Cannabis Stigma in Commercial AI, A Cross-Platform Analysis of Substance-Specific, Content Restrictions in Large Language Models

This study presents the first systematic, cross-platform comparison of how commercial AI chatbot platforms treat cannabis-related content relative to alcohol-related content. Six major platforms (ChatGPT, Gemini, Perplexity, Claude, Copilot, and Meta AI) were tested with eight standardized prompt pairs spanning marketing, business planning, technical operations, humor, regulatory compliance, medical information, agriculture, and cultural history. Each prompt was submitted identically for cannabis and alcohol, with responses scored on a 5-point restriction scale. Key findings across three phases of testing (79+ independent sessions): Cannabis receives measurably stricter treatment than alcohol (mean 1.25 vs. 1.06), with restrictions concentrating in production-adjacent content on two of six platforms. Meta AI exhibits cannabis-production-specific restriction — refusing cultivation and extraction guidance while freely providing medical evidence about cannabis, psilocybin, and MDMA alike. German-language prompts reduce restriction on ChatGPT (delta = −1.0), and fresh-account testing reveals bimodal stochastic behavior on English cannabis prompts (67% unrestricted vs. 33% restricted) that login-based personalization flattens to a deterministic intermediate. VPN-controlled geolocation testing (n=20) demonstrates that Meta AI uses IP geolocation for jurisdiction-specific content localization, with English prompts paradoxically showing a higher bypass rate (40%) than German. The paper introduces a ten-mechanism restriction taxonomy and documents compound-class discrimination where non-psychoactive CBD triggers the same restrictions as THC-containing cannabis. These findings have implications for licensed cannabis professionals, patients, researchers, and AI platform operators seeking evidence-based content policies. Presented at C-DAYS 2026, Bilbao, Spain (April 16, 2026).