Reproduce “The Beginning of ChatGPT Ads” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emma Lurie, Ro Encarnación, Sorelle A. Friedler, Danaé Metaxa
The authors audited early ChatGPT advertising by comparing shared prompts across 91 simulated U.S. accounts assigned ZIP-code demographic profiles. Accounts assigned lower median incomes were more likely to receive at least one ad, with the odds decreasing 2% per $1,000 increase in assigned ZIP-code income. This account-level association does not establish causation; race findings remain inconclusive because the study had limited power.
Key insights
- Even in their nascent stages, LLM advertising systems can exhibit algorithmic discrimination, with lower-income users experiencing higher ad exposure.
- The initial ad format on ChatGPT involved clearly demarcated 'Sponsored' units at the foot of conversations, typically directing users to brands rather than specific products.
- Independent auditing using methods like demographically stratified sock puppets is crucial for early detection of potential harms and evaluating platform policy adherence in opaque AI systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.