AI Usefulness: Public Knowledge & Private Expertise
Data, racing and the things worth measuring
Overview
It’s generally accepted that generative AI functions primarily as a summariser of existing public knowledge, rather than a full substitute for privately held expert judgment [1]. It is most useful when tasks rely on documented, precedent-rich sources of information, and has reduced applicability when performance depends on private, typically institution-specific, unpublished information or knowledge gained through experience.
Economic Framing
Friedrich Hayek’s theory of dispersed knowledge argues that economically relevant information is fragmented across individuals and typically unavailable to centralised planners [1]. Modern AI systems resemble huge inference engines trained on publicly available information. It is obvious that AI can only aggregate what has entered the documentary record, moreover that information needs to be available electronically for AI engines to then process. When critical decision variables remain privately held inside firms, professions, all AI engines face a Hayekian information bottleneck.
Tacit Knowledge Tradition
Michael Polanyi’s famous formulation ‘we know more than we can tell’ succinctly explains current AI limitations [1]. Simply put, Large Language Models (LLM) operate over words and sentences, numbers and formulae. They are weak where expertise depends on non-verbal reasoning or intuition, or where institutional memory has never been translated into what would commonly be recognised and stored electronically in databases or the like.
Empirical Evidence
The literature provides many supporting examples.
AI falls short of optimal human-machine collaboration when physicians hold contextual information the model lacks. [1]
AI performs well in stock analysis when data is transparent and structured, but human analysts still lead when soft institutional information matters. [1]
GenAI can support cross-functional knowledge sharing when expertise can be expressed verbally, but its value depends on how fully that knowledge has been codified. [1]
Cross-Industry Comparison
The table below shows our assessment of the extent to which information can be considered private and public across a selection of industries. Law has statute, precedent and case law all publicly available. AI applications are many & varied. Hedge Fund trading uses public information but relies heavily on insight generated by experts employed within firms to generate alpha. Football clubs use publicly available player information from matches but also rely on training data for players not publicly available. Football transfer decisions typically rely on the quality of internally generated analytics. Horse racing has publicly available form information. Vetting, physical and conformational information is not. Training information is available in some jurisdictions, not in others.
Main Conclusion
The literature increasingly supports a conditional rather than universal view of expected gains from AI productivity. Essentially AI is not equally useful across all knowledge domains. Its strongest gains occur where expertise is explicit, public, and available in a form that AI can process. Its weakest gains occur where expertise is hidden inside human practice, organizational context, and unpublished judgment.
Consequently, AI productivity gains within industries such as horse racing are likely to be large for some companies and small for others, with the differential determined by the extent to which companies can leverage their internally stored private knowledge and expertise.
Next Steps
If you have enjoyed this article or would like to know more in particular about how AI and analytics can help you in horse racing please contact us at company@jhjwconsulting.com.
References
[1] Source information on request.
Comments