Avi Goldfarb on AI and Predictive Analytics

Episode Summary

Executive Summary: Avi Goldfarb argues AI should be understood primarily as a prediction technology, not sci-fi general intelligence, and that its impact will depend on institutions, regulation, and redesigned workflows. Drawing on electricity, radiology, COVID testing, and Flint, he says the real bottleneck is often not prediction but turning predictions into trustworthy, scalable systems that humans can adopt.

Main Topics: AI as prediction, not machine autonomy (Priority: 5/5): Goldfarb frames AI as improved machine learning that fills in missing information. The book argues the key economic value is prediction, which then changes decisions rather than replacing humans outright. The 'between times' and slow diffusion of general-purpose technologies (Priority: 5/5): Using electricity as an analogy, the discussion emphasizes that transformative technologies often take decades to reshape industries because organizations must invent new processes, layouts, and roles. Rules, institutions, and regulatory barriers (Priority: 5/5): The conversation highlights how regulations built for old systems can slow adoption, even when the underlying prediction technology is ready. The main challenge is aligning rules, incentives, and trust. Human decision-making, bias, and accountability (Priority: 4/5): Goldfarb argues humans are often the source of discrimination and poor decisions, so AI can improve outcomes if humans remain responsible for decisions and oversight is properly allocated. Explainability, trust, and ex post auditing (Priority: 4/5): He is skeptical of broad real-time explainability requirements, calling them partly fictional, but supports auditability and post hoc investigation to detect bias or failures. Point solutions versus system solutions (Priority: 5/5): The hosts and Goldfarb distinguish incremental replacements of human tasks from broader system redesigns that create new value, like Uber or electrified factories. Case studies: COVID testing, Flint, radiology, and creative AI (Priority: 4/5): Examples show that a useful prediction can exist without immediate transformation because regulation, incentives, and workflows delay adoption or redirect AI into narrow workflow tasks.

Key Arguments: AI’s current economic importance comes mainly from better prediction, not from full machine autonomy or AGI. Like electricity, AI will take time to create large productivity gains because organizations must invent complementary processes and workflows. Rules are not external to AI adoption; they are 'glue' that can block or enable deployment depending on how they are designed. The biggest resistance often comes from incumbents and beneficiaries of current systems, not from technical limitations alone. Human bias is a major reason to adopt AI, since many existing human decision processes are discriminatory or inconsistent. Regulation should not demand no rules, but the right balance: enough oversight to build trust without freezing innovation. Explainability is useful only in limited cases; broad real-time explanations for complex models are often misleading, but ex post auditability is valuable. AI often replaces narrow tasks within workflows, but the bigger gains come when firms redesign the system around the technology. In medicine, algorithmic recommendations generally outperform human overrides when the model has been rigorously validated. Geopolitical competition may accelerate AI adoption because countries fear falling behind if others proceed faster.

Data Points: Electricity diffusion timeline: About 40 years - Goldfarb cites the lag between clear value in the 1880s and median household/factory electrification in the 1920s. COVID prevalence in summer 2020: About 1 in 1,000 Americans - Used to illustrate why workplace screening was an information problem for most people. Flint algorithm accuracy: About 80% accurate - University of Michigan professors’ model predicted which houses likely had lead pipes. Flint street-by-street replacement success rate: About 20% - Accuracy fell after politicians overruled the algorithm and used a fairness-based street-by-street approach. Flint post-court ruling success rate: About 70% - After legal pressure, the city had to follow the predictions again, improving targeting. Medical technology approval timeline: 5 to 10 years - Goldfarb notes health-tech regulation is slow, which is part of why radiology AI diffusion lags. Computers’ productivity lag: Around 40 years - He references the Solo-style argument that computers took decades to show up in productivity stats. Internet productivity lag: Around 25 years - He says the internet eventually appeared in productivity measures after a long delay. Taxi-driver improvement from AI tool: 14% - A cited NBER paper found AI reduced the gap between mediocre and top taxi drivers in pickup prediction by 14%.

Pivotal Quotes: "We should think about AI as prediction technology." — Avi Goldfarb: He defines the book’s central thesis early in the interview. "Rules are glue." — Avi Goldfarb: He uses this phrase to describe how institutions and regulation can either enable or block technological adoption. "There is no such thing as a machine decision. There is a human responsible." — Avi Goldfarb: He argues regulation should focus on human accountability, not the illusion that algorithms independently decide.

Implications: AI adoption will hinge less on raw model capability than on redesigning institutions, workflows, and accountability structures. Expect many near-term gains to look like narrow task automation, with larger productivity effects arriving only when firms and regulators adapt.