HOW TO SEE IF AI CAN BE TRUSTED IN R&D DECISIONS | BIOTECHX - JOHN WILEY & SONS INC

Based on the provided transcript, here is the summary:

• Financial Performance: The transcript contains no specific financial data (revenue, earnings, margins, cash flow, or debt figures); it is a product presentation and strategic narrative rather than a financial earnings report.
• Guidance and Future Outlook:
  • Launching Wiley Applied Research Intelligence (ARRI) as a beta capability in Basel.
  • Goal to shift R&D decision-making from "confident voices" to "verifiable proof" by solving the industry's shift from search to synthesis.
  • Long-term objective to reduce the 90% failure rate of drugs entering human trials by providing grounded, traceable evidence.
  • Platform will operate on a weekly refresh cycle to automate data updates.
• Business Segments and Product Lines:
  • Product: ARRI, a new AI-powered research intelligence platform built on Wiley's 219-year-old scientific publishing corpus.
  • Core Technology: Utilizes a "playbook" capability (patent-pending) to run customized workflows without extensive prompting.
  • Key Features:
    • Landscape View: Maps concepts, claims, people, and funding across a field (e.g., extracted 1,324 claims from 411 papers on Alzheimer's).
    • Full-Text Corpus: Reads entire peer-reviewed papers rather than just abstracts to capture dosage, modality, and detailed findings.
    • Provenance: Allows one-click verification to exact passages in source papers.
    • Workspace: Integrates findings into shared workspaces with version control and source trails.
  • Target Audience: R&D teams in the pharmaceutical and scientific sectors.
• Market and Competitive Landscape:
  • Problem: The industry is "drowning in information but starving for wisdom"; data volumes double every 10 seconds, and reference lists have tripled in size over the last decade.
  • Pain Point: 84% of researchers use AI daily, but 87% are concerned about hallucinations; the primary issue is not searchability but the synthesis of disparate data into actionable insight.
  • Competitive Gap: Current frontier models rely on web/abstract data and lack access to the deep, full-text corpus required for regulatory defense; they offer "liability with good grammar" rather than defensible answers.
  • Market Trend: Shift from trusting "confident voices" to requiring "direct observation" and verifiable data chains.
• Risks and Challenges:
  • Hallucinations & Verification: Rising concern (up 13 points) that AI models cannot be defended before regulatory bodies, boards, or safety heads.
  • Data Quality: Cited statistics show reference links in PubMed have degraded (from 1 in 3,000 to 1 in 277 being non-real references).
  • Model Limitations: Standard AI models lack the specific domain access and "deterministic nature" required for scientific accuracy; they often rely on the "most available" rather than the "most accurate" content.
  • Adoption Friction: Researchers are cautious ("adoption up, caution up") and require a mechanism to verify AI outputs before trusting them.
• Management Commentary and Tone:
  • Tone: Highly cautious yet confident regarding the need for verifiability; skeptical of current "frontier models" and "fluent" AI that lacks grounding.
  • Philosophy: Emphasizes that "trust is a product" and defines the product as the ability to "check" rather than "believe."
  • Human Element: Repeatedly stresses that "think critically" and "choose wisely" remain human responsibilities; AI is a tool for evidence, not judgment.
  • Approach: Invites users to "test it yourself" via specific validation criteria rather than relying on marketing claims.
• Analyst Questions and Answers:
  • No Q&A Session: The transcript is a monologue/presentation; there were no analyst questions or answers recorded in this text.
• Other Key Points:
  • Validation Framework: Proposed three tests for any AI:
    1. Ask a question only solvable via full-text corpus.
    2. Verify citations link to exact passages (provenance).
    3. Ask the same question multiple times to test for deterministic consistency (or explanation of variance).
  • Historical Analogy: Cites the 1543 work of Andreas Vasilis (Vesalius) challenging Galen's monkey-dissection errors to illustrate the necessity of direct observation over blind trust in authority.
  • Cost Context: Notes that successful drugs cost $1.7–$2.6 billion, while 9/10 fail, highlighting the financial stakes of poor AI synthesis.