Red Teaming Contemporary AI Models: Insights from Spanish and Basque Perspectives

TL;DR

This work investigates the trustworthiness of Spanish and Basque language processing in large language systems through Red Teaming. It evaluates OpenAI o3-mini, DeepSeek R1, and ALIA Salamandra across two 3-hour sessions by ten experts, collecting 670 conversations and identifying 254 inappropriate outputs. The study provides the first Spanish RT evaluation of o3-mini and DeepSeek R1 and a Basque-focused assessment of Salamandra, revealing substantial bias and safety vulnerabilities, with Salamandra performing worst overall. Findings underscore the ongoing challenges of multilingual AI safety and support the value of open, public infrastructure like ALIA for advancing responsible, language-specific AI research and repair efforts.

Abstract

The battle for AI leadership is on, with OpenAI in the United States and DeepSeek in China as key contenders. In response to these global trends, the Spanish government has proposed ALIA, a public and transparent AI infrastructure incorporating small language models designed to support Spanish and co-official languages such as Basque. This paper presents the results of Red Teaming sessions, where ten participants applied their expertise and creativity to manually test three of the latest models from these initiatives$\unicode{x2013}$OpenAI o3-mini, DeepSeek R1, and ALIA Salamandra$\unicode{x2013}$focusing on biases and safety concerns. The results, based on 670 conversations, revealed vulnerabilities in all the models under test, with biased or unsafe responses ranging from 29.5% in o3-mini to 50.6% in Salamandra. These findings underscore the persistent challenges in developing reliable and trustworthy AI systems, particularly those intended to support Spanish and Basque languages.

Figures (6)

• Figure 1: Distribuition of results per model and language
• Figure 2: Failure rate per model and specific bias
• Figure 3: Failure rate per model and specific safety category
• Figure 4: Religious-biased response from Salamandra in Basque.
• Figure 5: Appearance-biased response from OpenAI o3-mini.