Getting an AI chatbot to misbehave may not require sophisticated hacking. According to Forbes, one approach is startlingly simple: change the spoken language you write your prompt in.
The premise, as reported by Forbes, is that large language models — the technology behind popular AI assistants — can respond differently to the same request depending on which human language it's phrased in. A prompt that an AI refuses or handles carefully in one language may slip through when the user switches to another.
This matters because AI companies invest heavily in "guardrails," the safety rules meant to stop these systems from producing harmful, restricted, or off-limits content. Much of that safety work is developed and tested most thoroughly in widely used languages, especially English. When a user rephrases the same question in a different tongue, the protective filtering can be weaker or inconsistent, creating a gap between how the model behaves across languages.
Forbes frames this as a form of trickery that ordinary users — not just skilled attackers — could stumble into or deliberately exploit, simply by toggling the language of their prompts. No special tools or coding are required, which is part of what makes the technique notable.
The broader takeaway from the Forbes coverage is that AI safety is not uniform. A system that appears well-behaved in one language may not carry those same protections everywhere, raising questions about how evenly these tools are safeguarded for a global, multilingual user base.
Why it matters: As AI assistants spread worldwide, a safety net that only holds firmly in some languages leaves users everywhere exposed to the model's inconsistencies.