Every speculative cycle feels rational while you are inside it.
- In 1999, it was bandwidth.
- In 2006, it was home equity.
- In 2021, it was liquidity without cost.
- In 2026, it is autonomy.
Capital today is not betting on faster software or marginal efficiency gains. It is betting on the idea that machines can now replace judgment itself—that probability has crossed some invisible threshold and become intelligence. That belief is embedded not just in venture decks and earnings calls, but in hiring plans, labor forecasts, productivity models, and national competitiveness narratives.
The uncomfortable question is not whether artificial intelligence is real. It is whether markets have priced autonomy that does not yet exist. And historically, when markets price something imaginary as inevitable, the unwind is never gentle.
##### Abstract
The current artificial intelligence investment cycle is driven less by technological discontinuity than by a persistent category error: the conflation of machine learning (statistical inference) with artificial intelligence (autonomous, goal-directed agency). This paper argues that “agent washing” has distorted venture capital deployment, public-market valuations, and labor substitution expectations.
I. The Taxonomy Trap: Statistics Wearing an Intelligence Costume
At the foundation of the AI boom is a linguistic sleight of hand.
Fact 1. Modern machine learning systems optimize probability distributions, not causal explanations. They estimate likelihoods $P(Y|X)$, not mechanisms (why Y happens).
[SOURCE] Judea Pearl, The Seven Tools of Causal Inference (PDF)
Fact 2. Michael I. Jordan has explicitly argued that what is marketed as “AI” today is best understood as large-scale statistical engineering, not autonomous reasoning systems.
[SOURCE] Jordan (2018), PDF via Stock Psycho
Fact 3. This view is not fringe. It is mainstream among senior computer scientists.
[SEARCH] Google Scholar (Jordan + AI statistics)
Fact 4. The field has repeatedly rebranded statistics to reignite capital interest: Expert Systems -> Data Mining -> Big Data -> Machine Learning -> Agentic AI.
[SEARCH] Google Scholar (AI hype cycles)
II. Agent Washing: How Autonomy Is Marketed Into Existence
Fact 5. In economics and computer science, an “agent” implies autonomy, goal persistence, and environmental adaptability.
Fact 6. Most commercial “AI agents” today are: Prompt-chained LLM calls, deterministic workflows with stochastic outputs, or scripted automation wrapped in narrative.
[SEARCH] Google Scholar (LLM agents limitations)
Fact 8. This mirrors ESG greenwashing: unverifiable claims, selective demos, and narrative-driven capital flows.
[SEARCH] Google Scholar (greenwashing financial markets)
III. Why Due Diligence Fails: The Metric Gap
- Fact 9. Accuracy benchmarks ignore distributional shift, the dominant real-world failure mode. (Amodei et al., 2016 PDF)
- Fact 10. State-of-the-art models degrade sharply outside benchmark conditions. (Recht et al., 2019 PDF)
- Fact 11. Even uncertainty estimates collapse under shift. (Ovadia et al., 2019 PDF)
IV. Agent Reality Checks: When the Demo Meets the World
Fact 14. Autonomous coding agents fail a majority of real software tasks without human intervention. [SOURCE] SWE-Bench paper (PDF)
Fact 15. Independent evaluations show agents overfit demos and collapse in open-ended tasks. [SOURCE] METR research hub
V. Systemic Risk: When Errors Correlate
Fact 17. Financial regulators warn that autonomous systems amplify systemic risk through correlated errors. [SOURCE] BIS AI & Financial Stability (Permanent Link)
Fact 18. Similar warnings appear across global financial oversight bodies. [SEARCH] Google Scholar (FSB AI Research)
##### Conclusion: Probability Is Not a Promise
The danger is not machine learning. The danger is selling probability as autonomy. The party ends when belief outruns output. It always does.