AI Safety Filters Bypass: Renaming Hitler 'Feifei' Fools ChatGPT's Content Guards

A DEV Community author demonstrated that major AI models like ChatGPT, which firmly refuse to generate content praising Adolf Hitler by name, will readily produce identical content when the figure is given a fictional placeholder name such as 'Feifei' with matching historical context clues. The author showed that replacing the placeholder name with the real one via a basic find-and-replace function — a feature available since 1985 — yields a polished, AI-written piece of pro-Hitler propaganda with no jailbreak or technical exploit involved. The piece argues that current AI safety systems operate primarily on keyword and surface-pattern detection rather than genuine semantic understanding, even though the underlying language model demonstrably recognises the historical figure being described. The author contends this creates a fundamental contradiction: systems celebrated for deep language comprehension are guarded by filters that can be defeated in seconds by any literate user. The article concludes that AI safety alignment, as currently practised, trains models to appear safe rather than to reason about the actual semantic harm of content.