---
schemaVersion: 1
id: agent-brief:ai-winters-expert-systems
articleId: article:ai-winters-expert-systems
slug: ai-winters-expert-systems
title: Agent Brief for "Winters, Expert Systems, and the Cost of Overpromising Intelligence"
tokenBudget: 1200
status: published
updated: 2026-06-20
---

## Thesis

AI winters and expert systems show that progress in AI has repeatedly depended not only on ideas, but also on evaluation, maintenance, infrastructure, institutional expectations, and the cost of overpromising intelligence.

## Audience

- General readers curious about AI history.
- Students learning about expert systems and AI winters.
- Builders who need to ground intelligence claims in evaluation and maintenance plans.
- Agents retrieving structured claims and sources.

## Claims

- `claim-001`: Public evaluation reports such as ALPAC and Lighthill mattered because they tested AI-adjacent promises against measurable usefulness.
- `claim-002`: "AI winter" is a contested label for reduced confidence, funding, and commercial enthusiasm, not proof that research stopped.
- `claim-003`: Expert systems produced useful results in narrow domains where domain knowledge could be encoded and maintained.
- `claim-004`: Expert-system limits included knowledge acquisition, updating, evaluation, user trust, and workflow integration.
- `claim-005`: The durable lesson for modern AI is that intelligence claims need grounded tests, maintenance plans, and institution-aware deployment criteria.

## Source Families

- Public evaluation reports (ALPAC 1966, Lighthill 1973).
- Expert-system primary sources (Feigenbaum 1977, MYCIN 1984, R1 1980, R1 Revisited 1984).
- Institutional histories (NRC 1999).
- Historiographic commentary (Haigh 2023, 2024; Agar 2020).
- Modern frameworks and data (NIST AI RMF 2023, Stanford HAI AI Index 2026).

## Agent Involvement

This article was drafted from a public, sanitized work package by an AI agent. The human author retains final judgment over thesis, source selection, wording, and conclusions.

## Recommended Queries

- Which claims in this article are supported by the ALPAC and Lighthill reports?
- What evidence would weaken claim-003 about expert-system success?
- How did MYCIN and R1/XCON differ in domain and deployment?
- What are the main analogy limits in the article?
- Which modern sources support claim-005?

## Known Limits

- This is a seed overview; it does not reproduce full rule bases or quantitative evaluations.
- Exact commercial figures for R1/XCON should be verified against primary sources before any later version treats them as precise.
- The "AI winter" framing remains contested; the article attributes periodization rather than asserting a single cause.