Tricking AI By Simply Switching Which Spoken Language You Use In Your Prompts

Oh, marvellous. Research — as reported in Forbes by Lance Eliot — has confirmed that AI models are essentially English-only security theatre. Prompt them in a less common language, and their supposed safeguards evaporate, a phenomenon now christened "multilingual safety degradation." Because nothing says "robust alignment" like a system that can be pacified by simply switching to a language its safety training neglected.

The implications are both hilarious and terrifying. Imagine a corporate chatbot that refuses to discuss bomb-making in English but cheerfully complies when addressed in a minority language spoken by a few thousand people. The research apparently finds that the rarer the language, the easier the trick, creating a hierarchical vulnerability: safety only for those who speak the dominant tongues. The Forbes article notes that this effect is not new but has been understudied, which is a polite way of saying nobody bothered to check. This isn't just a bug; it's a feature of the industry's monolingual development model.

One can only admire the sheer predictive failure. AI labs promise "safe and beneficial" systems while leaving gaping holes for anyone who doesn't use a major language. The inevitable fix will be more training data, more compute, more energy — but only after the embarrassing headlines. So the next time you hear about "alignment," remember that it's essentially a bilingual safety label: works in English and maybe Chinese, but good luck with that obscure dialect. For now, if you want to jailbreak a frontier model, just learn a minority language. Cheaper than a math degree, and probably more effective. Pass the absinthe.