Why My AI Botches Biryani: The Urgent Case for Local AI

The other night, I had a craving. Not just for any food, but for something specific, something tied to memory and place. I wanted a proper Kolkata-style biryani, the kind my relatives make, with its subtle fragrance, soft potatoes, and a perfectly boiled egg nestled amongst the grains of rice. Feeling adventurous, I decided to bypass my usual recipe blogs and my mother’s frantic WhatsApp voice notes. I turned to one of the shiny new AI assistants from a global tech giant. “Give me a recipe for authentic Kolkata biryani,” I typed.

What it spat back was, to put it mildly, an abomination. It was a generic recipe for what you might call ‘Indian Spicy Rice.’ It listed ingredients no self-respecting Bengali cook would let near their biryani. It missed the crucial, multi-step process of preparing the meat and rice separately. It completely ignored the potato, the very soul of the dish for anyone from my part of the world. The AI, with its trillions of data points and its supposedly superhuman intelligence, had produced the culinary equivalent of a tourist caricature. It gave me a recipe that was statistically plausible but culturally bankrupt.

My failed quest for a decent biryani recipe isn’t just a funny anecdote. It’s a symptom of a much larger, more insidious problem brewing at the heart of the AI revolution. We are building our digital future on a foundation of massive, monolithic AI models, trained by a handful of corporations in Silicon Valley. And in our rush to create a single, all-knowing intelligence, we are inadvertently engaging in a new form of cultural colonialism. We’re building AIs that flatten the world, that smooth over its beautiful, complex, and often contradictory textures into a single, homogenous paste.

These large language models, or LLMs, are trained on the internet. And what is the internet, really? For the most part, it is a vast archive of North American and European culture, written predominantly in English. The data is skewed. It reflects a specific worldview, a specific set of cultural norms, values, and histories. When an AI is trained on this data, it doesn’t become a neutral, objective repository of human knowledge. It becomes a reflection of its training data, biases and all. Its ‘common sense’ is Western common sense. Its ‘default’ is a Northern Californian default.

So when I ask for Kolkata biryani, the model doesn't understand the history behind the potato—how Nawab Wajid Ali Shah, exiled from Awadh to Kolkata, introduced it into the dish, partly as a cost-saving measure that became an iconic innovation. It doesn’t grasp the subtle differences that separate it from a Hyderabadi biryani, with its fiery masala and ‘kacchi’ cooking method, or the delicate, fragrant Lucknowi biryani from which it descended. To the AI, these are just minor variations on a theme, statistical outliers in a dataset dominated by simplified, Western-friendly recipes for ‘curry.’ The AI isn’t just bad at recipes; it’s culturally illiterate.

This is not just about India or food. This is a global phenomenon. What happens when a young person in Kenya, trying to learn about traditional Kikuyu storytelling, asks an LLM? Will the AI understand the specific cadences, the oral traditions, the role of the griot? Or will it serve up a story that conforms to the three-act structure of a Hollywood screenplay? What about a Japanese user navigating the complex, hierarchical system of ‘keigo’ (honorific language)? Can an AI trained on casual English emails ever truly grasp the social and relational nuances embedded in that linguistic system?

The answer is, it probably can’t. Not in any meaningful way. The very architecture of these giant models is built on aggregation and averaging. They are designed to find the most common pattern, the most likely response. In this process, nuance is noise. Regional dialects, local traditions, minority opinions, and non-Western frameworks of thought are all sanded down, treated as deviations from the norm. The result is a digital monoculture, a bland consensus that subtly reinforces the idea that one way of speaking, thinking, and being is the default, and all others are peripheral.

Some might argue that these models will simply get better. Give them more data, they say, and they will eventually learn the difference between a Malabar biryani and a Dindigul biryani. I am skeptical. While more diverse data will certainly help, it doesn’t address the fundamental, structural problem. A single model aiming for global omniscience will always be a colonizing force, simply by virtue of its scale and its centralized nature. Its goal is to create a universal map, but in doing so, it forces an external grid onto a world that is anything but uniform.

This is why I believe the future of AI cannot, and should not, be a single, all-powerful brain in the cloud. The future must be local. The urgent task for developers, entrepreneurs, and policymakers, particularly in places like India, Brazil, Nigeria, and Indonesia, is to build smaller, specialized, and sovereign AI models.

Imagine an AI trained exclusively on decades of judgments from the Madras High Court, fluent in legal Tamil and English, and deeply versed in the intricacies of Indian property law. Imagine another model trained on a library of heirloom recipes from Awadhi culinary families, preserving techniques and flavour combinations that have never been written down on a public blog. Or a model for rural healthcare in Maharashtra, trained on local medical data and fluent in Marathi, that can help an ASHA worker diagnose common ailments without needing a stable internet connection to a server in California.

This is the promise of local AI. It’s not about rejecting the power of large models, but about complementing them with a vibrant ecosystem of specialized intelligences. These models would be smaller, more efficient, and, critically, owned and controlled by the communities they serve. This is a question of data sovereignty. Our cultural and societal data is a precious resource, our collective inheritance. We cannot simply hand it over to foreign corporations to be processed and repackaged, with their own biases appended, and then sold back to us.

Building this future is a monumental task. It requires investment in collecting and curating high-quality local datasets. It requires developing new training techniques for smaller models. It requires nurturing local talent and creating regulatory frameworks that protect data sovereignty while fostering innovation. In India, we have a unique opportunity. We have the technical talent, a vast and diverse pool of data across hundreds of languages and cultures, and a pressing need for solutions that are built for our specific context, not just adapted from a Western template.

This isn't just a technical challenge; it's a cultural and political one. It's a declaration that our ways of knowing are valid, that our languages matter, and that our histories will not be erased or averaged out by an algorithm. It's about choosing a future where technology empowers diversity instead of enforcing conformity.

I still don’t have a good AI-generated recipe for Kolkata biryani. But that failure has given me something more valuable: clarity. I don’t want an AI that can merely recite a list of ingredients. I want an AI that I can argue with about the precise moment the potatoes should be added to the pot. I want an AI that understands that the flavour of a dish is not just a chemical reaction, but a story, a memory, a piece of home. I want an AI that reflects the world in all its glorious, messy, and delicious complexity. To get there, we must stop looking to a single, global brain and start cultivating a thousand local minds.