Wiley Embeds Research Feeds in AI Systems to Cut Errors

Confident machines: Why AI's trust problem starts before the model ever runs with Wiley — Ai4

Ai4

Unchecked knowledge foundations cause compound failures in multi-step AI agents, requiring full-text peer-reviewed data to ensure reliable outputs.

AI-assisted summary of Confident machines: Why AI's trust problem starts before the model ever runs with Wiley — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published on YouTube by Ai4 on August 19, 2026, this presentation features Sue Weng-Cross, Director of Strategic AI Partnerships at Wiley. The video explores AI's trust problem, detailing how modern models fail when built on unreliable, non-peer-reviewed knowledge layers and proposing a foundation grounded in verified scientific data.

## TL;DR Sue Weng-Cross addresses the paradox of accelerating information overload and rapid AI adoption, pointing out that 90% of citations in major AI tools come from non-peer-reviewed sources like Reddit, Wikipedia, and YouTube. While the industry concentrates on the model and deployment layers, she argues that true trust requires a robust knowledge layer built on full-text, peer-reviewed research rather than surface-level abstracts or web scrapes. By grounding agentic multi-step workflows in Wiley's expert-curated knowledge feeds, organizations significantly reduce hallucinations, avoid compounding errors, and discover critical scientific insights.

## Atmosphere & style The presentation carries a professional, authoritative, and engaging tone, blending hard statistical evidence with lighthearted humor about infamous AI blunders like pizza glue and hallucinated legal cases. Weng-Cross presents a clear, structured argument that flips standard responsible AI conversations on their head, transitioning smoothly from alarming industry metrics to compelling real-world enterprise success stories. Watching the talk feels like an insightful, crisp executive briefing that delivers a urgent wake-up call regarding data quality in autonomous AI workflows.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) Sue Weng-Cross introduces the modern research dilemma: humanity produces 3.3 million scientific papers annually, with medical output doubling every 73 days. - [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) A survey by Wiley shows 89% of corporate researchers use AI and 86% expect ubiquitous adoption within two years, despite serious trust concerns. - [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) A study of 150,000 AI citations revealed that 90% originated from Reddit, Wikipedia, and YouTube, leading to ungrounded model outputs like adding non-toxic glue to pizza. - [04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) Corporate researchers (83%) and medical clinicians (74%) report major anxiety over hallucinations and data privacy issues. - [04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) In multi-step autonomous AI agent workflows, a 95% accuracy rate per step drops to a 36% success rate across a 20-step chain due to compounding errors. - [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) Trust in AI operates across three layers: Deployment, Model, and Knowledge, but the industry heavily neglects the foundational Knowledge Layer. - [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) A robust Knowledge Layer requires breadth of content, depth beyond basic paper abstracts, and explicit machine-readable trust signals. - [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) Wiley delivers curated full-text Knowledge Feeds via APIs and Model Context Protocol (MCP) standards to 7 out of the top 10 global pharmaceutical companies. - [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) A proof-of-concept with BioRelate showed that full-text scientific articles yielded 63% more novel biological relationships per article than abstract-only indexing. - [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) Wiley partnered with Microsoft Discovery to build a Life Sciences Research Agent that integrates into multi-agent workflows to streamline drug discovery.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) Information Overload and the AI Adoption Paradox

Sue Weng-Cross, Director of Strategic Partnerships at Wiley, introduces her role leading a team that builds partnerships with AI developers, technology platforms, and industry leaders. Her team's primary goal is integrating trusted scientific research into the AI tools that researchers and clinicians rely on to accelerate scientific progress. She frames the central dilemma facing modern research: humanity publishes 3.3 million scientific research articles every year. In medicine specifically, total research output doubles every 73 days, creating an overwhelming workload where a primary care physician would need 27 hours a day just to read and stay up to date with new clinical guidelines and evidence. Across every sector, individuals and teams are drowning in raw information while remaining starved for actionable insights.

This information bottleneck is driving rapid AI adoption across research-intensive fields. A survey conducted by Wiley among 1,000 corporate and academic researchers reveals that 89% of corporate researchers already use AI tools, and 86% believe AI adoption will be ubiquitous within the next two years. However, Weng-Cross emphasizes that widespread adoption does not imply these systems pull from trustworthy foundations. A June 2025 study analyzing 150,000 citations across ChatGPT, Perplexity, Gemini, and Google AI revealed that Reddit accounted for over 40% of all citations, followed by Wikipedia at 26% and YouTube at 24%. Collectively, 90% of AI citations originated from these three non-peer-reviewed platforms.

While platform communities like Reddit have value, they were never designed as rigorous, evidence-based foundations for scientific experiments or clinical care. When AI models fail to differentiate between joke posts and valid evidence, bizarre errors occur—such as Google AI recommending users add non-toxic glue to pizza sauce so the cheese won't slide off, a tip scraped directly from a humorous Reddit comment. Weng-Cross quips that while glue on pizza might be a minor tragedy for pizza lovers, ungrounded AI outputs are terrifying in contexts like drug interactions or patient care.

- **3.3 million** research articles published globally every year. - Medical research knowledge doubles every **73 days**. - Primary care doctors need **27 hours per day** to keep pace with new literature. - **89%** of corporate researchers surveyed by Wiley currently use AI. - **86%** anticipate widespread industry-wide AI adoption within two years. - **90%** of citations across major AI tools came from Reddit (40%), Wikipedia (26%), and YouTube (24%).

[04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) The Trust Gap and the Hidden Costs of Bad Data

Despite high usage rates, significant anxiety persists among professionals regarding system reliability. In Wiley's research, 83% of corporate researchers reported serious concerns about hallucinations, model inaccuracies, and data privacy. Similarly, while 81% of medical clinicians utilize AI in their workflows, 74% cite hallucinations as their leading worry. When asked what would encourage deeper integration, both groups overwhelmingly demanded validated safety and stronger data assurances. Organizations want AI, but hesitation surrounding untrustworthy outputs throttles full operational deployment.

``` UNTRUSTWORTHY DATA AT RETRIEVAL │ ┌──────────────────┼──────────────────┐ ▼ ▼ ▼ Financial Loss Lost Time Reputational Damage ($13M/yr avg) (Rework/Fix) (e.g. Mata v. Avianca) ```

The financial, operational, and reputational risks of flawed AI data are substantial. Data quality issues cost organizations an average of $13 million annually. Operationally, the promised speed efficiencies of AI are wiped out by human rework, verification, and error correction. Weng-Cross highlights the infamous legal case *Mata v. Avianca*, where attorneys submitted a court motion containing three completely hallucinated judicial precedents generated by AI. The model even confused Washington, D.C. with Washington State—an error Weng-Cross jokes that even she, as an Australian, easily spotted.

- **83%** of corporate researchers fear model hallucinations, inaccuracies, and privacy breaches. - **81%** of clinicians use AI, but **74%** rank hallucination as their primary concern. - Poor data quality in AI costs companies an average of **$13 million per year**. - Verification attempts break down in practice: a physician seeing 10 to 20 patients daily will not leave their primary software to manually cross-check AI claims across secondary databases. - Autonomous agent compounding error: a multi-step model that is **95% accurate** at each individual decision step drops to a **36% success rate** (64% failure rate) across a 20-step chain.

[08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) The Three Layers of Trust and Wiley's Knowledge Feeds

To address these vulnerabilities, Weng-Cross introduces the three distinct operational layers that dictate trust in an AI system: the Deployment Layer, the Model Layer, and the Knowledge Layer. The industry currently focuses most of its governance on the top two layers while ignoring the foundation.

``` +-------------------------------------------------------+ | DEPLOYMENT LAYER (Governance, Compliance, Audits) | <-- High Focus +-------------------------------------------------------+ | MODEL LAYER (Architecture, Fine-Tuning, Guardrails) | <-- High Focus +-------------------------------------------------------+ | KNOWLEDGE LAYER (Full-Text Grounding, Peer Review) | <-- Neglected +-------------------------------------------------------+ ```

1. **Deployment Layer:** Manages real-world execution, user access, regulatory compliance, privacy, and audit trails. 2. **Model Layer:** Focuses on underlying architecture, reasoning performance, guardrail alignment, and hallucination suppression. 3. **Knowledge Layer:** Sits at the bottom as the foundation. It dictates the exact facts and raw data the model can access before reasoning or safety filters execute.

Weng-Cross argues that no amount of deployment oversight or model tuning can make up for missing or corrupt baseline data; a model cannot reason over facts it never received. Evaluating the quality of this Knowledge Layer requires assessing three dimensions:

* **Breadth:** Moving beyond free web scrapes to index full-text research, specialized databases, and verified primary sources. * **Depth:** Reading past basic paper abstracts. Many AI systems rely exclusively on free abstracts, which omit critical methodology, experimental protocols, raw result tables, limitations, and causal linkages. * **Trust Signals:** Machine-readable metadata indicating peer-review status, citations, version histories, and post-publication corrections.

``` WILEY KNOWLEDGE LAYER ARCHITECTURE

[ 5M+ Research Articles | 2,000+ Journals | 30k Books ] │ ▼ [ Expert SME Curation & Indexing ] │ ▼ [ Industry Knowledge Feeds (APIs/MCP) ] │ ┌────────────────┴────────────────┐ ▼ ▼ [ Enterprise Workflows ] [ Autonomous AI Agents ] ```

As a major scientific publisher, Wiley leverages its catalog of over 5 million research articles across 2,000 journals and 30,000 books to feed specialized, full-text Knowledge Feeds. Curated by subject-matter experts, these feeds connect via APIs and Model Context Protocol (MCP) standards directly into agentic workflows for life sciences, healthcare, chemistry, and finance. Currently, 7 out of the top 10 global pharmaceutical companies integrate Wiley’s Knowledge Feeds into their R&D pipelines.

- Wiley partners with **900+ scientific societies** and reaches **50 million researchers** across **50+ countries**. - Full-text knowledge feeds deliver raw data alongside explicit **trust signals** (citations, peer-review tags, versioning). - Serves key research verticals: **Life Sciences, Healthcare, Chemistry, Engineering, and Finance**.

[13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) Real-World Proof Points: BioRelate & Microsoft Discovery

To demonstrate the quantitative impact of grounding models in full-text knowledge, Weng-Cross highlights two enterprise partnerships. In a proof-of-concept with biomedical firm BioRelate—which maps complex causal relationships for drug discovery—Wiley provided a small test sample of full-text life sciences articles to compare against PubMed Central's publicly available abstracts.

The full-text corpus revealed **63% more novel biological relationships per article** than abstract-only indexing. Across the small sample, BioRelate extracted **160,000 unique biological relationships** that were completely invisible when scanning abstracts alone. This proved that earlier reasoning limitations stemmed from missing access to primary data rather than model capability gaps.

In a second partnership with Microsoft Discovery, Wiley built a dedicated Life Sciences Research Agent. This agent searches, retrieves, and synthesizes Wiley’s pharmacology and life sciences databases. Rather than functioning solely as a isolated search bar, it operates inside an orchestrated network of multi-agent workflows. A drug discovery team can prompt the Wiley agent to generate research hypotheses and then hand those findings off to neighboring agents that generate lab protocols, interpret dry-lab simulations, and manage wet-lab experiments.

Weng-Cross concludes by reminding the audience that AI agents designed to eliminate information overload will only compound errors faster if grounded in poor data. True enterprise differentiation comes from grounding systems in reliable knowledge at the source. She invites attendees to visit Wiley at Booth 1144 or connect via LinkedIn to evaluate their AI data architecture.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) In medicine specifically, total research output doubles every 73 days, creating an overwhelming workload where a primary care physician would need 27 hours a day just to read and stay up to date with new clinical guidelines and evidence. — Sue Weng-Cross

> [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) Across every sector, individuals and teams are drowning in raw information while remaining starved for actionable insights. — Sue Weng-Cross

> [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) Collectively, 90% of AI citations originated from these three non-peer-reviewed platforms. — Sue Weng-Cross

> [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) When AI models fail to differentiate between joke posts and valid evidence, bizarre errors occur—such as Google AI recommending users add non-toxic glue to pizza sauce so the cheese won't slide off, a tip scraped directly from a humorous Reddit comment. — Sue Weng-Cross

> [04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) Organizations want AI, but hesitation surrounding untrustworthy outputs throttles full operational deployment. — Sue Weng-Cross

> [04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) Operationally, the promised speed efficiencies of AI are wiped out by human rework, verification, and error correction. — Sue Weng-Cross

> [04:18](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=258s) A multi-step model that is 95% accurate at each individual decision step drops to a 36% success rate (64% failure rate) across a 20-step chain. — Sue Weng-Cross

> [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) The industry currently focuses most of its governance on the top two layers while ignoring the foundation. — Sue Weng-Cross

> [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) No amount of deployment oversight or model tuning can make up for missing or corrupt baseline data; a model cannot reason over facts it never received. — Sue Weng-Cross

> [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) The full-text corpus revealed 63% more novel biological relationships per article than abstract-only indexing. — Sue Weng-Cross

> [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) This proved that earlier reasoning limitations stemmed from missing access to primary data rather than model capability gaps. — Sue Weng-Cross

> [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) AI agents designed to eliminate information overload will only compound errors faster if grounded in poor data. — Sue Weng-Cross

## Mentioned Sue Weng-Cross, Wiley, Ai4, ChatGPT, Perplexity, Gemini, Google AI, Reddit, Wikipedia, YouTube, Mata v. Avianca, Model Context Protocol (MCP), BioRelate, PubMed Central, Microsoft Discovery

## Video description Sue Weng-Cross - Strategic AI Partnerships, Wiley

Knowledge is compounding faster than anyone can keep up with — in medicine, research output is doubling every 73 days; in almost every other field, the story is the same: more papers, more reports, more data than any person or team can read. People are drowning in information, yet starved for valuable insights they can trust.

Every model builds or breaks trust at three layers: what it knows, how it reasons, and how it's deployed. Right now, the entire industry is obsessing over tuning guardrails and alignment, while the foundation underneath goes almost completely unchecked. The result: AI systems that sound certain while citing Reddit threads as gospel, hallucinating peer review out of thin air, and manufacturing false confidence at exactly the moments people need real answers.

And the stakes are rising fast. We’re no longer just prompting a chatbot, we're deploying agents that search, retrieve, reason, and act in multi-step chains. If the knowledge layer is flawed, an agent doesn't make one mistake; it makes a decision, then builds more decisions on top of it, each one inheriting and compounding whatever it got wrong at the start. And the cost compounds too: now the mistake isn't caught early, it's caught downstream, where it's far more expensive to unwind.

This session breaks down three layers of trust, and makes the case that the one nobody's watching is the one that decides everything else.

This talk flips the responsible AI conversation upside down — from outputs to foundations — showing exactly what breaks when the knowledge layer is rotten, what happens when agentic systems build on it, and what becomes possible when it's solid.

[Source](https://www.youtube.com/watch?v=GlIIs9d_YQU)

中文简报

Wiley在AI系统中嵌入科研数据流以减少错误未经检验的知识基础会导致多步AI智能体出现复合性失效，因而需要经过同行评审的全文数据来确保输出可靠性。 Sue Weng-Cross指出了信息过载持续加剧与AI快速普及之间的矛盾，并指出主流AI工具中90%的引用来自Reddit、维基百科和YouTube等非同行评审来源。她认为，尽管当前行业主要聚焦于模型层和部署层，但真正的信任需要建立在由同行评审的全文研究构建的稳健知识层之上，而非表面化的摘要或网络抓取内容。通过将智能体多步工作流扎根于Wiley经专家策展的知识数据流中，各机构能够显著减少幻觉，避免复合性错误，并发现关键的科学洞见。 Sue Weng-Cross介绍了当代的科研困境：人类每年发表330万篇科学论文，医学科研产出每73天便翻一番。 (0:04) Wiley的一项调查显示，尽管存在严重的信任顾虑，仍有89%的企业研究人员使用AI，且86%的人预计AI将在两年内无处不在。 (0:04) 一项针对15万条AI引用的研究显示，其中90%来自Reddit、维基百科和YouTube，导致模型输出缺乏依据的内容，例如建议在披萨中加入无毒胶水。 (0:04) 企业研究人员（83%）和临床医生（74%）均表示对AI幻觉和数据隐私问题深感担忧。 (4:18) 在多步自主AI智能体工作流中，由于复合性错误的影响，即便每一步的准确率达95%，在经历20步的链条后整体成功率也会降至36%。 (4:18)

Reading summary

Sue Weng-Cross of Wiley argues that enterprise AI failures stem primarily from an unvetted 'Knowledge Layer' rather than algorithmic flaws, warning that grounding models in shallow web data creates compounding operational and safety risks in research-intensive industries. Information Overload and AI Sourcing Deficits (0:04) Modern research output is expanding faster than professionals can absorb: global scientific publishing produces 3.3 million articles annually, and medical knowledge doubles every 73 days, requiring primary care physicians to spend an impossible 27 hours daily to stay current. This bottleneck has spurred rapid adoption, with a Wiley survey of 1,000 researchers indicating that 89% of corporate researchers currently utilize AI tools and 86% expect ubiquitous adoption within two years. However, high adoption rates mask a fundamental sourcing deficit. A June 2025 analysis of 150,000 citations across ChatGPT, Perplexity, Gemini, and Google AI found that 90% of citations originated from non-peer-reviewed platforms: Reddit accounted for over 40%, followed by Wikipedia at 26% and YouTube at 24%. While suitable for casual queries, models relying on these platforms fail to separate humor from empirical evidence, producing dangerous recommendations when applied to clinical guidelines or drug interactions. Economic Costs and Agentic Compounding Errors (4:18) User anxiety remains high despite pervasive usage. Wiley's findings show that 83% of corporate researchers and 74% of clinicians identify hallucinations and inaccuracies as their primary barrier to deeper AI integration. Flawed data quality imposes an estimated average cost of $13 million annually per organization, while promised efficiency gains evaporate when professionals must manually audit outputs—such as in the legal case Mata v. Avianca, where attorneys submitted AI-hallucinated judicial precedents. These vulnerabilities scale exponentially within autonomous, multi-step systems. In clinical practice, physicians seeing 10 to 20 patients daily lack the time to verify AI claims across secondary databases. Furthermore, Weng-Cross points out that an autonomous agent operating at a 95% accuracy rate at each individual decision step suffers a compounding degradation, resulting in a 36% overall success rate—and a 64% failure rate—across a 20-step chain. Rebalancing Trust Across AI Architectural Layers (8:33) Weng-Cross segments enterprise AI systems into three distinct operational tiers: the Deployment Layer (governing compliance, privacy, and audit trails), the Model Layer (managing architecture, fine-tuning, and guardrails), and the foundational Knowledge Layer. Current industry governance focuses disproportionately on the top two layers while neglecting the third. Weng-Cross asserts that guardrails and fine-tuning cannot correct for absent or corrupt baseline facts. Evaluating the Knowledge Layer requires addressing three dimensions: breadth (indexing specialized primary literature over web scrapes), depth (accessing full-text methodologies, raw result tables, and protocols rather than abstracts alone), and machine-readable trust signals that communicate peer-review status, citation metrics, and post-publication corrections. Measurable Benefits of Full-Text Grounding (13:54) To demonstrate the practical value of complete literature access, Wiley partnered with biomedical firm BioRelate to benchmark full-text research articles against publicly available PubMed Central abstracts. Grounding models in full-text literature uncovered 63% more novel biological relationships per article, yielding 160,000 unique relationships that were entirely undetectable in abstract-only indexing. This indicates that prior analytical shortcomings stemmed from data visibility constraints rather than model reasoning capacity. In an enterprise application with Microsoft Discovery, Wiley deployed a Life Sciences Research Agent connected via Model Context Protocol standards. Instead of acting as an isolated search interface, the agent functions within an orchestrated multi-agent workflow, generating hypotheses from vetted pharmacology data and passing structured findings directly to downstream agents responsible for lab protocols and experimental simulations. Weng-Cross concludes that enterprise differentiation depends on grounding AI systems in verified primary literature at the source, noting that autonomous agents built to resolve information overload will otherwise merely accelerate the spread of systemic errors.

精炼阅读摘要

Wiley 战略合作总监 Sue Weng-Cross 指出，面对全球科研文献与临床指南爆炸式增长带来的信息过载，企业加速引入 AI，但主流模型底层高度依赖未经同行评审的公开网络内容；要消除幻觉并建立可信系统，关键在于从源头治理知识层，以经过同行验证的高质量全文数据作为推理基石。信息过载与底层语料不可靠的双重矛盾 (0:04) 人类目前每年公开发表 330 万篇科研论文，其中医学领域的知识总量每 73 天便翻倍一次，初级保健医生若要跟进所有临床指南和最新证据，每天需耗费 27 个小时进行阅读。这种信息瓶颈促使研究密集型行业迅速拥抱 AI。Wiley 对 1,000 名企业和学术研究人员的调查显示，89% 的企业研究人员已在工作流中使用 AI 工具，86% 的受访者预计 AI 将在未来两年内实现全面普及。然而，工具的广泛应用并不意味着输出结果建立在可靠基础之上。2025 年 6 月一项针对 ChatGPT、Perplexity、Gemini 和 Google AI 涉及 15 万次引用的分析表明，Reddit 占所有引用来源的 40% 以上，维基百科占 26%，YouTube 占 24%，三者合计构成了 90% 的引用基础。Weng-Cross 强调，这些缺乏同行评审的网络平台原本就并非为严谨科学研究所设计，当模型无法区分幽默调侃与严肃事实时，就会出现将 Reddit 搞笑评论作为建议（如在披萨酱中掺入无毒胶水防滑）等荒唐错误；在药物配伍或临床诊疗等场景中，缺乏事实依据的生成结果将直接带来严重危害。幻觉风险、连锁错误与低质数据的经济代价 (4:18) 高采用率的背后伴随着强烈的信任危机。Wiley 的调查显示，83% 的企业研究人员高度担忧模型幻觉、不准确性以及数据隐私问题；81% 的临床医生虽已使用 AI，但其中 74% 将模型幻觉列为最主要顾虑。实际业务流程中，医生每天接诊 10 至 20 名患者，根本无法在主系统外逐一跨库手动核查 AI 给出的结论，导致自动化验证在现实操作中难以实现。劣质数据引发的财务、运营和声誉损失极为显著。企业因数据质量缺陷平均每年蒙受高达 1,300 万美元的损失，原本追求的效率提升也被人工排错和返工所抵消。例如在 Mata v. Avianca 诉讼案中，律师提交了包含 AI 捏造虚假判例的诉状，模型甚至混淆了华盛顿特区与华盛顿州。而在多步骤自主智能体链条中，误差还会呈现指数级复合放大：即使单步决策准确率高达 95%，经过 20 步的推理传递后，整体任务成功率将骤降至 36%，失误率高达 64%。信任三层架构与被忽视的基础知识层 (8:33) 为拆解可信度缺陷，Weng-Cross 提出了构建 AI 信任的三个关键层次：负责合规、权限与审计的部署层（Deployment Layer）；负责基础架构、微调与对齐护栏的模型层（Model Layer）；以及处于底座位置的知识层（Knowledge Layer）。目前业界治理主要偏重于前两层，却忽略了知识层才是决定模型推理上限的基础。如果缺乏准确输入，无论后续护栏如何严密，模型都无法对从未接收到的事实展开正确推理。评估知识层的质量需涵盖三个核心维度：广度（超越免费网络抓取，纳入经过同行评审的专有数据库与一手信源）、深度（穿透浅层摘要，完整收录包含实验方案、原始数据表格、研究局限性与因果链条的全文内容）以及信任信号（提供机器可读的同行评审标识、引用脉络、版本迭代及勘误信息）。Wiley 依托覆盖 2,000 余种期刊、3 万部图书和超 500 万篇论文的数据库，通过 API 和模型上下文协议（MCP）向生命科学、医疗、化学和金融等领域的自主智能体输送结构化全文知识流，目前全球排名前 10 的制药公司中有 7 家已将其接入研发管线。全文数据赋能与多智能体研发协作验证 (13:54) 产业合作数据证实了以同行评审全文知识替代公开发表摘要对模型性能的实质提升。在与生物医药公司 BioRelate 开展的药物发现因果图谱概念验证中，Wiley 提供了小规模生命科学全文文献样本，与 PubMed Central 的公开摘要检索进行对比。测试结果显示，全文语料使单篇文献提取出的新颖生物学因果关系数量提升了 63%，并成功挖掘出纯摘要检索完全无法获取的 16 万个独特生物学关联，证实了过往知识挖掘的瓶颈本质上在于底层数据的缺失而非模型算法能力的局限。此外，Wiley 与 Microsoft Discovery 合作开发了专门的生命科学研究 Agent。该系统并非单一的文献检索框，而是深度嵌入多智能体编排网络中：科研人员提示 Wiley Agent 检索并合成药理学数据库以提出假说，随后可直接将研究成果流转给协同 Agent，用以自动生成实验协议、处理干实验模拟并调度湿实验执行，展现了高可靠性数据对端到端科研自动化的支撑能力。 Weng-Cross 总结指出，若 AI 智能体构建在不可靠的数据地基之上，消除信息过载的工具只会成倍加速错误的产生与扩散；真正实现企业级竞争壁垒的前提，是在模型运行之前从源头接入经权威验证的专业知识。然而，如何在保障知识产权与合规的前提下降低全文接入成本，并在更多复杂跨学科领域实现知识流的动态同步，仍是企业深度部署 AI 面临的持续挑战。

Detailed notes — zh

- [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) 科研信息面临严峻超载：全球年发表论文超 330 万篇，医学产出每 73 天翻倍，基层医生完全跟进指南每天需 27 小时。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) 主流 AI 模型的检索来源极度依赖公众平台：Reddit、Wikipedia 和 YouTube 垄断了高达 90% 的引用量，其中 Reddit 占比超 40%。 - [03:26](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=206s) 企业与医疗科研界普遍存在 AI 信任鸿沟：超八成企业科研人员及 74% 的医生对模型幻觉和准确性心存高度顾虑。 - [03:26](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=206s) 数据质量低下导致企业每年在 AI 系统上蒙受约 1300 万美元损失，基础大模型的商品化使得受信任的知识层成为唯一护城河。 - [06:55](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=415s) 在现实高强度工作流中，“人机协同”（Human-in-the-loop）核验机制一旦脱离当前系统，在实践中往往彻底失效。 - [06:55](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=415s) 自主智能体的多步骤推理会放大错误：即使单步准确率达 95%，经过 20 步级联后，系统的综合失败率仍将飙升至 64%。 - [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) AI 信任架构包含部署层、模型层和知识层，其中最被业界忽视的底层“知识层”决定了推理边界与模型能力的上限。 - [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) 评估知识层质量需考量三大核心维度：广度（权威来源覆盖）、深度（全文本 vs 仅依赖摘要）以及信任信号（同行评议与元数据）。 - [11:45](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=705s) Wiley 推出的知识流（Knowledge Feeds）通过 API 与 MCP 协议提供经清洗的权威全文本语料，目前已被全球前十大药企中的 7 家采用。 - [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) Biorate 概念验证显示，基于全文本的文献挖掘比公开摘要使每篇论文发现的新颖生物因果关系提升 63%，额外发掘 16 万个独特关联。 - [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) Wiley 与微软合作构建的科研智能体支持跨系统自主编排，能够打通从科学假设生成到干湿实验设计的完整推理闭环。 - [16:34](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=994s) 构建智能体若脱离可信知识底座，只会以更快的速度极其自信地输出错误答案，知识基座的质量决定了整个 AI 系统的成败。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) 信息过载危机与学术界的 AI 采纳悖论 Wiley 战略 AI 合作总监 Sue Weng-Cross 开篇介绍了自己的团队职责：与 AI 开发者、技术平台及行业伙伴建立深度合作，将权威可信的科学研究成果注入到科研人员与临床医生使用的 AI 平台中，以加速科学发现的步伐。当前人类正面临极其严峻的知识爆炸难题——全球科研界每年发表超过 330 万篇学术论文；在医学领域，科研产出更是以每 73 天翻一番的速度激增；初级保健医生如果想完全跟上最新的临床指南与循证医学证据，每天需要耗费 27 个小时，这在生理上根本无法实现。我们正在被海量的信息所淹没，却极度渴求真正值得信赖的高价值洞察（"drowning in information, but starved for insight"）。正是这种信息超载促使各行各业加速拥抱 AI。Wiley 针对 1000 名学术界与企业科研人员开展的调研显示，企业界已有 89% 的研究人员在日常工作中使用 AI，且 86% 的受访者认为未来两年内 AI 将被全面普及。然而，极其广泛的采纳并不代表 AI 系统在摄取最优质、最可信的知识源。2025 年 6 月针对 ChatGPT、Perplexity、Gemini 及 Google AI 的 150,000 条引用分析揭示了一个惊人的现状：Reddit 竟然是引用率最高的单一来源，占比超过 40%；排在其后的是维基百科（26%）与 YouTube（24%）。这意味着高达 90% 的 AI 检索引用均来自这三大公众平台。 - 全球每年发表科研论文达 330 万篇；医学领域研究产出每 73 天翻倍。 - 基层医生若要完全跟进最新指南，每天需花费 27 小时。 - 企业科研人员中 AI 使用率已达 89%，86% 预计两年内实现广泛应用。 - 头部 AI 模型检索引用来源分布：Reddit 占 40% 以上，Wikipedia 占 26%，YouTube 占 24%，三者合计垄断了 90% 的引用量。 ### [03:26](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=206s) 信任鸿沟与生成错误的连锁代价 Reddit 或许能在日常问答中提供参考，但它绝非为了支撑严肃的科学实验与临床医疗而设计的严密证据库。当 AI 无法分辨严肃事实与网络恶搞时，荒谬的灾难便会发生：Google AI 曾因采信 Reddit 恶搞帖子，堂而皇之地建议用户在自制披萨底上涂抹无毒胶水以防芝士滑落。对于披萨爱好者而言这或许只是个笑料，但如果发生在患者护理、药物相互作用或前沿科学实验中，后果不堪设想。这种对输出可靠性的担忧直接导致了巨大的“信任鸿沟”：Wiley 的调研指出，83% 的企业科研人员对现有 AI 模型心存顾虑，焦点集中于幻觉（Hallucination）、模型不准确性以及数据隐私；而在医学领域，81% 的医生虽使用 AI，却有 74% 将模型幻觉列为最主要的担忧。 AI 犯错所带来的代价是极其具体且昂贵的。首先是金钱损失：企业每年因 AI 系统数据质量低下而造成的平均损失高达约 1300 万美元。其次是时间沉没成本：AI 的核心承诺是提效降本，但底层基石薄弱导致的大量返工、人工核对和修正，反而抵消了其带来的效率收益。第三是声誉受损，典型的如 Mata 诉 Avianca 航空案中，律师借助 AI 撰写法律文书，结果文书中凭空捏造了三起虚假判例，甚至连华盛顿哥伦比亚特区（Washington, D.C.）和华盛顿州（Washington State）都混淆不清——作为澳大利亚人的演讲者调侃道，连自己都能分清这两者的差别。 - 83% 的企业科研人员担忧现有 AI 模型的幻觉、不准确性及隐私问题。 - 81% 的临床医生采纳 AI，但 74% 极度担忧其幻觉倾向。 - 数据质量低下导致企业每年在 AI 系统上蒙受约 1300 万美元损失。 - 基础大模型正在高度同质化与商品化，单纯依靠模型选型无法为企业构筑护城河，差异化的核心在于受信任的知识层。 ### [06:55](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=415s) 检索缺陷的级联放大与智能体复合错误在现实落地中，检索阶段引入的错误几乎不可能在下游被完全挽回。行业普遍推崇在流程中引入“人机协同”（Human-in-the-loop）核验机制，但这完全取决于人类是否真有精力执行核查。一名每天需要穿梭巡诊 10 到 20 位病患的临床医生，根本不可能为了验证 AI 给出的诊断建议而去单独打开第二个系统查验原文。一旦验证机制要求用户脱离现有的工作流，这种核验在实践层面就等同于完全失效。更关键的是，随着技术由“问答式聊天机器人”向能够自主规划并采取行动的“自主智能体（Autonomous Agents）”演进，试错的赌注正在呈指数级攀升。智能体检索到的错误信息不仅不会止步于单次回答，反而会作为输入传递到第二步、第三步的推理链中，造成每一步决策都在前一步错误基石上层层累加。数学概率在此处展现了残酷的现实：即便一个模型在单步推理中的准确率高达 95%，但在一个长达 20 步的多步骤自动化工作流中，累积到第 20 步时的系统综合失败率将飙升至 64%。 ``` [ 单步准确率: 95% ] │ ├── 步骤 1 (准确率 95.0%) ──> 错误输入下游 ├── 步骤 2 (准确率 90.3%) ──> 错误逐步累积 ├── 步骤 3 (准确率 85.7%) ──> 推理偏差扩大 │... └── 步骤 20 (累积失败率高达 64.2%!) ``` ### [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) AI 信任的三层架构与知识层的核心维度要从根源上解决 AI 的可靠性危机，必须审视支撑 AI 系统的“信任三层架构”。处于最顶层的是**部署层（Deployment Layer）**，它决定系统如何在现实世界中合规运行，涵盖治理架构、审计跟踪、问责制、隐私保护以及法规合规，目前受到业界最为密集的监管与审视；中间是**模型层（Model Layer）**，关注模型本身的能力指标、行为特征、幻觉率以及检索准确度，同样备受算法工程师瞩目；而垫在最底端、决定一切的则是**知识层（Knowledge Layer）**。知识层决定了模型在运行推理和安全护栏触发之前究竟能够获取什么信息——再严密的部署护栏也无法弥补系统关键数据的缺失，模型绝不可能对其从未接触过的知识进行逻辑推理。 ``` +---------------------------------------------------------+ | 部署层 (Deployment) | | 合规审计、访问治理、问责机制、安全护栏 (关注度极高) | +---------------------------------------------------------+ │ +---------------------------------------------------------+ | 模型层 (Model) | | 推理性能、幻觉率、检索能力、对齐微调 (关注度高) | +---------------------------------------------------------+ │ +=========================================================+ | 知识层 (Knowledge) | | 决定检索边界与推理上限的基础 (关注度最低，却最致命) | +=========================================================+ ``` 评估底层知识层的质量取决于三大维度： 1. **广度（Breadth）**：AI 检索的知识视野。系统究竟是在摄取最多元、最具权威性且持续更新的全文期刊与专业数据库，还是仅仅抓取 Reddit 等公开免费的网络数据。 2. **深度（Depth）**：专业领域知识的穿透力。至今仍有大量科研 AI 工具仅依赖论文摘要（Abstracts）进行检索，这仅仅是因为摘要免费且容易获取。但摘要本质只是高度概括，遗漏了论文正文中的完整实验数据、试剂方案、操作协议、研究局限性以及深层因果关系。 3. **信任信号（Trust Signals）**：包括规范引用、源头归属、同行评议状态（Peer-reviewed Status）、版本迭代与勘误记录。这些元数据向 AI 系统和最终用户明确传递了知识的时效性、合法性与学术置信度。 ### [11:45](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=705s) Wiley 的知识基建：从全文本到知识流（Knowledge Feeds）作为全球排名前三的科学出版集团以及全球最大的学术学会出版商（合作学会超 900 家），Wiley 拥有 2,000 余种学术期刊、超 500 万篇科研论文全文以及近 30,000 部学术专著，触达全球 50 多个国家的 5,000 万科研人员，也是最早服务 AI 与数据市场的科学出版商。针对 AI 所急需的广度、深度与信任信号，Wiley 推出了面向生命科学、医疗健康、化学工程以及金融商业等高研发密集型行业的“知识流（Knowledge Feeds）”。这些知识流由 Wiley 内部各学科的领域专家（SMEs）精心策划清洗，均保留论文完整正文（Full Text），并通过 API 以及 MCP（Model Context Protocol）协议直接交付，附带完整的可溯源引用，能够无缝插入到机构现有的任意智能体框架与工作流中。目前全球排名前 10 的制药巨头中有 7 家正在使用 Wiley 的知识流支撑其核心药物发现任务，且越来越多药企不再局限于购买数据插件，而是主动寻求与 Wiley 进行深度的联合开发（Co-build）。 - Wiley 拥有超过 2,000 本学术期刊、500 万篇全文文献及近 30,000 种专著。 - 覆盖生命科学、医疗、化工工程、金融商业等研发密集型行业。 - 知识流支持以 API 及 MCP 协议向各类智能体注入带完整引用的全文本语料。 - 客户已涵盖全球前十大药企中的 7 家。 ### [13:54](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=834s) 落地验证：Biorate 因果发现与微软 Discovery 智能体编排为了证明高质量知识对 AI 表现的决定性作用，演讲者分享了两个具体案例。首先是与生物医学公司 Biorate 的概念验证（PoC）合作。Biorate 擅长利用 AI 挖掘药物研发中的因果生物学关系。Wiley 仅为其提供了一小批生命科学领域的“全文”文献样本，结果却极其震撼：基于全文本的文献挖掘相比于从 PubMed Central 抓取的公开摘要，每篇文章识别出的新颖生物学因果关系数量提升了 63%；仅在这批微小的全文样本中，系统就新发掘出 160,000 个在摘要模式下完全隐匿的独特生物学关系。这充分证明，之前洞察的缺失根本不是因为 AI 模型不够强大，而是因为模型自始至终无法接触到底层的完整实验数据。第二个案例是 Wiley 与微软科学探索平台（Microsoft Discovery）的合作。双方联合打造了“Wiley 生命科学研究智能体”，该智能体不仅能对 Wiley 庞大的生命科学与药理学语料库进行检索、召回与长文本合成，更具备与其他专用智能体进行跨系统编排的能力。研发人员在进行药物发现任务时，可借助该智能体加速科学假设生成，随后直接在统一环境中联动其他智能体，顺畅衔接实验方案设计、结果解读以及干湿实验（Dry Lab / Wet Lab）推演，形成连续的闭环科研推理链。 - Biorate PoC 结果：全文挖掘使每篇论文发现的新颖生物学因果关系比公开摘要提升 63%。 - 样本集内额外发现 160,000 个未曾被摘要收录的独特生物关联。 - 联合微软构建的科研智能体支持跨系统自主编排，打通从假设生成到干湿实验设计的端到端科学推理闭环。 ### [16:34](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=994s) 结论：审视智能体的立足之基在演讲的结尾，Sue Weng-Cross 强调了三项核心启示： 1. 人类构建 AI 智能体的初衷是解决海量信息的超载问题，但一个依赖垃圾信息运行的系统根本不能解决问题，它只会“以更快的速度、极其自信地给出错误答案”（"confidently wrong at a faster rate"）。 2. 在由部署层、模型层和知识层构成的信任三角中，最被业界忽视的底层知识层，恰恰决定了上面两层能否真正发挥作用。 3. 知识层的失效将沿着工作流造成系统级雪崩，而随着自主智能体的深入普及，这一层面的风险与代价只会成倍放大。她呼吁所有正在采购、构建与部署 AI 系统的组织与开发者跳出对教科书老牌出版商的固有印象，将 Wiley 视为企业级 AI 赖以生存的可信知识基座。无论是在设计提示词还是部署多智能体网络，每个人都应向自己发问：“你的系统究竟立足于何种知识基石之上？”因为对这个问题的回答，直接决定了整个 AI 系统能否真正被赋予信任。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) drowning in information, but starved for valuable insights they can trust — Sue Weng-Cross > Every model builds or breaks trust at three layers: what it knows, how it reasons, and how it's deployed. — Sue Weng-Cross > If the knowledge layer is flawed, an agent doesn't make one mistake; it makes a decision, then builds more decisions on top of it, each one inheriting and compounding whatever it got wrong at the start. — Sue Weng-Cross > [00:04](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=4s) 初级保健医生如果想完全跟上最新的临床指南与循证医学证据，每天需要耗费 27 个小时，这在生理上根本无法实现。 — Sue Weng-Cross > [03:26](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=206s) 基础大模型正在高度同质化与商品化，单纯依靠模型选型无法为企业构筑护城河，差异化的核心在于受信任的知识层。 — Sue Weng-Cross > [06:55](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=415s) 一旦验证机制要求用户脱离现有的工作流，这种核验在实践层面就等同于完全失效。 — Sue Weng-Cross > [06:55](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=415s) 即便一个模型在单步推理中的准确率高达 95%，但在一个长达 20 步的多步骤自动化工作流中，累积到第 20 步时的系统综合失败率将飙升至 64%。 — Sue Weng-Cross > [08:33](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=513s) 再严密的部署护栏也无法弥补系统关键数据的缺失，模型绝不可能对其从未接触过的知识进行逻辑推理。 — Sue Weng-Cross > [16:34](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=994s) 一个依赖垃圾信息运行的系统根本不能解决问题，它只会‘以更快的速度、极其自信地给出错误答案’（confidently wrong at a faster rate）。 — Sue Weng-Cross > [16:34](https://www.youtube.com/watch?v=GlIIs9d_YQU&t=994s) 无论是在设计提示词还是部署多智能体网络，每个人都应向自己发问：‘你的系统究竟立足于何种知识基石之上？’ — Sue Weng-Cross ## Mentioned Sue Weng-Cross, Wiley, Ai4, ChatGPT, Perplexity, Gemini, Google AI, Reddit, Wikipedia, YouTube, Mata v. Avianca, Avianca, MCP (Model Context Protocol), Biorate, PubMed Central, Microsoft Discovery, Microsoft ## Video description Sue Weng-Cross - Strategic AI Partnerships, Wiley Knowledge is compounding faster than anyone can keep up with — in medicine, research output is doubling every 73 days; in almost every other field, the story is the same: more papers, more reports, more data than any person or team can read. People are drowning in information, yet starved for valuable insights they can trust. Every model builds or breaks trust at three layers: what it knows, how it reasons, and how it's deployed. Right now, the entire industry is obsessing over tuning guardrails and alignment, while the foundation underneath goes almost completely unchecked. The result: AI systems that sound certain while citing Reddit threads as gospel, hallucinating peer review out of thin air, and manufacturing false confidence at exactly the moments people need real answers. And the stakes are rising fast. We’re no longer just prompting a chatbot, we're deploying agents that search, retrieve, reason, and act in multi-step chains. If the knowledge layer is flawed, an agent doesn't make one mistake; it makes a decision, then builds more decisions on top of it, each one inheriting and compounding whatever it got wrong at the start. And the cost compounds too: now the mistake isn't caught early, it's caught downstream, where it's far more expensive to unwind. This session breaks down three layers of trust, and makes the case that the one nobody's watching is the one that decides everything else. This talk flips the responsible AI conversation upside down — from outputs to foundations — showing exactly what breaks when the knowledge layer is rotten, what happens when agentic systems build on it, and what becomes possible when it's solid. [Source](https://www.youtube.com/watch?v=GlIIs9d_YQU)

Source quotes

[0:04] Sue Weng-Cross: drowning in information, but starved for valuable insights they can trust Sue Weng-Cross: Every model builds or breaks trust at three layers: what it knows, how it reasons, and how it's deployed. Sue Weng-Cross: If the knowledge layer is flawed, an agent doesn't make one mistake; it makes a decision, then builds more decisions on top of it, each one inheriting and compounding whatever it got wrong at the start. [0:04] Sue Weng-Cross: 初级保健医生如果想完全跟上最新的临床指南与循证医学证据，每天需要耗费 27 个小时，这在生理上根本无法实现。 [3:26] Sue Weng-Cross: 基础大模型正在高度同质化与商品化，单纯依靠模型选型无法为企业构筑护城河，差异化的核心在于受信任的知识层。 [6:55] Sue Weng-Cross: 一旦验证机制要求用户脱离现有的工作流，这种核验在实践层面就等同于完全失效。 [6:55] Sue Weng-Cross: 即便一个模型在单步推理中的准确率高达 95%，但在一个长达 20 步的多步骤自动化工作流中，累积到第 20 步时的系统综合失败率将飙升至 64%。 [8:33] Sue Weng-Cross: 再严密的部署护栏也无法弥补系统关键数据的缺失，模型绝不可能对其从未接触过的知识进行逻辑推理。 [16:34] Sue Weng-Cross: 一个依赖垃圾信息运行的系统根本不能解决问题，它只会‘以更快的速度、极其自信地给出错误答案’（confidently wrong at a faster rate）。 [16:34] Sue Weng-Cross: 无论是在设计提示词还是部署多智能体网络，每个人都应向自己发问：‘你的系统究竟立足于何种知识基石之上？’

Source quotes

[0:04] Sue Weng-Cross: In medicine specifically, total research output doubles every 73 days, creating an overwhelming workload where a primary care physician would need 27 hours a day just to read and stay up to date with new clinical guidelines and evidence. [0:04] Sue Weng-Cross: Across every sector, individuals and teams are drowning in raw information while remaining starved for actionable insights. [0:04] Sue Weng-Cross: Collectively, 90% of AI citations originated from these three non-peer-reviewed platforms. [0:04] Sue Weng-Cross: When AI models fail to differentiate between joke posts and valid evidence, bizarre errors occur—such as Google AI recommending users add non-toxic glue to pizza sauce so the cheese won't slide off, a tip scraped directly from a humorous Reddit comment. [4:18] Sue Weng-Cross: Organizations want AI, but hesitation surrounding untrustworthy outputs throttles full operational deployment. [4:18] Sue Weng-Cross: Operationally, the promised speed efficiencies of AI are wiped out by human rework, verification, and error correction. [4:18] Sue Weng-Cross: A multi-step model that is 95% accurate at each individual decision step drops to a 36% success rate (64% failure rate) across a 20-step chain. [8:33] Sue Weng-Cross: The industry currently focuses most of its governance on the top two layers while ignoring the foundation. [8:33] Sue Weng-Cross: No amount of deployment oversight or model tuning can make up for missing or corrupt baseline data; a model cannot reason over facts it never received. [13:54] Sue Weng-Cross: The full-text corpus revealed 63% more novel biological relationships per article than abstract-only indexing. [13:54] Sue Weng-Cross: This proved that earlier reasoning limitations stemmed from missing access to primary data rather than model capability gaps. [13:54] Sue Weng-Cross: AI agents designed to eliminate information overload will only compound errors faster if grounded in poor data.

Original transcript

[0:04] Hi everyone, thank you for being here today. Um, as Austin said, my name is Suen Cross and [0:10] I am the director of strategic partnership at WY. And so what this means is that I lead a team that [0:17] builds deep partnerships with AI developers, um, technology platforms and industry partners [0:26] to bring trusted scientific research into AI solutions and platforms that researchers and [0:33] clinicians use um, and and all with the goal of accelerating the pace of scientific research. [0:42] And so today I want to have a conversation with you about trust um and the three layers in which [0:50] trust in an AI system is built and make the case that the layer that that's receiving the least [0:58] attention which is the knowledge layer is the one that determines everything else that follows. And [1:06] so let's start with the problem that um AI is actually being asked to solve here. So every [1:14] year we publish 3.3 million research articles. We as in all the researchers in the world and so [1:22] um in medicine alone research output doubles every 73 days. And also in um medicine, [1:32] primary care physicians needs 27 hours a day, that's not possible to learn and stay up to [1:40] date with the latest clinical guidelines and um and evidence. And so these statistics capture the [1:50] scale of the problem that we face which is that we are publishing more research, more reports [1:56] and more data than any individual or teams could possibly read, let alone synthesize and apply. And [2:04] so we are drowning in information but starved for insight. And so this is the reason why we as an [2:13] industry are accelerating towards AI adoption. And so in a study that was done by WY where [2:19] we surveyed a thousand academic and corporate researchers, we found that 89% of researchers [2:27] in the corporate sector already use AI and in fact 86% of researchers believe that AI adoption will [2:36] be widespread in the next two years. But here is what's interesting. Widespread adoption does not [2:43] mean that AI and AI systems are drawing from the best and most trustworthy resources. And [2:50] so here is a analysis that was done in June 2025 of 150,000 citations that was um across Chat GPT, [3:01] Perplexity, Gemini, and Google AI. And what it found was that Reddit was by far the most [3:08] trusted the the most um cited source with over 40% of citation. This was followed by Wikipedia [3:16] at 26% and YouTube at 24%. So collectively 90% of citations came from these three sources. [3:26] Now, Reddit can be an extremely valuable resource, but it was never designed to be a scientific and [3:33] clinical evidence-based to support scientific experimentation and clinical care. And we've [3:40] all seen what happens when AI cannot distinguish between what is trusted and reliable source versus [3:46] something that's a joke. So, Google AI famously told its user to put non-toxic glue um in it pizza [3:55] base to stop the cheese from sliding off. And it's a it's an it's an output that was tied to [4:02] a Reddit post. Now, that example might be funny, it might not if you're a pizza lover, but it is [4:09] definitely not funny when it's related to patient care or drug interaction or scientific experiment. [4:18] And so this then brings us to the trust gap, right? And so in the same explanation survey [4:24] that WY conducted, we found that 83% of corporate researchers express concerns about the AI models [4:31] that they use, specifically hallucination, u model inaccuracies and data privacies. [4:39] Also in medicine, 81% of clinicians use AI, but 74% sites hallucination as a leading concern. [4:49] And so when these same corporate researchers and clinicians were asked what would drive greater [4:55] adoption and usage, they consistently say the same things, which is validated safety and greater data [5:03] um assurance. And so the meaningful takeaway here is while we all agree that we need AI, [5:11] concerns around the reliability of its output is preventing organizations and [5:16] users from fully trusting these systems and ultimately limiting greater adoption and use. [5:28] And also there are some real cost and consequences when AI system gets it wrong. [5:33] So first is money. On average, an organization can lose about $13 million a year because of [5:40] poor data quality in its AI systems. Second is time, right? So the promise of AI is [5:49] um speed and efficiency, but a weak foundation can cancel out the gains through things like rework [5:58] um validation and corrections. And third is reputational cost. So in this example with [6:06] Marta versus Aviank um the lawyers prepared a court filing with the help of AI but that [6:14] court filing contained three fabricated cases and it even confu may it even got confused between [6:21] Washington DC and Washington state and I'm Australian and I even know the difference. So, [6:31] and so let's also not forget that powerful AI models are becoming increasingly commoditized, [6:38] non-differiated and accessible. And so what is going to give your organization [6:44] um that competitive edge will no longer come from the model you choose to deploy [6:49] and it's the trusted knowledge layer that is going to create the differentiated outcome. [6:55] The other point that I want to make here is what we see in reality is that errors that [6:59] enter at the point of retrieval is difficult to be corrected downstream. Now we build we have [7:07] verifications and human in the loop right that is supposed to provide a level of assurance but that [7:13] only provides that level of assurance if this the human actually does it. And so a physician [7:21] that is running between 10 20 patients a day is unlikely to open a second system to verify if the [7:29] recommendation that's being made is accurate. So verification that requires a user to leave [7:36] its existing workflow is a verification that in practice doesn't happen. And the last point [7:44] that I want to make about the problems that we all face is that the stakes are getting higher, right? [7:50] Because we are moving from agents or AI systems that answers questions to autonomous agents that [7:58] acts on our behalf. And so in an autonomous agent, a bad retrieval doesn't just stop with one answer. [8:06] it becomes the input into step two and step three with every decision building from what came before [8:12] it. And so this is what's interesting. A model that is right 95% of the time at every step still [8:21] has a chance of failing still has a chance of failing 64% by step 20 across a 20step process. [8:33] And so how do we fix all this? Well, by taking a deep look at um the three layers of trust that [8:41] AI systems depends on. So at the top you have the deployment layer. This is the layer that [8:48] determines how an AI system can operate in the real world and it includes looking at things [8:54] like governance, audit trail, accountability, privacy and and compliance with regulations. [9:01] This layer is heavily watched. Then beneath that you've got the model layer. And this is [9:07] the layer that looks at things like how does a model perform? What is its capability? How [9:13] does it behave? Looks at things like hallucination rate, retrieval of reliability. This layer is also [9:20] heavily looked at. But then you have the knowledge layer which sits at the bottom. And the knowledge [9:25] layer is a layer which determines what the model can actually retrieve and reason and before you [9:32] know the deployment layer even um you know before the guardrails even fires off right and so a sound [9:42] deployment layer doesn't compensate for the fact that a model that that a system is missing [9:51] uh certain data and a model can't reason over information that it never had in the first place. [9:58] And so while all three layers are critical, the f the knowledge layer is the foundation and it's the [10:04] layer that we at WY build. And so you may ask the question of well what determines knowledge quality [10:12] in that knowledge layer? Well, it comes down to three things. Breath, depth, and trust signals. [10:18] So breath is where the AI system can look. Is it drawing from the most diverse authoritative [10:26] current sources, right, like full text research, uh, specialized databases, or is it just simply [10:33] pulling information that is readily available online and free like Reddit? Depth on the other [10:40] hand looks at how far deep a discipline or topic area the knowledge can go. And what I find really [10:47] interesting is in this day and age, a lot of the AI systems that research researchers use are still [10:55] pulling information from abstract, right? And so abstracts are are being used because they are free [11:02] and they're readily accessible, but they are just a summary and they're they don't intend to incre [11:09] include um critical informations like um full experiment results, protocols and methodologies, [11:18] research limitations and causal relationships that's contained in the body of the research. [11:24] And then the third dimension is trust signals. So they include things like citations, attributions, [11:30] peer review status, versioning and um corrections. And so these send signals to the AI system and the [11:38] user whether where that information come from and whether that information is reliable. Um [11:45] it it's valid and it is current. And so this is where WY comes in. So for those of you who [11:52] are not familiar with WY, we are the top three scientific publisher in the world. We are also [11:58] the largest society publisher in the world and we have over 900 society um publishers. We have [12:07] um a very unmatched uh and unique breath of content including over 5 million research [12:14] articles across 2,000 journals. We also published just under 30,000 books. We reach over 50 million [12:23] researchers um over 50 countries and we are the first scientific publisher. We were the first [12:29] scientific publisher to work and to serve the AR and data market. And so how do we leverage our [12:37] scale expertise and our um trusted knowledge to address the breath, depth and trust signals that [12:46] AI systems so desperately need? well through our industry specific knowledge feeds. And so these [12:53] knowledge feeds serve some of the most research intensive verticals including life sciences, [13:00] healthcare, chemistry and engineering and finance and um business. And so these knowledge feeds are [13:08] curated by our in-house subject matter experts. They're in full text and um they uh are also [13:16] delivered through API and MCP with full citations into any agents, tools and workflows that your [13:25] organization already use and we're already seeing value in these um knowledge feeds. [13:31] We serve seven out of the top 10 pharmaceutical companies for some of their most critical drug [13:38] discovery use cases. And increasingly these pharmaceutical companies are not just buying [13:44] our knowledge feeds and plugging into their solutions. They're asking us to build with them. [13:54] And lastly, I want to go through a couple of examples of that demonstrates the benefit of [14:01] trusted knowledge in AI systems. So in this particular example, WY partnered with a company [14:07] called Biorate in a proof of concept. So Biorate is a biomed company um that has a capability to [14:16] identify causal relationships in drug discovery use cases. Um and in this particular proof of [14:22] concept partnership, we gave bio relate a very small sample of of our full text research um in [14:31] life sciences and what we found was significant. We found that in full text re in full text [14:38] research we identified 63% more novel biological relationships per article compared to publicly [14:49] available abstract in PubMed central. In fact, across that entire um very small sample that we [14:56] gave bio relate, we identified 160,000 new unique biological relationships that would have otherwise [15:05] not been identified if the system was only only retrieving content from abstracts in PubNet [15:12] Central. And so what this example demonstrate demonstrated is that the insight was missing not [15:19] because um the model was not powerful enough. The insight was missing because the model never had [15:25] access to the research where this relationship existed. And the last example that I want to [15:32] take is an an example of Wley's partnership with Microsoft discovery. And so in this partnership [15:39] we built the WY life sci w we research agent in life sciences and um this agent has the capability [15:47] to search retrieve and synthesize across wley's large corpus of life sciences and pharmarmacology [15:54] content. But what is very special about this example is that not only can this agent operate on [16:01] a standalone basis, it also works in orchestration with other agents in Microsoft discovery. And so [16:08] when a life sciences and pharmaceutical company does an important drug discovery task, it can use [16:15] this agent to help um accelerate hypothesis generation, but at the same time they can [16:20] then stay in this environment, right? to then um continue with it additional tasks like experiment [16:27] protocol generation result interpretation, dry lab and wet lab experimentation across the continuous [16:34] scientific reasoning loop. And so as I wrap up my talk today, there are a few things I want to [16:43] leave you with. First, we build AI agents to help solve the problem of information overload. But an [16:52] AI system that is pulling bad information doesn't solve that problem. It just it's just confidently [17:00] wrong at a faster rate. Second, there are three layers of trust that AI agents depends on and the [17:09] layer that's receiving the least attention. The the knowledge layer is the one that determines [17:15] whether other two layer even gets to work. And three, um, getting that knowledge layer wrong [17:22] has consequences across the entire system and the stakes are only getting higher with autonomous [17:28] agents. And so you may remember WY from your days, from your university days, uh, with, [17:36] you know, your textbooks or from research articles that you've previously read. But what I want you [17:42] to remember is we are that trusted knowledge layer. And we have made the leap by investing [17:48] and building on top of those trusted knowledge with knowledge feeds with real time intelligence [17:55] that can be plugged into your AI system in the moments you need it and inside the workflow you [18:01] are already working in. Oops. Um, and so as you buy, build and deploy agents and AI systems, [18:11] ask yourself the question, what is your system grounded in? Because the answer to that question [18:17] is the answer to whether it can be trusted. Um, so if you want, if you're interested in [18:25] learning more, you can find us at booth 1144. You can also feel free to connect with me on LinkedIn [18:32] or use the QR code on the left to learn about to learn more about WY's um AI solutions. Thank you.