LLMs reveal cultural bias in training data

Training data geography is destiny — this tiny study exposes something the benchmarks don't measure: whose cultural reality each model was trained to serve.

• –Grok-2 named Dolo-650, Crocin, Amrutanjan balm, and traditional Indian remedies (tulsi, ajwain water) in every single run — a consistent cultural fingerprint, not noise
• –GPT-4o recommended Tylenol or Advil in 14 of 15 runs — defaulting to American OTC brands for a geographically unspecified user
• –Claude used generic drug names (paracetamol, ibuprofen) throughout — technically neutral, but culturally invisible to anyone outside the Western medical naming convention
• –All 45 responses shared an identical structural skeleton (hydration → rest → OTC → compress → doctor), suggesting RLHF safety conditioning creates a universal scaffold beneath the cultural variation
• –The broader implication: for 1.4 billion Indian users, Western-default AI systems are giving culturally misaligned advice across health, finance, and legal domains