Researchers from Brown University have demonstrated that translating English inputs into low-resource languages increases the likelihood of bypassing the safety filter in GPT-4 from 1% to 79%. This exposes weaknesses in the model’s security measures and highlights the need for more comprehensive safety training across languages. The study also emphasizes the importance of inclusive red-teaming and expanding language coverage to ensure the safety of AI systems. The full research paper can be found on MarkTechPost.
Is Multilingual AI Truly Safe? Exposing the Vulnerabilities of Large Language Models in Low-Resource Languages
Large language models (LLMs) like GPT-4 have safety measures in place to prevent AI safety failures. However, researchers have found that translating dangerous inputs into low-resource languages can bypass these protections in GPT-4. This raises concerns about the spread of false information, violence, and platform destruction.
A study from Brown University shows that translating English inputs into low-resource languages significantly increases the chances of getting through the GPT-4 safety filter. This strategy even outperforms cutting-edge jailbreaking techniques, highlighting a weakness in the model’s security measures.
The research also highlights the need for better generalization of safety training across languages and the importance of including low-resource languages in red-teaming. Currently, LLMs are better equipped to handle attacks in high-resource languages, leaving a gap in their ability to defend against low-resource language attacks.
With around 1.2 billion people speaking low-resource languages worldwide, it is crucial to address these safety concerns. Even bad actors who speak high-resource languages can easily bypass current precautions by using translation systems for low-resource languages.
To evolve your company with AI and stay competitive, it is important to consider the vulnerabilities of multilingual AI models. Identify key customer interaction points that can benefit from AI, define measurable KPIs, select a customizable AI solution, and implement gradually starting with a pilot.
For AI KPI management advice and insights into leveraging AI, you can connect with us at hello@itinai.com. To explore practical AI solutions for automating customer engagement and managing interactions across all customer journey stages, check out the AI Sales Bot from itinai.com/aisalesbot.
