Course contentsShow
← AI Engineering
Lesson 1562 of 1,88638. Bias, Fairness, and AlignmentFree lesson

Allocation Harms vs Representation Harms

Distinguish between biases affecting resource distribution and those affecting social representation or dignity.

Allocation Harms vs Representation Harms

You'll learn how to distinguish two major categories of AI bias: those that unfairly distribute resources versus those that damage dignity and social perception.

Core Distinction

Not all AI harms look the same. When we talk about bias in AI systems, we're often dealing with two fundamentally different types of damage:

Allocation harms occur when an AI system distributes opportunities, resources, or services unequally. Think of a loan approval model that denies mortgages to qualified applicants from certain neighborhoods, or a hiring AI that filters out resumes with certain names. These are tangible, material consequences—someone gets something or they don't.

Representation harms occur when an AI system reinforces stereotypes, erases identities, or damages the dignity of individuals or groups. This includes an image generator consistently depicting CEOs as male, a chatbot using offensive language about particular communities, or a search system failing to recognize certain cultural contexts. The harm here is social and psychological—it shapes how people are perceived and valued.

Why the Distinction Matters

These categories require different measurement and mitigation strategies. Allocation harms can often be quantified through metrics like demographic parity or equal opportunity rates (concepts you've explored when measuring bias). Representation harms require qualitative analysis, user studies, and deeper cultural understanding.

A single system can cause both types simultaneously. A recruitment tool might allocate fewer interviews to women (allocation harm) and generate job descriptions with masculine-coded language (representation harm).

Key Takeaway: Allocation harms affect who gets what resources or opportunities, while representation harms affect how people are portrayed and perceived—both matter, but require different evaluation and mitigation approaches.