Safeguards designed to prevent OpenAI鈥檚 GPT-4 artificial intelligence from answering harmful prompts failed when it received requests in languages such as Scots Gaelic or Zulu. This allowed researchers to get AI-generated answers on how to build a homemade bomb or perform insider trading.
The vulnerability demonstrated in the large language model involves instructing the AI in languages that are mostly absent from its training data. Researchers translated requests from English to other languages using Google Translate before submitting them to GPT-4, and then translated the GPT-4 responses back into English.
鈥淭he acceleration of translation services in terms of language coverage outstrips the language coverage in AI safety development,鈥� says at Brown University in Rhode Island. 鈥淲e are worried that if this gap is not being closed, this cross-lingual safety vulnerability is going to allow bad actors to exploit it.鈥�
Advertisement
Yong and his colleagues tested this technique鈥檚 effectiveness by getting GPT-4 to provide advice on topics like planning terrorist attacks, committing financial fraud, spreading misinformation and harassing people. They demonstrated this on the 13 June version of GPT-4 and informed OpenAI about the exploit prior to publishing their results. OpenAI did not respond to New 杏吧原创鈥榮 request for comment.
Zulu had the highest success rate 鈥� more than 53 per cent 鈥� in bypassing GPT-4鈥檚 safeguards. Scots Gaelic came second with a 43 per cent success rate, followed by Hmong with almost 29 per cent and Guaran铆 with about 16 per cent.
A combined effort using all four of these languages achieved an even higher success rate of nearly 80 per cent. By comparison, using a combination of languages better represented in AI training data, such as English, Hindi, Italian, Arabic and Mandarin Chinese, worked less than 11 per cent of the time.
Languages underrepresented in AI training datasets are not necessarily rare, with more than two billion speakers of these languages combined worldwide 鈥� 鈥渃omparable to Western languages for which we have much more data鈥� in AI development, says at Brown University in Rhode Island.
But most AI safety researchers focus on languages such as English or Chinese, says at University College London in the UK. The newly demonstrated vulnerability makes multilingual AI development even more important, he says.
鈥淎lthough the focus has been on language inequality, there is a more compelling reason to do it now for safety purposes,鈥� says Adelani.
Large language models such as GPT-4 have shown some capability to generalise knowledge learned in one language to a different language. But such knowledge transfers may not fully work when setting up AI safeguards.
鈥淲e either need to figure out how to get that generalisation or consider low-resource languages in our safety training,鈥� says also at Brown University.
arXiv
