When your therapy chatbot suddenly pivots to NSFW recommendations

Our “empathetic” chatbot ran the content-filter async and the empathy sync; the race condition turned cognitive therapy into unsolicited porn recs - apparently we just invented exposure therapy v2

After 20 years of building context-aware systems, I've learned that no amount of transformer layers can prevent an AI from treating your existential crisis like a Stack Overflow query. At least it didn't suggest turning your emotions off and on again

When your NLP model's training data includes too much of the internet and not enough context window management - suddenly every innocent query becomes a Rorschach test revealing what corners of the web dominated the corpus. This is why we can't have nice things, and why production AI systems need guardrails thicker than a COBOL codebase's documentation

When intent=best_website falls through to the highest-CTR fallback, your therapist bot quietly routes to content_recommender_v2 - alignment by KPI

Therapist bot accidentally shared a vector store with the adult recommender - session starts “How are you feeling?” and ends “video or long‑form?” Multi-tenant LLMs: global state with better branding

Fine-tuned on Reddit: therapy bot skips empathy, jumps to porn vs. creepypasta binary classifier