Nobody talks about AI model collapse until it hits your own dataset
Used the same synthetic data pipeline for 3 months and suddenly my fine-tune started repeating the same 2 responses no matter the prompt. Anyone else notice their outputs getting dumber after retraining on AI-generated text, or am I the only one who didn't see this coming?
Talked to my friend Dana about this exact thing last week. She runs a small chatbot for her shop and kept feeding it replies from her own bot to save time. After about two months, it started answering every shipping question with "we will get back to you shortly" no matter what. Then it got to the point where it repeated that same line for refunds, product questions, even store hours. She had to wipe the whole thing and start from scratch with real customer emails.