When LeetCode problems were modified slightly, answers were still correct 67% percent of the time. When a problem was custom, however, ChatGPT answers were correct only 25% of the time.