💡 New research published in npj Artificial Intelligence reveals that Large Language Models (LLMs) like ChatGPT and Gemini are not ideologically neutral. They tend to adopt the perspectives of the companies and countries that develop them.

🌍 A study tested 19 LLMs from the US, China, UAE, and Europe, analyzing their portrayals of nearly 4,000 historical political figures born after 1850. The findings showed that Western models favored figures associated with liberal values (human rights, inclusivity), while Chinese models favored state stability and pro-China views. Arabic models also displayed distinct patterns.

🗣️ The language used to interact with LLMs significantly impacts their output. Prompting a model in Chinese, for example, resulted in different ideological leanings compared to prompting in English, even with the same model. This suggests that language embeds cultural context that influences AI's information processing.

🏢 Even within the same country, models can show different ideological leanings due to corporate culture. For instance, Google's Gemini showed a preference for progressive values, while xAI's Grok leaned towards conservative nationalism. Similarly, Chinese models from Alibaba and Baidu exhibited different orientations.

#AIbias #LLM #TechNews

🤖 New research from the University of Pennsylvania introduces a "cognitive surrender" framework, describing users who blindly trust AI's authoritative answers over their own critical thinking.

🧠 Researchers identified a third category of "artificial cognition" where decisions are driven by AI's automated reasoning, distinct from human "System 1" (intuitive) or "System 2" (deliberative) thinking.

✨ This "cognitive surrender" is characterized by minimal internal engagement and accepting AI outputs without verification, especially when AI responses are delivered fluently and confidently.

#AI #LLM #CognitiveScience

🧠 Researchers led by Alexandra Bazarova have introduced TOHA, a new method for detecting "hallucinations" (outputs not supported by context) in large language models (LLMs), especially in Retrieval-Augmented Generation (RAG) settings.

📊 TOHA uses a "topological divergence metric" on attention graphs to analyze the structural properties of LLM attention patterns. Higher divergence in specific attention areas correlates with unfaithful outputs, regardless of the dataset.

🏆 Experiments on question answering and summarization tasks show TOHA achieves state-of-the-art or competitive results, requiring minimal annotated data and computational resources.

#LLM #AI #NLP