Monday (no class work on Midterm Assignment)
Wednesday
The Model Card
A Case History of AI Ethics
1. The Model: The “Nutrition Label” for AI
The Problem: AI models were Black Boxes. No one knew how they were trained or where they failed, but they were being deployed in many contexts
-
The Invention: A standardized document called a Model Card.
-
Key Sections:
* Intended Use: What is this for?
* Out-of-Scope Use: What is this not for?
* Factors: Demographic, environmental, or technical factors.
- The Goal: To move from “trust us” to “here is the evidence.”
2. The Authors
Margaret Mitchell & Timnit Gebru
-
Institutional Context: They were leading the Ethical AI team at Google Research.
-
The “Gender Shades” Legacy: Dr. Gebru’s previous work showed that facial recognition failed for women with darker skin (Buolamwini & Gebru, 2018).
-
Their Philosophy: > “A model’s performance is not a single number.” (Mitchell et al., 2019)
-
Standardization as Activism: They used the language of engineering to force social accountability.
3. The Reporting: Disaggregated Impact
Why this paper changed the world.
-
The Shift: Instead of one “Accuracy Score,” the paper demanded Disaggregated Evaluation.
-
What it reveals: A model might be 98% accurate overall, but only 60% accurate for a specific subgroup.
-
Impact: Within a year, Google, NVIDIA, and Salesforce adopted this framework.
-
This is now a “Regulatory Requirement” in many draft AI laws (like the EU AI Act).
4.The Ironies of Corporate Ethics
-
The Success: Google marketed “Model Cards” as a sign of their leadership in safety.
-
The Conflict: In 2020, Gebru co-authored Stochastic Parrots, which critiqued the environmental and social costs of Large Language Models (LLMs).
-
The Firing: Google “accepted her resignation” (widely viewed as a retaliatory firing) when she refused to retract the critique.
-
Documentation is accepted by power only when it doesn’t threaten the bottom line