Inference, not prediction — Prof. Michael I. Jordan on what modern AI is still missing
Michael I. Jordan, a leading figure in machine learning and statistics, argues for reframing AI from a race for disembodied superintelligence to the design of collective economic systems. He critiques the AGI hype, advocates for integrating economic principles and robust uncertainty quantification into ML, and proposes a new intellectual framework for building technology that augments, rather than replaces, human systems.
Michael I. Jordan, who trained as a statistician and cognitive scientist, has never considered himself an “AI researcher.” His work has focused on building real-world machine learning systems for domains like supply chains, commerce, and healthcare. He argues that the recent rebranding of the field as AI, and then AGI, has been a “distortionary effect” on research, business models, and technology.
A Collectivist, Economic Perspective on AI
Jordan proposes a “collectivist, economic perspective” on AI, viewing it not as an attempt to mimic a single human brain but as a large-scale system involving billions of humans as producers and consumers. This perspective emphasizes that intelligence is social, contextual, and deeply intertwined with economic principles.
- Critique of the AGI Narrative: AGI is dismissed as a “PR term” that creates a false dichotomy for young builders: either be exuberant about superintelligence or alarmist about extinction. Jordan finds these narratives, promoted by “thought leaders,” to be demoralizing and lacking in economic or intellectual depth. He argues this sci-fi framing distracts from building useful systems that can create jobs and opportunities.
“It’s going to wipe out humanity with a with a high probability or it’s super intelligent arrive soon so there’s nothing left to do. That’s in your lifetime. That is so demoralizing.”
- Intelligence as a Social System: Much of human intelligence arises from aggregation, culture, and social context. Economics and game theory provide the mathematical language to formally study these interactions, including cooperation, competition, and incentives. This is contrasted with the view of the brain as a computer, which he finds to be a limited and often goal-less starting point.
Beyond Hype: Systems Thinking and Predictability
Current large language models (LLMs) are powerful statistical engines, but they are not systems. The prevailing vision of an “assistant-on-your-shoulder” is a weak business model. Instead, we should focus on integrating ML into complex ecosystems like healthcare, transportation, and finance.
-
Predictability Over Faux Understanding: It’s not necessary to fully understand the internal mechanics of a complex model, just as we don’t understand the detailed workings of a human brain. What is crucial is understanding its input-output behavior and ensuring it’s predictable. The goal is not mechanistic interpretability for its own sake, but creating actionable explanations. For instance, if a person is denied a loan, they don’t need to see a circuit diagram; they need to see examples of similar people who were approved to understand what they could change.
-
AlphaFold and Prediction-Powered Inference: While admiring AlphaFold, Jordan points out its limitations. When used to test novel scientific hypotheses at the edge of knowledge—where training data is sparse—it can be highly biased and produce overly confident, incorrect results. AlphaFold provides predictions but lacks the necessary error bars for the specific scientific question being asked. To address this, his group developed prediction-powered inference, a methodology that uses a small amount of ground-truth data to recalibrate the model’s outputs, providing statistically valid confidence intervals. This allows scientists to leverage the power of large models without being led astray by their biases.
The Role of Economics, Incentives, and Uncertainty
To build robust and beneficial systems, we must move beyond pure optimization and incorporate economic principles.
-
Drug Discovery as an Incentive Problem: The FDA drug approval process is not just a statistical problem of controlling false positives and negatives. The data comes from self-interested pharmaceutical companies. Without the right incentives, companies might submit many low-quality candidates, hoping for a false positive that leads to a profitable drug. The system must be designed using economic thinking (mechanism design) to incentivize companies to submit their most promising candidates, thereby improving the overall error rate of the system.
-
Data Markets and Contracts: A simple “three-layer data market” model illustrates these dynamics:
- Users provide data to a platform.
- Platforms use the data to improve their service.
- Third-party buyers purchase the data for market research. This creates a conflict: users lose privacy. An economic solution involves platforms offering tunable privacy levels (e.g., differential privacy). A user might choose a platform with higher privacy, which in turn makes their data less valuable to the buyer. This isn’t an optimization problem but an equilibrium problem, where statistical concepts (prediction accuracy) and economic concepts (incentives, social welfare) must be modeled together.
-
Quantifying Uncertainty: LLMs are notoriously bad at quantifying their own uncertainty. True uncertainty quantification is not just about producing a single error bar. It involves understanding the context:
- Economic Context: A single Bayesian duck might go to the side of the lake with a 2/3 chance of having more food. But in a population of ducks, the Nash equilibrium is for each duck to randomize, going to that side with 2/3 probability to avoid overcrowding.
- Information Asymmetry: In contracts, one party knows more than the other, creating a different kind of uncertainty that can’t be resolved by simple sampling.
- Provenance: Data should be tagged with metadata (e.g., how old it is), which should quantitatively factor into the uncertainty calculation. Current systems, including LLMs, do none of this.
A New Foundation for the AI Era
Jordan concludes by advocating for a new “liberal arts” for the modern era, a triangle of foundational disciplines:
- Computational Thinking: Abstractions, modularity, APIs.
- Inferential Thinking: Statistical reasoning, understanding uncertainty, data gathering.
- Economic Thinking: Incentives, markets, strategic interaction.
By integrating these three modes of thought, the next generation of builders can move beyond the narrow, hyped-up narratives and create systems that are robust, socially responsible, and genuinely beneficial to humanity.