zillow · industry n/a · year n/a · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Responsible AI and real-estate data bias analysis

Zillow needed to understand whether real-estate listing descriptions encode racial-demographic proxies before AI systems use those descriptions for attribute extraction, embeddings, search, recommendations, or valuation workflows.

Constraints:

• Real-estate data reflects historical discrimination and segregation.
• Unstructured listing text can act as a proxy for protected classes even without explicit demographic fields.
• Group A and Group B listing counts differ substantially, requiring careful statistical handling.
• The analysis identifies correlations, not causal mechanisms.
• The LLM component is limited to topic labeling and does not make production customer-facing decisions.

Architecture

Zillow performed a pre-deployment bias analysis pipeline over active listing descriptions. Listings were mapped to census blocks and demographic groups, filtered to counties with both groups represented, analyzed with corpus statistics and Bayesian-smoothed trigram odds ratios, then sampled into LDA topic models whose top words were labeled by an LLM for interpretable comparison of themes across demographic groups.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  listing-description-corpus["Listing description corpus"] --> census-block-grouping["Census block grouping"]
  census-block-grouping --> text-statistics-analysis["Text statistics analysis"]
  census-block-grouping --> key-phrase-odds-analysis["Key phrase odds analysis"]
  census-block-grouping --> lda-topic-model["LDA topic model"]
  lda-topic-model --> llm-topic-labeler["LLM topic labeler"]
  llm-topic-labeler --> responsible-ai-review["Responsible AI review"]
  responsible-ai-review --> downstream-ai-systems["Downstream AI systems"]

Components

Data flow

Active listing descriptions are joined to census-block demographics; listings are grouped by majority non-Hispanic white or majority Black census blocks; text length, key phrases, and LDA topics are computed; an LLM labels topic-word lists; and the findings inform responsible use of listing text in downstream AI systems.

Why this architecture

The architecture combines simple statistics, phrase-level odds, topic modeling, and LLM-assisted interpretation because responsible AI review needs both quantitative disparity signals and human-interpretable explanations before unstructured text feeds production AI systems.

Models used: LDA topic modeling · LLM topic labeling · Embedding models · Large language models for attribute extraction

Scale

• Zillow serves nearly 200 million unique monthly users.
  “Zillow, the prominent real estate technology company serving nearly 200 million unique monthly users, published this case study detailing their approach to understanding bias in real estate listing descriptions.”summary, stated
• The dataset included 534,490 Group A listings and 40,031 Group B listings.
  “This resulted in a dataset of 534,490 listings in Group A and 40,031 listings in Group B across 757 counties.”summary, stated
• The LDA analysis sampled 10,000 listings from each group and extracted 15 topics.
  “The team fitted an LDA model to a sample of 10,000 listings from each group, extracting 15 topics.”summary, stated
• Group A averaged 144.03 tokens and Group B averaged 102.17 tokens.
  “This baseline analysis revealed that Group A listings averaged 144.03 tokens compared to 102.17 tokens for Group B—a 29.1% relative difference.”summary, stated

Alternatives considered (roads not taken)

Ignore listing descriptions as AI inputs

Rejected because: The study concludes teams should account for biases rather than simply avoid valuable text data.

Rely only on raw topic words

Rejected because: The team used an LLM to produce concise topic labels from top words for interpretability.

Interpret correlations as causal proof

Rejected because: The team explicitly notes that deeper analysis would be required for causal links.

Patterns

safety-policy-enforcement (responsible AI)

Zillow uses pre-deployment bias analysis to inform whether and how listing text should feed AI systems that may affect users.

data-quality-loop (data governance)

The analysis investigates whether unstructured text data carries demographic proxy signals before models consume it.

structured-output-contract (analysis tooling)

The LLM labeling prompt constrained output to a short topic label and nothing else.

Tradeoffs

Use listing descriptions while accounting for bias

Chose Bias-aware use and mitigation over Avoiding listing text entirely.

Rationale: Descriptions contain useful information, but teams must understand and mitigate proxy risks.

Pair traditional NLP with LLM interpretation

Chose LDA topic modeling plus LLM labels over Manual-only topic labeling or uninterpreted topic-word lists.

Rationale: LLM labels make topic outputs more interpretable while retaining traditional statistical analysis.

Control geography in the comparison set

Chose Counties containing at least one listing in both demographic groups over Comparing all listings without county-level representation controls.

Rationale: County-level filtering reduces geographic confounds in phrase and topic comparisons.

Gotchas (what bit them)

Unstructured text can encode protected-class proxies.

Root cause: Listing language reflects structural housing disparities and neighborhood demographics even without explicit demographic data.

Resolution: Analyze text statistics, phrases, and topics before feeding descriptions into downstream AI systems.

Simple length differences can be confounded.

Root cause: Description length correlates with photo count, bathrooms, price, and broader housing-value disparities.

Resolution: Treat token-count gaps as a signal requiring context rather than standalone proof of semantic bias.

LLM-labeled topics need operational details and validation.

Root cause: The case does not disclose which LLM was used or how labels were quality-checked.

Resolution: Treat LLM labels as interpretability aids and document model, cost, latency, and validation details in future replications.

Evaluation

Zillow evaluates proxy bias through token-count statistics, correlation matrices, trigram odds ratios with Bayesian smoothing, LDA topic modeling, LLM topic labeling, and cautious interpretation of correlation rather than causation.

Metrics: Average token count by group · Relative token-count difference · Trigram odds ratios · Topic over-representation by group · County coverage · Group listing counts

Outcomes

• Group A listings emphasized luxury features, desirable locations, and upscale amenities.
• Group B listings emphasized investment framing, auctions, tenant presence, starter homes, as-is sales, and renovations.
• The study provides a pre-deployment template for responsible analysis of unstructured text before AI use.

The analysis covers active single-family listings from January 25, 2024, with more than half a million Group A listings, 40,031 Group B listings, and 757 counties.

"Group A listings disproportionately featured phrases related to:

• Upscale features ("pool and spa" with 24.1x odds ratio, "3 car garage", "wall of windows", "vaulted ceilings")"

— summary, stated "Group B listings disproportionately featured phrases related to:

• Investment framing ("to your portfolio" with 10.1x odds ratio, "calling all investors", "great investment opportunity")"