了基于区块链的负责任 AI 治理方案。演讲阐明了如何利用区块链的不可篡改性与智能契约，强制执行企业级开发规范，全流程存证数据采样、特征选择与审批记录并作为生产发布的硬性门禁。针对生成式 AI 与大模型，他主张构建剔除冗余杂音的“聚焦语言模型（FLM）”并通过种子数据深度硬化合规逻辑，同时引入“谦逊 AI”机制让模型在不确定场景下主动放弃决策。面向多智能体（Agentic AI）时代，FICO 规划了层级化的“智能体任务区块链”架构，通过白名单与环境快照锁定执行边界，为重大自动化决策沉淀终身可溯源的“决策 DNA”。 ## Atmosphere & style 演讲整体洋溢着深厚严谨、务实笃定的工业级工程与合规底色，展现了一位在风险建模领域深耕数十年的资深专家的技术洞察。Scott Zoldi 摒弃了当下大模型领域的盲目狂热叙事，以生动犀利的行业类比直击黑盒模型的治理痛点，论证层层推进且逻辑严丝合缝。整场分享不仅传递出对算法伦理与金融底线的敬畏之情，更以冷静沉着的专业气场，为复杂的自主智能体系统勾勒出一条高度可落地的技术信任路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=4s) 信贷审批与欺诈检测等高风险 AI 应用直接影响个人的重大生活和财务成果，必须建立包含伦理测试、可解释性与全流程审计的强力治理机制。 - [01:40](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=100s) 企业必须确立统一的负责任 AI 开发标准，杜绝数据科学家自由炫技和各行其是，确保所有模型在算法准入与解释方法上严格对齐合规要求。 - [03:36](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=216s) 区块链的密码学不可篡改性为模型合规提供了强制约束，杜绝了开发团队在后期因指标未达标而暗中调低初始标准的侥幸行为。 - [04:53](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=293s) 私有区块链完整记录了开发者、测试者与审批者的全生命周期协同过程，作为生产发布的硬性门禁，未完全达标的模型绝对禁止上线。 - [07:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=439s) “谦逊 AI（Humble AI）”理念强调深层潜在特征激活监控，预设规则让模型在超出能力边界或处于高风险场景时主动放弃决策并交由人工处理。 - [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) 通用大模型因语料黑盒无法满足监管审查，受监管行业需自建“聚焦语言模型（FLM）”，从源头剔除无关垂直领域的语料以抑制偏见与幻觉。 - [10:48](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=648s) 构建聚焦数据集的全流程（包括业务焦点设定、解码器打分与编码器筛选）均写入区块链，使企业在模型训练前即可主动完成地域与人口偏见审计。 - [13:30](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=810s) 模型合规严禁依赖 Prompt“角色扮演”，必须基于业务种子数据合成数十亿级词元进行行为硬化，并将调优历程链上存证以备终身复核。 - [16:06](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=966s) 面对自主智能体动态环境不可复现的失控风险，FICO Agent 区块链通过链上合约强制锁定工具白名单、获准执行路径与瞬时环境快照。 - [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) 针对多智能体协同决策，层级化“任务区块链”汇总各子 Agent 的状态与判定，生成独一无二的“决策 DNA”，在保障审计穿透性的同时化解公众信任赤字。 - [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) AI 治理已从单纯的数据溯源演进为对多智能体交互路径与任务成效的追踪，通过链上度量证明在监管与客户之间构筑不可篡改的信任体系。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=4s) AI 的信任危机与高风险决策背景 在人工智能迅速演进的浪潮中，业界正面临一个核心拷问：我们何时才能真正信任这项技术，让它从实验室里的“科学实验”转变为切实改变人类生活的生产力工具？在金融服务、电信和保险等受到高度监管的行业中，AI 模型如果涉及决定个人能否获得房屋贷款，或是孩子能否进入心仪的大学，这种决策将对个人的未来财务和人生轨迹产生深远影响。信用风险评估（Credit Risk）与欺诈检测（Fraud Detection）等应用，在风险层级上与用于警务司法的 AI 同属“高风险 AI 应用”。 在缺乏严格治理的情况下，企业随时可能受制于产生幻觉（Hallucination）、存在偏见甚至得出错误结论的黑盒模型。企业必须具备溯源和剖析模型决策依据的能力。作为在 FICO 工作了 27 年并主导部署过数百个机器学习模型的首席分析官，Scott Zoldi 指出，解决这一信任问题的关键，在于建立满足伦理测试、可解释性（Explainability）、可理解性（Interpretability）以及可审计性（Auditability）的全流程治理机制。 - **高风险 AI 的界定**：信贷审批、欺诈防御与 AI 警务同属高风险类别，直接决定个人的重大财务和生活成果。 - **监管与合规核心要求**：必须涵盖伦理测试、决策可解释性、逻辑可理解性以及完整的审计链条。 --- ### [01:40](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=100s) 确立企业级负责任 AI 开发标准 大约十年前，业界开始认识到：任何部署可能对人类生活产生重大影响的 AI 的组织，都必须制定明确的企业级 AI 开发标准（Corporate AI Development Standard）。负责任的 AI 不能仅仅停留在口头誓言或道德倡议上，而必须通过具象化的规范来证明其落地能力——明确规定哪些算法被允许使用、采用何种可解释性技术，以及如何对模型进行监督与强制执行。 对于拥有数百名数据科学家的组织而言，缺乏统一标准是极其危险的。数据科学团队的目标不应是“用各种炫酷的新算法互相炫耀”，而是以负责任的方式交付能对业务产生实际价值的模型。如果完全放任自由，组织内可能会出现 300 种各行其是的建模方式，其中绝大多数可能并不符合组织层面的合规要求。通过将组织内顶尖人才的智慧固化为统一的开发标准，企业能够根据监管政策的变化集体调整技术路线，确保所有模型在算法准入、伦理测试与解释方法上保持高度一致。 ```mermaid flowchart TD accTitle: 负责任 AI 开发标准制定与执行流程 subgraph S1["顶尖专家与治理团队"] A["定义算法准入范围"] B["确立可解释性与伦理指标"] C["响应监管调整标准"] end subgraph S2["统一开发标准规范"] D["企业级 AI 开发标准<br/>(消除300种随意做法)"] end subgraph S3["开发团队落地"] E["开发者遵循规范构建"] F["严禁算法自嗨与科学实验"] end A --> D B --> D C --> D D --> E D --> F ``` --- ### [03:36](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=216s) 利用区块链实现规则的强制约束与不可篡改性 即便制定了完善的开发标准，核心挑战依然在于“如何强制执行（Enforce）”。区块链技术在这里展现了其超越加密货币范畴的核心价值——它本质上是一套具备强制力的数字化契约系统。在过去，管理者只能依赖数据科学家在纸质文件上手写签名来确立合规契约；而如今，私有分布式区块链能够以技术手段确立不可撤销的责任体系。 区块链的不可篡改性（Immutability）来自于每个区块对前序数据的密码学依赖。如果在后续阶段有人试图修改链上的任何历史参数或记录，整条链条就会立刻断裂失效。这意味着，如果在项目初期设定了严苛的合规指标，而在开发后期发现指标过高难以达成，团队绝无可能在事后悄悄调低标准以混过审查——“达不到要求就是没达到，没有任何篡改标准的可能”。这种不可篡改的特性为模型的构建过程提供了铁一般的问责机制与诚实度保障。 - **从纸质契约到链上契约**：将传统的开发人员纸质签字合规承诺升级为区块链智能契约。 - **不可篡改性带来的问责制**：后置修改历史要求会导致区块链断裂，杜绝事后为了达标而降低要求的侥幸行为。 --- ### [04:53](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=293s) 链上记录模型生命周期全貌与发布准入机制 在 2018 年提交相关专利时，FICO 便将区块链构想为一部记录模型构建全过程的“完整电影记录”，而非仅仅记录一个最终是否合格的静态结果。从确定构建一个信贷审批模型开始，链上就会记录初始需求，包括针对伦理测试设定的特定指标（如差异性度量 Disparity Metric）。在整个开发生命周期中，系统围绕三大核心角色展开协同：开发者（Developer）、测试者（Tester）与审批者（Approver）。 区块链不仅记录成功跨越的里程碑，更忠实记录整个过程中遇到的技术挑战、架构调整的决策路径，以及每一个环节的具体责任人签字。当模型在多年后上线运行且面临外部审查时，企业无需为“当年是谁拍板、为何这样设计”而苦恼，因为所有的治理凭证早在开发阶段就已实时写入区块链。在模型发布阶段，私有链直接充当“守门人”：唯有链上完整证明了开发、测试与审批的每一个环节都严丝合缝地达标，模型才被允许发布到生产环境。 ```mermaid flowchart LR accTitle: 基于区块链的模型准入与发布审查流程 subgraph Roles["三大核心协同角色"] R1["开发者 (Developer)"] R2["测试者 (Tester)"] R3["审批者 (Approver)"] end subgraph Chain["私有分布式区块链"] B1["初始业务与伦理指标"] --> B2["技术挑战与架构调整"] B2 --> B3["阶段性责任签字"] end subgraph Gate["发布准入审查"] G1{"链上记录是否<br/>完全达标？"} G2["准予上线生产部署"] G3["严禁发布 / 驳回整改"] end Roles --> Chain Chain --> G1 G1 -- 是 --> G2 G1 -- 否 --> G3 ``` - **全生命周期动态记录**：记录开发过程中遭遇的挑战与技术路线转向，提供完整的审计上下文。 - **硬性门禁控制**：如果无法在链上拿出满足全部治理标准的不可篡改证明，无论开发耗时多久，模型绝对禁止上线。 --- ### [07:19](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=439s) 运行期模型监控与“谦逊 AI（Humble AI）”理念 合规模型的构建只是第一步，更关键的是确保模型在投入实际使用后依然保持负责任的状态。当前行业在模型监控上面临巨大挑战，许多团队往往只停留在监测数据漂移（Data Drift）等可能与业务表现弱相关的浅层指标上。事实上，设计模型的架构师才是定义监控维度的最佳人选。 以基于神经网络的模型为例，设计者清楚应当监控哪些潜在特征（Latent Features）被激活、激活的频率与数值范围如何，以及哪些特征会在特定元组（Tuples）组合中同时触发。区块链伴随模型整个生命周期流动，不仅固化了模型的构建参数，还内置了运行时监控规则与退役阈值——明确规定模型在何种情况下必须被强制下线，或在遇到何种交易模式时应主动放弃决策并交由人工处理。这种在不确定场景下主动承认局限性的设计，正是“谦逊 AI（Humble AI）”的核心所在。 - **深层特征级监控**：超越浅层数据漂移，直接深入神经网络内部监测潜在特征的激活范围与组合触发状态。 - **谦逊 AI 机制**：预设规则让模型在超出能力边界或处于特定高风险交易场景时主动忽略自身输出，确保系统的公平与安全。 --- ### [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) 生成式 AI 的合规瓶颈与聚焦语言模型（FLM） 随着生成式 AI 与大语言模型的爆发，各大科技巨头纷纷投入算力竞赛并争相刷新基准测试评分。然而在受监管的金融服务领域，生成式 AI 早已应用数十年，当前开源或商用基础模型（Foundation Models）的“黑盒”属性却带来了严重的合规挑战。对消费者进行信用评估的前提是必须完全理解其训练数据的构成；如果直接调用外部无法溯源的基础模型，企业就根本无法证明模型是否对特定人群存在隐性歧视，也无法保证特定业务场景下的评分覆盖率。 因此，受监管行业正在转向从零自建专有模型的发展路径。FICO 提出了“聚焦语言模型”（Focused Language Model, FLM）的构想：构建模型的第一步必须明确其专属业务边界。如果一个模型专门用于处理金融服务业务，那么它就不应该包含制作鸡汤的菜谱、辣妹合唱团（Spice Girls）的八卦或波兰的历史百科。剔除一切与业务无关的冗余语料，不仅能从源头上抑制模型幻觉与不必要的偏见输入，而且这一由业务驱动的“聚焦边界”设定，本身就是写入区块链的第一道核心决策。 - **通用基础模型的监管困境**：因训练数据与权重生成过程不可知，无法在受监管领域自证无歧视与合规性。 - **聚焦语言模型原则**：精准剔除与垂直领域无关的知识语料，通过缩小上下文范围从根本上减少模型偏见与幻觉。 ### [10:48](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=648s) 构建聚焦语言域数据集与区块链全流程存证 决定模型应当承担何种职责并非单纯的数据决策，而是一项严肃的业务决策。确立这一业务目标后，团队便可以接入底层数据源（例如从 Common Crawl 海量网页中提取样本数据）。在此阶段，研发人员会引入解码器模型（Decoder Model）对样本数据进行打分：得分高的内容意味着与既定业务焦点高度契合；得分低的内容则说明完全不相干——用 Scott Zoldi 生动的比喻来说，低分内容充斥着像“辣妹组合（Spice Girls）或鸡汤食谱”这类偏离主题的杂音。在这一过程中，业务焦点的定义、初始采样方案、研究所选用的解码器模型，以及最终打好分数的样本集，全部都被不可篡改地记录在区块链上。 随后，团队基于解码器模型的评分逻辑训练出一个轻量高效的编码器模型（Encoder Model），用于将学到的判别规则固化下来，并对全部可用数据进行大规模评分与筛选。经过这道工序筛出的高相关性语料被称为“聚焦语言域数据集”（Focused Language Domain Dataset）。由于从焦点设定、样本采样、解码器打分到编码器过滤的全流程都记录在链，整个数据流水线具备完全的可控性与可审计性。企业可以清晰验证数据在地理分布上是否对怀俄明州有足够代表性，或在人口属性上是否存在针对非裔美国人与白人群体的系统性偏见。在强监管行业中，模型必须满足安全、公平和可信的严苛底线，而链上全流程追溯正是实现这种审计与控制的基石；打好这个基础后，工程师才能放心地调出 PyTorch 开始训练基础领域模型。 ```mermaid flowchart TD accTitle: 聚焦语言域数据集与链上审计流程 subgraph Data_Pipeline["数据筛选流水线"] A["定义业务焦点 Focus"] --> B["Common Crawl 样本采样"] B --> C["解码器模型打分<br/>筛选高相关语料"] C --> D["构建高效编码器模型"] D --> E["大规模数据评分过滤<br/>聚焦语言域数据集"] end subgraph Chain_Audit["区块链不可篡改存证"] BC[("私有区块链底层")] A -.-> BC B -.-> BC C -.-> BC D -.-> BC E -.-> BC end E --> F["启动 PyTorch 训练领域模型"] ``` - **全流程链上资产**：业务焦点规格书、数据抽样算法、研究用解码器模型版本、样本评分结果与编码器参数权重全部上链。 - **合规审计能力**：在模型训练前即可主动评估地域代表性（如怀俄明州样本量）与群体公平性（如非裔与白人群体样本分布），杜绝黑盒训练。 --- ### [13:30](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=810s) 任务微调、合规硬化与生成式 AI 治理 仅仅训练出一个领域模型还远远不够，因为它此时只反映了特定语料的分布特征，并不具备执行具体业务任务的能力。模型必须接受任务导向型微调（Task Training），以满足具体的业务与合规诉求。以澳大利亚银行业为例，澳大利亚证券与投资委员会（ASIC）制定了近九项针对银行如何与提出诉求的客户沟通的严格法规。开发团队必须将这些合规条款转化为专门的训练词元（Tokens），并与业务部门深度协同，搜集真实业务场景中处理得极其出色以及违规失败的典型案例，提炼为高纯度的“种子数据”（Seed Data）。 从这些宝贵的种子数据出发，系统会合成生成数十亿词元的高质量合成数据，对模型展开持续的强化训练。Scott Zoldi 强调，这种工程方法与目前流行的轻率做法有本质区别：企业绝不能仅仅依赖一段 Prompt，让大语言模型去“扮演一个熟知 ASIC 法规的客服代表”并祈祷它做出正确决策；相反，企业必须通过种子数据和合成数据对模型进行行为“硬化”（Hardening），迫使模型严格按照预期规范行事。所有训练数据、合规规则与调优历程同样全部写入区块链，配合部署阶段的“AI 信任评分”（AI Trust Scores）等二级监控模型，实时判断其生成内容是否值得信赖。当监管部门、治理委员会或客户日后提出质疑时，链上记录能够彻底消除“开发者当时到底在想什么”、“训练语料中到底有没有包含哈萨克斯坦”、“残疾人群体是否遭受偏见”等历史悬案，为关乎人类重大福祉的决策提供无可辩驳的证据链。 - **拒绝角色扮演式 Prompting**：严禁用大模型虚构客服角色代替合规控制，必须使用业务最佳实践的种子数据进行深度硬化。 - **数十亿级合成词元**：基于少量金牌真实样本拓展海量合成数据，定向覆盖合规边缘案例。 - **事后问责机制**：为未来监管调查、客户维权及模型治理审查提供即时、不可推卸的技术复核依据。 --- ### [16:06](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=966s) 智能体（Agentic AI）的失控挑战与 FICO Agent 区块链约束 生成式 AI 固然复杂，但行业正加速迈向由自主智能体构成的 Agentic AI 时代。在 Scott Zoldi 看来，这是一场让系统复杂性成倍剧增的“新型科研项目”。与传统单向输出的模型不同，自主智能体具有环境感知、调用工具、动态规划执行路径以及根据任务反馈自主迭代自适应的特性。这带来了巨大的系统风险：智能体在某一时刻被调用时所感知到的瞬时外部环境极难复现；倘若缺乏严格边界，智能体完全有可能在互联网上失控乱窜、抓取黑客入侵指令或做出危险举措。 为了防止自主 AI 演变成无法无天的“狂野西部”，FICO 提出了“FICO Agent 区块链”（FICO Agent Blockchain）架构。企业不能任由智能体自主决定行动策略，必须通过链上智能合约与规则字典对其进行强约束：明确限定其允许调用的特定工具白名单、获准通行的执行路径以及严格的操作权限级别。在运行监控端，区块链会精确锁定智能体的内部认知状态、实际引用的数据源及每一步合规指标。正如 Scott Zoldi 举例指出的那样，系统如果需要调取关于他的个人信贷数据，合法路径只能去他的发卡行或已知可信数据源提取，绝不允许智能体擅自调用来自罗马尼亚不知名服务器的信息。对于直接关乎个人财产与金融健康的严肃决策，这种技术严谨性不可或缺。 ```mermaid flowchart LR accTitle: FICO Agent 区块链的执行路径与权限约束 subgraph Agent_Control["约束规范 (事先定义)"] T["工具调用白名单"] P["合规执行路径"] A["操作权限级别"] end subgraph Runtime["智能体运行时"] Env["瞬时外部环境"] --> AG["自主智能体 Agent"] AG --> Dec{"路径与权限核验"} end subgraph Agent_BC["FICO Agent 区块链"] BC1[("记录瞬时环境状态")] BC2[("存证内部思维与引用源")] BC3[("记录执行日志与合规状态")] end Agent_Control ==> Dec Dec -- "违规路径 (如不明海外源)" --> Deny["阻断执行"] Dec -- "合规路径 (如授信银行数据)" --> Allow["允许执行并沉淀结果"] Runtime -.-> Agent_BC ``` - **动态环境锁定**：由于调用时刻的实时网络环境瞬间即逝，必须在链上对智能体所见的环境状态打快照，确保可复现性。 - **自适应边界控制**：防止智能体为了追求“任务完成度”而脱轨自适应，强制其在既定的受限路径内演进。 --- ### [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) 决策 DNA：Agentic Task 区块链与多智能体系统信任体系 当场景从单一智能体演进到由数十个智能体协同运作的多智能体系统时，失控风险会被进一步放大。在现实业务中，合成数据与数字孪生早已不是新鲜概念——FICO 从 1992 年起就开始为全球大部分跨行银行卡交易构建反欺诈模型，而 Scott Zoldi 本人早在 27 年前加入 FICO 时就专注于利用合成建模来推演新型欺诈手法的演变与支付方式的采纳。面对复杂现实，多智能体之间的级联调用极易引发连锁故障。例如，一名商旅用户将出差行程规划完全交托给 Agentic AI，事后账单上却莫名出现一笔 24 美元的网约车费用和高达 800 美元的机票，如果无法穿透层层智能体追溯其决策依据，信任将荡然无存；而对于那些申请首套房抵押贷款的年轻夫妇来说，一旦背后的智能体网络产生误判，将对其一生带来沉重打击。 德勤（Deloitte）的研究数据显示，全社会对自主 AI 的信任赤字极为严峻：仅有 3% 的“狂热信徒”愿意将控制权毫无保留地交给智能体；多达 20% 的人群坚决拒绝此类技术；而绝大多数人则坚持要求人工全程介入每一个裁决环节。但倘若事事都需要人肉审查，Agentic AI 所承诺的高效与自动化将化为泡影。为破解这一僵局，FICO 提出了正在推进的蓝图——构建层级化的“智能体任务区块链”（Agentic Task Blockchain）。在现有的信用评分体系中，客户被拒绝时可以拿到具体的分数、拒绝代码（Reason Codes）与数据出处，从而进行有效申诉与信用修复；未来的多智能体决策同样需要这种穿透性与可解释性。在一个由 20 个智能体协同做出的重大决策中，这 20 个智能体各自底层的私有区块链将全部向顶层的“任务区块链”如实汇报其运行状态与决策分支。这套架构构成了该次判决不可磨灭的“决策 DNA”（Decision DNA），让复杂的自主系统不仅能够自我运转，更能随时被人类复盘、拆解与完全审计。 - **多智能体层级区块链**：每个参与执行的子智能体拥有独立的链上运行时记录，顶层“任务区块链”汇总所有子智能体的输入输出与逻辑判定。 - **公众信任与效率平衡**：避免“3% 盲目信任 vs 97% 要求全流程人工审查”的两极化困境，用可追溯的审计机制释放智能体的自动化潜能。 - **“决策 DNA”的未来愿景**：未来社会中每一项影响人类关键利益（如授信、抵押贷款、风险定价）的重大自动化裁决，都将绑定一段独一无二、终身可溯源的区块链指纹。 ### [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) 智能体交互溯源与建立 AI 信任的治理架构 区块链在 AI 治理中带来的不仅仅是基础数据来源的溯源（provenance），更核心的价值在于记录多智能体系统（Agentic AI）中各组件之间的交互路径与执行成效。随着大模型和智能体架构的发展，单纯追踪“数据来自何处”已经无法满足监管与合规的要求；企业必须掌握每个 Agent 是如何协同交互的、各个任务单元是如何流转推进的，以及各项任务究竟以何种成功率完成。通过在底层引入区块链的不可篡改记录，系统能够忠实还原智能体生态中每一步协作的来龙去脉。 当前在 AI 与生成式 AI（Generative AI）领域中，“信任危机”依然根深蒂固，整个行业距离彻底解决这一难题仍有很长的路要走。面对这种现状，行业最急需攻坚的关键课题之一，就是围绕 Agentic AI 建立起严密的治理结构。单凭模型厂商的口头承诺无法换来信任，企业必须依赖技术手段在底层为复杂的自主协作构筑约束框架与审计依据。 ```mermaid flowchart TD accTitle: Blockchain and AI Trust Framework subgraph S1["智能体协作追踪 (Agentic AI)"] A["智能体交互轨迹记录"] --> B["任务完成状态与度量"] end subgraph S2["区块链治理层 (Blockchain)"] C["不可篡改审计证据链"] D["负责任 AI 标准校验"] end subgraph S3["合规与信任结果 (Trust & Accountability)"] E["内部标准严格对齐"] F["客户信赖与合规达标"] end B --> C A --> C C --> D D --> E D --> F ``` - **双重可信度量**：区块链与 AI 的结合能够建立双向的信任与问责机制（trust and accountability）——对内能够强制对齐开发标准与合规要求，对外则为客户提供透明度，证明模型完全是“严格按声明的方式构建完成”。 - **满足合规与落地证明**：借助链上留存的度量证明（proof of measure），企业在面对严苛的行业监管指引（regulatory guidelines）时能够从容应对，证明自身具备无可挑剔的治理与审计链条。在演讲的最后，主讲人表示时间已到，希望本场关于区块链赋能负责任 AI 的探讨能够为听众带来切实的思考与启发。 ## Notable quotes > [03:36](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=216s) 达不到要求就是没达到，没有任何篡改标准的可能 — Scott Zoldi > [08:54](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=534s) 如果一个模型专门用于处理金融服务业务，那么它就不应该包含制作鸡汤的菜谱、辣妹合唱团（Spice Girls）的八卦或波兰的历史百科 — Scott Zoldi > [13:30](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=810s) 企业绝不能仅仅依赖一段 Prompt，让大语言模型去“扮演一个熟知 ASIC 法规的客服代表”并祈祷它做出正确决策 — Scott Zoldi > [16:06](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=966s) 系统如果需要调取关于他的个人信贷数据，合法路径只能去他的发卡行或已知可信数据源提取，绝不允许智能体擅自调用来自罗马尼亚不知名服务器的信息 — Scott Zoldi > [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) 仅有 3% 的“狂热信徒”愿意将控制权毫无保留地交给智能体；多达 20% 的人群坚决拒绝此类技术；而绝大多数人则坚持要求人工全程介入每一个裁决环节 — Scott Zoldi > [18:08](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1088s) 未来社会中每一项影响人类关键利益（如授信、抵押贷款、风险定价）的重大自动化裁决，都将绑定一段独一无二、终身可溯源的区块链指纹 — Scott Zoldi > [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) 区块链在 AI 治理中带来的不仅仅是基础数据来源的溯源（provenance），更核心的价值在于记录多智能体系统（Agentic AI）中各组件之间的交互路径与执行成效 — Scott Zoldi > [21:41](https://www.youtube.com/watch?v=BqaH9gAhJ3g&t=1301s) 严格按声明的方式构建完成 — Scott Zoldi ## Mentioned Scott Zoldi, FICO, Common Crawl, PyTorch, ASIC (澳大利亚证券与投资委员会), Deloitte (德勤), Spice Girls, Focused Language Model (FLM), FICO Agent Blockchain, Agentic Task Blockchain, Humble AI ## Video description Scott Zoldi - Chief Analytics Officer, FICO AI has a trust problem. Research shows that most Americans are more concerned than excited about AI in daily life, and that concern is well-founded: without proper governance, AI models can introduce bias, produce unexplainable outcomes, and be nearly impossible to audit after the fact. To build genuine trust in AI, organizations must ensure that their models are interpretable, auditable, and enforceable, and blockchain technology offers a practical, operational path to achieving all three. In this session, I will discuss how FICO uses blockchain-based AI models to help reduce hallucinations, while ensuring adherence to our Responsible AI standards. By embedding governance rules directly into a private blockchain, we have created an immutable record of every decision made throughout the model development lifecycle, from variable selection and algorithm approval to ethics testing and final sign-off. [Source](https://www.youtube.com/watch?v=BqaH9gAhJ3g)

Source quotes

[3:36] Scott Zoldi: 达不到要求就是没达到，没有任何篡改标准的可能 [8:54] Scott Zoldi: 如果一个模型专门用于处理金融服务业务，那么它就不应该包含制作鸡汤的菜谱、辣妹合唱团（Spice Girls）的八卦或波兰的历史百科 [13:30] Scott Zoldi: 企业绝不能仅仅依赖一段 Prompt，让大语言模型去“扮演一个熟知 ASIC 法规的客服代表”并祈祷它做出正确决策 [16:06] Scott Zoldi: 系统如果需要调取关于他的个人信贷数据，合法路径只能去他的发卡行或已知可信数据源提取，绝不允许智能体擅自调用来自罗马尼亚不知名服务器的信息 [18:08] Scott Zoldi: 仅有 3% 的“狂热信徒”愿意将控制权毫无保留地交给智能体；多达 20% 的人群坚决拒绝此类技术；而绝大多数人则坚持要求人工全程介入每一个裁决环节 [18:08] Scott Zoldi: 未来社会中每一项影响人类关键利益（如授信、抵押贷款、风险定价）的重大自动化裁决，都将绑定一段独一无二、终身可溯源的区块链指纹 [21:41] Scott Zoldi: 区块链在 AI 治理中带来的不仅仅是基础数据来源的溯源（provenance），更核心的价值在于记录多智能体系统（Agentic AI）中各组件之间的交互路径与执行成效 [21:41] Scott Zoldi: 严格按声明的方式构建完成

Source quotes

[0:04] Scott Zoldi: impress each other with different algorithms [5:10] Scott Zoldi: movie [7:19] Scott Zoldi: Humble AI [8:54] Scott Zoldi: focused language models [8:54] Scott Zoldi: how to make chicken soup [8:54] Scott Zoldi: history of Poland [10:48] Scott Zoldi: the Spice Girls or chicken soup [13:10] Scott Zoldi: pretend to be a customer service representative who's aware of ASIC [15:19] Scott Zoldi: wild west [18:08] Scott Zoldi: zealots [18:08] Scott Zoldi: Decision DNA [21:41] Scott Zoldi: proof of measure

Original transcript

[0:04] Thank you for being part of this afternoon  session. I appreciate it very much. I'm going to [0:08] talk about the intersection of AI and blockchain  and why we really need to consider leveraging [0:13] these two fundamentally different technologies  to establish trust. A little bit about me. Um, [0:20] I've been with the company FICO for 27  years. Uh, during this time I have deployed [0:27] hundreds and hundreds of AI machine  learning models within financial services, [0:30] telecommunications, insurance companies. Companies  that require trust and have to meet regulations, [0:37] including ethics, testing, explainability,  interpretability, and auditability. [0:43] Organizations that have huge model  governance requirements, right? And, [0:48] you know, with all the evolution of AI, right?  There's been a lot of challenges in terms of when [0:53] will we be able to trust this technology, when  will we be able to deploy it, and when will it [0:56] move away from being a science experiment  to a useful tool that impacts our lives? [1:02] This slide kind of sums it up, right?  Um, we have a trust issue with AI. [1:09] If AI is making a decision about whether or not  you get a home loan, or whether or not your child [1:14] goes to the college that they want to go to,  it's a serious, serious impact, and you have [1:19] to have an ability to ensure that that AI was  built appropriately, and that you can triage [1:24] what that went into that decision. If you  can't, right? Then you're at the mercy of [1:30] an AI that may hallucinate, may be biased, right?  May be wrong. So, how do we establish trust? [1:40] Well, about 10 years ago I talked about the fact  that every organization that's going to deploy AI [1:46] that impacts a human life in a substantial way,  right? And things like credit risk and fraud, [1:51] um, are considered high-risk AI applications as  long in the same category as policing with AI. [1:59] Why? Because they make big impacts in whether the  future sort of financial outcomes of your life. [2:04] These organizations have to come up with  a corporate AI development standard. [2:08] Essentially, right, moving beyond  an oath of being responsible with AI [2:14] to demonstrating how they will do  that. What algorithms are allowed, [2:19] what explainability methods will be used,  how will that be enforced and monitored. [2:25] Now, this is a challenge. I don't know about each  of your organizations. My organization is have [2:29] hundreds and hundreds of data scientists, right?  So, the very first thing we had to do is say, [2:33] we are not here to impress each other with  different algorithms. We What we are doing [2:38] here is building AI in a responsible fashion  so that we can deliver models that impact [2:43] these businesses. And so, part of this was to  codify what is this AI development standard. [2:50] What does it involve? It involves saying that  these are the algorithms that we at FICO would [2:53] use. This is how we do explainability,  this is how we do interpretability, [2:57] this is how we're going to do ethics testing.  Now, the value and beauty of this is that [3:02] all the best minds in the organization focus  on the development standard, right? They make [3:07] decisions about when to move to a different type  of AI technology or or when to adjust uh based on [3:13] way regulations adjusting. If you left that open  to each and every data scientist, you'd have 300 [3:18] different ways of doing AI. Most of them, right,  would be the wrong way or maybe not the right [3:23] way that you'd want from an organizational  perspective. Now, you can have a standard, [3:30] but how do you enforce the standard?  And this is where blockchain comes in. [3:36] Blockchain is perfectly designed to enforce  a corporate responsible AI standard. [3:43] Enforce is the right word here, right? Blockchain  is not just about a record that's immutable, but [3:48] it's around enforcing that we can show that models  and AI were developed to that particular standard. [3:55] Now, for those of you unfamiliar with blockchain,  right, many many of you may think of it from [3:59] a cryptocurrency perspective, um it's a very  powerful technology because it is establishing [4:06] a contract. In the old days, before we had  blockchain, I actually had my scientist sign [4:10] a piece of paper. Right, that was a contract  between me and my scientist they were going to [4:14] meet the requirements associated with these model  builds. But now we can use blockchain to do that. [4:19] The beauty of it is is as you add information  to the blockchain, right, um it is predicated [4:25] on all the previous um pieces of information  within those blocks. So, if you went in later [4:30] on and changed anything in that blockchain, you'd  break the entire blockchain. So, it's immutable. [4:37] So, if I set a set of requirements for a build  and later on it was too hard to meet, well, tough, [4:42] right? We did not meet the requirements, right?  There's no ability to change those requirements. [4:48] That's accountability and that's, you know,  honesty in terms of how we develop these models. [4:53] That immutability is really critical because it  allows us to make sure that we have confidence [4:58] that when we want to ask a question about how  the AI was developed or how to monitor that AI, [5:02] that that chain is in has integrity and  the information recorded to it is accurate. [5:10] Now, I filed the first of the patents around AI  blockchains in the year 2018. Okay, so this is not [5:17] brand new. There are some new ideas I'll share  with you towards the end of the presentation. [5:22] But what's really important here is that the  way you want this to be is you want it to be [5:27] essentially a movie, right? You want it to be a  record of what will happen during a model build. [5:33] So, for example, you start off and  say this is a model for lending. [5:37] This model has these sets of requirements. One  of these requirements is around ethics testing. [5:42] We define a metric called disparity. We  want to measure disparity. Right? And you [5:47] have three major roles. You have a developer,  you have a tester, and you have an approver. [5:52] And so, what you record on the blockchain is  not just that you met the the standard, but you [5:57] you record the journey. And that emphasizes where  do we have challenges? How did we have to pivot? [6:04] And who is signing off for each and every  step? Right? This is going to be important [6:10] because these models go out in production  and when we have to ask the question, [6:13] how it was it built? Who made the decision? Why  did we do that? Then none of that gets persisted. [6:19] But if you define what that needs to be on the  blockchain up front, you have what you need for [6:23] a model governance function. In fact, much  of model governance goes away because you've [6:27] preloaded the information that you need in the  blockchain while you're developing this model. [6:34] Now, I said I was an enforcer, right? Essentially,  the blockchain is associated with the model build. [6:41] So, you you review that blockchain, which  is on this private distributed chain. [6:46] You look at those requirements.  You have to see all the steps you [6:50] met in terms of the development, the testing, the  approval. And if you cannot demonstrate on that [6:55] blockchain that every one of those things were  met satisfactory, you don't release the model. [7:00] That's the enforcement piece of using blockchain  and AI together as a tool. So, that if we release [7:07] 300 models or 400 models, we know they are  developed to that standard and we have proof. [7:12] And that proof is within the blockchain,  it's immutable, and it's accessible later. [7:19] Moreover, blockchain plays another role.  Right? In some sense, it doesn't matter if [7:24] you build a responsible model, right? It matters  whether when you're using it, is it responsible? [7:30] Which means that we have to address also  model monitoring and other concepts like that. [7:35] Today, model monitoring is one of the things that  is one of the biggest challenges in our industry. [7:39] Many people don't know how to model how to  actually monitor these models appropriately. [7:43] They look at data drift and things like this,  which may or may not be at all relevant. But the [7:48] model designer who built that model is also the  perfect person to define what is it to monitor. [7:55] If it's a neural network based model, they're  going to look at which latent features are firing, [7:59] how often, what is the range of  activation, and which latent features [8:05] fire at the same time or in different tuples  of activity. They're also going to talk about [8:11] when do you have to pull that model, right?  So, the blockchain moves with the model. [8:17] It discusses how the model gets developed,  but also establishes how to monitor that [8:23] model to ensure that while it's in production, it  continues to behave in an ethical and fair way. [8:32] Now, that's traditional AI. This has been in  process for a long time. It's one of these [8:37] things that we do to make sure that we follow  the model development standard. But moreover, [8:42] right, that we know when to the model needs to  be replaced, or for certain types of transactions [8:48] where the model should be ignored, right?  That's the concept of humble AI. Now, [8:54] everyone's excited about generative AI. So, let's  look at what this looks like for generative AI. [9:00] Now, the first thing I'll say about  generative AI is that, you know, [9:03] it's not new. I think we all know that. What's new  is, you know, the computing power and some very, [9:09] very large companies generating lots of foundation  models and comparing against each other and [9:13] trying to beat each other's benchmarks,  right? But the use of generative AI has been [9:19] been used in the financial services industry and  other industries for many decades already, right? [9:24] And so, it's a concept of how do we use generative  AI responsibly? One of the tenets of any model [9:33] that you use to score a consumer is that you have  to understand the data on which it was built. [9:38] It means that you have to be involved in  the building of the model. And that's a [9:41] fundamental problem when we look at some of these  foundation models out there which you can download [9:46] or you can get access to, but you don't truly  understand how they were built and therefore [9:52] you don't truly understand whether it is  discriminatory towards certain groups of people [9:56] or not, whether you have proper coverage for  the types of scoring that needs to occur. And so [10:01] in financial services organizations and many  organizations where you are regulated use of AI, [10:07] there's a departure. And that departure is  towards building your own models from scratch. [10:12] And we'll talk through what that looks like here.  So, at FICO, the way this works for us is that we [10:18] have something called a focused language model.  Okay. Now, what why is it focused language? [10:23] Well, the very first thing you do when you build  any model is you say, "What is the model supposed [10:26] to do?" That's the focus. Okay. So, if I'm going  to if this is supposed to know about financial [10:31] services, then it shouldn't know how to make  chicken soup and it shouldn't know about the [10:34] Spice Girls or it shouldn't know the history  of Poland. Like that is not relevant. It's not [10:39] going to be helpful. It's going to create bias in  the in the outcomes. And so, the very first thing [10:43] is to define the focus. That's the first thing in  the blockchain. Okay. That's a business decision. [10:48] It's not a data decision. That's a business  decision around what that model is supposed to do. [10:53] Now, once you've done that, right, you go to  the next step. And the next step would be, well, [10:58] you know, let's let's go to whatever our data  source is. Here the data we may want to use is, [11:02] you know, what is the all the information  available to us through Common Crawl. [11:07] You can pull a sample of that  data and start to work with that. [11:11] And if you work with that data, right, you can  apply a decoder model to score. Now, what what do [11:17] you do here? What you're basically doing is you've  uh placed on the blockchain now the fact that what [11:22] your focus is, what is it supposed to be. You've  placed on the blockchain how you did the initial [11:26] sample. You placed on the blockchain which decoder  model you're using from a research perspective. [11:31] And you've placed on a blockchain, what is the  scored sample data set? That score, if you have a [11:38] high score, good score, that means it's very well  related to the focus. If you have a low score, [11:43] it means it's the Spice Girls or chicken soup  and things that are not relevant to the focus. [11:49] From there, the next step in this is to build an  encoder model. That encoder model is essentially [11:55] the most efficient way that we codify the  learnings that we derive from applying the decoder [12:02] model. Also on blockchain. All of this  is auditable, all of it's changeable. [12:09] And the changeable part comes when we use that  encoder model to then score all the available data [12:15] and select out only the data that is of relevance  for the problem that we're trying to solve. [12:20] We'd call this a focus language domain data set.  It is directly a result of the focus, the sample, [12:30] the decoder, the scored data, and the development  of this encoder model. This data now is in your [12:38] full control. You can understand whether or  not it has enough representation for Wyoming. [12:43] Or you can test whether or not it is biased  towards African-American versus white. [12:49] Right? You can decide that the focus  is not precise enough, crisp enough, [12:54] and we're going to be better in the way that we  prescribe the sort of data that we want to build. [12:59] This is the level of auditability and  control that we demand in regulated [13:04] environments where we're trying to make sure  that models are trustworthy and safe and fair. [13:10] And then from there, you can build your domain  model. Pull out your pie torch, go to town. [13:19] Now, that model right now does nothing  other than codify what that data looks like, [13:23] right? And that data is specialized based on  the focus. And we can apply them task training. [13:30] Cuz we have to then go and say, "Well, listen,  actually, I need this model to do a job, [13:35] and that job here is going to be around customer  compliance." And so, let's look at Australia. [13:39] Australia has a huge amount of regulation, almost  nine different regulations around how you interact [13:44] with a banking client when they call you with  a concern. Right? You codify that. You put [13:49] You put that into as additional training  tokens related to ASIC regulation. [13:57] Then you sit down with the business again, and you  say, "Business, give me the examples of how this [14:02] is done well, and where it is not done well, and I  want to train this model to do only one thing now. [14:08] I want it to be as good as your very best examples  of this done today. Right? Or what I expect to do [14:15] in the future." That's called the seed data.  And from the seed data, you generate billions [14:20] of tokens of a synthetic data. The purpose of that  is to continue to train that model, so you get a [14:25] task-based model. This is not an agent where we  go ask a large language model to pretend to be [14:32] a customer service uh representative who's aware  of ASIC and try to make a good decision. This is [14:39] hardening [snorts] with examples from that seed  data how it has to behave. Also on blockchain, [14:46] also part of how we train these models to behave  the way we expect them to behave and to be able to [14:51] monitor. So, that's blockchain for generative AI.  We put the entire development process in there. We [14:59] have added assets that we need to ask questions  about that. And then in deployment, we can [15:04] determine what we want to use, things like AI  trust scores for and other technologies that we [15:09] really monitor these models to determine whether  we should trust the output. Doesn't mean the [15:12] output is right or wrong. It's whether we should  trust it. And as we usually secondary models. [15:19] Now, as I said, these These ideas  have been in play for quite a while. [15:23] What does it establish? Establishes that there  are no more questions about what was that person [15:27] thinking when they built the model. Did we  include Kazakhstan in this corpus or not? [15:32] Like did somebody sufficiently look at disability  and whether or not the model is going to be biased [15:38] towards that. It provides an opportunity to ask  questions that only come afterwards when customers [15:43] have challenges or regulators have a question,  when your model governance team has a concern. [15:50] I'm a big fan and believer that blockchain  is critical to this. It provides a level of [15:54] accountability and rigor um that is required  when you're going to allow a model to make a [15:59] decision about a human life. Now, what about a  gentle guy? We're doing a gentle guy. Everyone's [16:06] doing a gentle guy. It's the next science project,  right? Um it gets worse. It gets harder. So, what [16:14] are some of the challenges? No, I won't talk about  how, you know, it goes and posts on Reddit and and [16:20] gets information from hackers how to compromise  uh organizations. That's just bad practice, right? [16:27] But these post challenges, right? There are  specific data sources and what the providence [16:31] of that data is. There's tool usage that has to  be specified. There's execution paths that are [16:35] allowed or not allowed, right? There's real-time  environmental status, right? When an agent gets [16:42] called, it's going to respond to the environment  that it sees at that moment. And that moment is [16:45] very hard to replicate unless you've recorded  it. And there's this continual self-adaptation. [16:52] As they learn, any of these agents are allowed to  adapt based on their success at completing a task. [16:59] We've introduced this concept of a FICO agent  blockchain. And there's more complexity to it, [17:07] right? Where you specify what are the tools, what  are the execution paths, what are the permissions. [17:14] On the monitoring side, you talk about whether  it was within compliance. What was that internal [17:18] state? What were the references that were actually  used or not used in the data when it when it [17:22] executed its path? So, the fact that these AI now  have more autonomous, they have an ability to talk [17:29] to the world around them means that very often  we have to constrain what they're allowed to do. [17:33] We're not leaving it to like the wild west so that  each of these models can just decide what they'd [17:37] like to do because that is the very very difficult  to control as we've seen in the press repeatedly. [17:44] But, we want to say these are the allowable  paths. I'm not going to get information about [17:49] Scott that came from Romania. We're going to get  information from about Scott that came from his [17:54] bank or some other sort of known data source. Now,  this takes more rigor, right? But again, if we're [18:01] making decisions that impact humans and their  financial well-wellness, it's important to do. [18:08] Now, it's going to get harder even. That's  just one agent. What about agentic AI? [18:15] And I'll just give you an example. You have your  own examples, but you know, one of the things I [18:18] do all the time is focused on synthetic data  generation. Digital twin is a term that people [18:25] leverage and use more and more, but this has been  done time over time over time. When I joined FICO [18:30] 27 years ago, the first things I was doing is  building fraud models. FICO's been building models [18:35] for fraud monitoring most of the world's  inner card transactions since the year 1992. [18:41] To do that, we have to synthetically understand  what fraud looks like, how fraud will adjust, [18:46] right? How do a new payment method, how it will  be adopted. There are a lot of things that go into [18:52] successfully generating data. And so, this is just  an example, right? If she wanted to go to or to a [18:58] business trip, right? And she's going to give over  her ability to make those decisions to agentic AI, [19:03] there are lots of different considerations.  Each of those agents might be off the rails [19:07] or they might be completely within range, right?  But when she has a question about why there's a [19:12] $24 ride share and almost $800 airplane  ticket, she's going to need the answers. [19:18] And if she doesn't have a way to triage with those  answers, she's not going to trust this agentic AI. [19:25] So, where are we with trusting agentic AI  systems? Right? What if you were that couple [19:30] that were trying to get their first home? That's  the type of people that I'm thinking about are [19:34] people that have big fundamental decisions to  make and if AI goes wrong, it they're impacted. [19:41] Trust is a big issue. This is from Deloitte.  I don't think it's changed that much, [19:44] right? 3% are zealots and they love it and  they're, you know, they're going to give away [19:48] all their sort of control to agentic AI, but the  rest, right, are going to basically say, "Hey, [19:54] I need to be involved in every single  judgment, which means that we're not [19:57] going to move at the speed and the pace  that we would expect with agentic AI." [20:02] Another 20% um say forget about it, right? I'm not  there yet. Why? Because we do not have the right [20:08] frameworks. So, I don't have a solution for this.  We're not at this point yet in the evolution [20:14] of these technologies, but I'll  show you uh an idea that we have [20:18] that we will be implementing. Right? As we have  agents, we're going to have this concept of a [20:24] agentic task blockchain. Essentially,  for any meaningful decision, [20:29] right? There will be a blockchain of everything  that was involved in making that decision. So, [20:34] if there are 20 agents, there'll be 20 agents  that are reporting from their blockchains [20:37] the information that talks about how they  operated and whether it was successful or not. [20:44] This is going to be very different than the  way the world road looks today. Today, you [20:47] basically get a score, some reasons, right? Some  other outputs. And so, when you complain, like [20:52] I don't understand why I didn't get this loan,  you can go triage. Where did the data come from? [20:56] Who do I have to talk to? What do I have to do to  improve my credit worthiness? Right? These are all [21:01] things that we're sort of entitled to, right? If  we're going to be scored by automated decisioning. [21:06] In the future, we're going to need technology  that records this. I'm going to need technology [21:10] that can unwind it and make it explainable  and interpretable for our humans. But, the [21:15] first step is to set up the frameworks of these  blockchains and and hierarchies of blockchains [21:19] so that we can record that information if agentic  AI moves to where many people want it to be. [21:29] Essentially, right, that agentic  task blockchain is a decision DNA. [21:36] We may see a world in the future where every  every decision is associated with a blockchain. [21:41] So, you have that providence. Not a providence of  just where the data was, but the providence of how [21:45] each of these agents interacted with each other  and how successfully they completed their tasks. [21:53] We still have a trust issue in AI. Both  AI and generative AI, so we're far away. [21:59] But, I think this is one of the most  critical things that we need to be [22:01] thinking about are the governance structures  around things like agentic AI um and agent AI. [22:09] Blockchain and AI together be established  trust and accountability. Both your internal [22:15] accountability to a standard uh and also for your  customers to have an expectation that models are [22:22] built the way they say they're built, that you  have the proof of measure, that you can meet the [22:26] regulatory guidelines, and ensure that you have  the right governance around it. And with that, [22:32] I'm at time. I want to thank you and hopefully  I've given you something to think through.