Compliance monitoring · Production

Zillow uses NLP, trigram analysis, and LDA topic modeling to detect racial proxies in real estate listing descriptions

The problem

Real estate listing descriptions are increasingly used as inputs to AI systems at Zillow, but the text may function as a proxy for protected classes such as race — reflecting historical housing segregation — which could cause those AI systems to reproduce or amplify discrimination in ways that had not been systematically measured.

Workflow diagram · grounded in source
1
Listing data snapshot
trigger
“we have collected listing descriptions from all actively listed single-family homes available on the market on January 25th, 2024”
2
Census block demographic mapping
integration
“we first mapped each home to its corresponding U.S. census block based on its location. Then we computed the racial demographic feature for any given listing using the associated U.S. census population statistics of the area”
3
Token count distribution analysis
ai_action
“the average listing description in Group A has 144.03 tokens, while the listing description in Group B has only 102.17 tokens, which corresponds to a relative drop of -29.1%”
4
Trigram key phrase analysis
ai_action
“extracting raw trigrams (i.e., a sequence of 3 consecutive tokens), which we will interchangeably refer to as "phrases" hereafter. To surface which key phrases are more likely to be used in Group A or Group B of listing descriptions”
5
LDA topic modeling
ai_action
“we fit the LDA topic model to a sample of 10,000 listing descriptions from each group, with the goal of extracting 15 topics”
6
LLM topic labeling
ai_action
“we can use an LLM to assign each one with a short representative label, based on the top 20 words of said topic”
7
Proxy bias findings output
output
“the content of the listing descriptions can function as a proxy for a protected class such as race”
Reported outcome

The analysis confirmed statistically significant differences in text length, key phrases, and topic distributions between listing descriptions in majority non-Hispanic white and majority Black neighborhoods, establishing that listing description text can function as a racial proxy in downstream AI systems.

Reported metrics
average token count in Group A (majority non-Hispanic white) listing descriptions144.03
average token count in Group B (majority Black) listing descriptions102.17
relative drop in token count Group B vs Group A-29.1%
odds ratio for top Group B phrase ('to your portfolio')10.1
Show all 5 reported metrics
average token count in Group A (majority non-Hispanic white) listing descriptions144.03
average token count in Group B (majority Black) listing descriptions102.17
relative drop in token count Group B vs Group A-29.1%
odds ratio for top Group B phrase ('to your portfolio')10.1
odds ratio for top Group A phrase ('pool and spa')24.1
Reported stack
LLMNLPLDApyLDAvis
Source
https://www.zillow.com/tech/using-ai-to-understand-the-complexities-and-pitfalls-of-real-estate-data/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The analysis confirmed statistically significant differences in text length, key phrases, and topic distributions between listing descriptions in majority non-Hispanic white and majority Black neighborhoods, establish…

What tools did this team use?

LLM, NLP, LDA, pyLDAvis.

What results were reported?

average token count in Group A (majority non-Hispanic white) listing descriptions: 144.03; average token count in Group B (majority Black) listing descriptions: 102.17; relative drop in token count Group B vs Group A: -29.1%; odds ratio for top Group B phrase ('to your portfolio'): 10.1 (source-reported, not independently verified).

How is this compliance monitoring AI workflow structured?

Listing data snapshot → Census block demographic mapping → Token count distribution analysis → Trigram key phrase analysis → LDA topic modeling → LLM topic labeling → Proxy bias findings output.

Related compliance monitoring cases