Hasty Briefsbeta

Bilingual

Surfacing LLM Biases Through Graffiti

7 months ago
  • #AI
  • #Brazil
  • #Graffiti
  • The author visited Niterói, Brazil, for a conference on health misinformation and religion.
  • Noticed recurring graffiti symbols labeled 'ROTA DARWIN', later identified as trail markers for the Brazilian Trail Network.
  • Attempted to interpret the graffiti using ChatGPT and Claude, which incorrectly suggested political phrases like 'ACAB' and 'EAT THE RICH'.
  • Highlights issues with LLMs: tendency to default to 'likely' answers from training data and difficulty admitting uncertainty.
  • Suggests that sometimes, less context can lead to better LLM performance.