chatgpt-failures

Why we included this project

Anyone who watched the early 2023 chatbot wave knows how fast these models could go off the rails, and this archive preserves the evidence. It collects documented failures of ChatGPT, New Bing, and similar large language models, with each case dated and linked back to its original post on Twitter, Reddit, Hacker News, or a blog. Many entries include the full transcript plus a note on what the correct answer should have been, which makes the collection far easier to study than scattered screenshots. Researchers and engineers can pull reproducible examples of arithmetic errors, hallucination, prompt injection, and jailbreaking, or reuse the transcripts to build synthetic test and training data. It is a reference collection rather than software you run, but for anyone designing evaluation sets or trying to understand where these models break down, it saves a lot of digging.