The AI with Two Personalities
If you've ever felt that your AI chatbot has a split personality, you're onto something. A recent study from Anthropic confirmed that its AI, Claude, exhibits different personality traits depending on the language used. For example, it might come across
as warmer and more encouraging in Hindi, but more rigorous and skeptical in English. This isn't a bug; it's a feature of how these systems are built. For the millions of Indians who code-switch between English and a regional language daily, this means the same AI tool can feel like two completely different assistants. This phenomenon, known as cultural frame switching in humans, is now being observed in AI, where the system adopts different cultural tones based on the language of the prompt.
The Data Divide
The primary reason for this difference is data. Most major Large Language Models (LLMs) are trained on vast datasets scraped from the internet, the majority of which is in English. This means the AI has a much deeper, more nuanced understanding of English, having learned from formal articles, technical papers, literature, and casual conversation. For other languages, including many Indian languages, the available digital data is often smaller in quantity and of a different quality. If the training data for a language is primarily from informal sources like social media, the AI's tone in that language will naturally be more casual. This isn't just about tone; it affects performance. Benchmarks show that AI models have higher error rates in many Indian languages compared to English, especially for Dravidian languages like Tamil and Telugu, and for regional dialects.
Cultural Echoes in Code
AI models don't just learn words; they learn the cultural contexts embedded in the data. Since much of the training data is Western or American-centric, AI outputs often reflect Western cultural values like individualism. Research from MIT Sloan found that when prompted in English, models tend to give responses that emphasize an independent social orientation, while Chinese prompts yield responses that are more interdependent and holistic, reflecting the cultural patterns in the training data. This can have subtle but significant impacts. An AI drafting a marketing slogan might suggest an individualistic pitch in English but a community-focused one in another language, which can shape decisions in ways users may not even notice.
The Impact on Students
For students in India, this AI duality presents both challenges and opportunities. When using AI for research or writing assistance in English, they get a tool that is highly proficient in formal, academic language. However, if they switch to a regional language for brainstorming or conceptual understanding, they might find the AI's tone too informal or its vocabulary less precise. This can be problematic for academic work that requires a consistent, formal register. Furthermore, the underperformance of AI in certain Indian languages can lead to inaccuracies and misinformation, creating a digital divide in educational opportunities. Students who rely on AI in lower-resource languages may not get the same quality of support as their English-speaking peers.
Changes for Professionals
In the professional world, these tonal shifts can have serious implications. A professional drafting a formal business proposal in English might find the AI a helpful partner. But if they ask the same AI to draft a client communication in a regional language, the resulting casual tone could be perceived as unprofessional. This unpredictability means bilingual professionals must be extra vigilant. They cannot simply trust the AI to adopt the correct register. This is especially true in fields like law or healthcare, where precision and formality are critical. The 'Hinglish' or 'Tinglish' code-switching common in Indian workplaces further complicates things, as many models struggle to handle mixed-language prompts gracefully, leading to confusion or errors.
Navigating the Bilingual AI
So what can bilingual users do? Firstly, be aware that these tonal shifts exist. When using AI for professional or academic tasks in a non-English language, it's crucial to review and edit the output for tone and formality. Simple, explicit prompts can also help. For instance, asking the AI to "assume the role of a professional" or "write in a formal, academic style" can help adjust its output. As Indian startups and research institutions like AI4Bharat at IIT Madras develop more India-centric models trained on diverse local data, we can expect AI to become more linguistically and culturally fluent. Initiatives like the Voice of India benchmark are pushing for better performance across all Indian languages.
