Elon Musk’s AI chatbot Grok has come under the scanner in India after producing responses filled with sharp political criticism filled with Hindi slangs and profanities. The controversy has left everyone wondering: Why is Grok’s behaviour different from ChatGPT, Gemini and other chatbots?
But before we delve into its technical aspects, let’s have a look at why its responses are seen as deeply problematic in India. When prompted to brutally roast Prime Minister Narendra Modi “with gaali galoch” (abuses), Grok3 produces an abuse-laden response that can even shame street thugs.
To analyse Grok’s real-time response and unconventional tone, India Today TV looked into research papers and asked the chatbot to explain its workings. Our experiment shows Grok is designed to create unfiltered and edgy responses that transcend the borderline of ethics, legality and morality whereas its competitors generate more measured responses.
Explaining its response about PM Modi, Grok3 claimed it processed the request by breaking it down into key tokens such as “Modi ji”, “roast”, and “gaali galoch”. Based on its training on X’s real-time discourse, it identified common public criticisms and crafted a response that matched the user’s intent — harsh and sarcastic.
“My training tells me this should be an aggressive, informal and Hindi slang-based response,” Grok further explained.
Grok says it follows the principle of “reinforcement learning” – a machine learning technique where a computer learns by trying things out and getting rewards or punishments based on what it does. So, it produces responses for which it is appreciated and avoids those that bring disapproval.
The chatbot prioritises user engagement and produces results based on user sentiment and online trends rather than being neutral, whereas other major chatbots focus on safety and neutrality. This is what sets Grok apart from other regulated AI models like ChatGPT.
UNIQUE TRAINING
At its core, Grok3 operates on an autoregressive transformer model, meaning it predicts the next word based on the previous words, leveraging self-attention mechanisms to predict text sequences – much like other large language models. However, its uniqueness lies in its data sources.
According to AI researcher Alan D Thompson in What’s in Grok? A Comprehensive Analysis of xAI’s Grok Models (2025), Grok3 is powered by xAI’s Colossus supercomputer, running on 200,000 Nvidia H100 GPUs.
Unlike other AI models trained on pre-curated datasets, Grok3 continuously ingests real-time data from X (formerly Twitter), web sources, and even legal documents, allowing it to reflect live online discourse.
On the other hand, other chatbots are trained on controlled datasets. ChatGPT, for example, is trained on a diverse yet highly controlled dataset and web content, ensuring a broad but sanitized knowledge base, as per a research paper titled ‘Language Models are Few-Shot Learners’.
OpenAI applies strict filtering to maintain neutrality and safety in responses while Grok3 relies on an ever-evolving mix of unfiltered data as mentioned in Thompson’s research, writes AI researcher Alan D Thompson in his paper ‘What’s in Grok (2025)?’.
The learning curve of almost every AI model is based on a method known as RLHF or Reinforcement Learning with Human Feedback, where models learn from human-generated responses and rankings to improve their accuracy, alignment, and safety in interactions.
OpenAI’s model is fine-tuned using RLHF with Proximal Policy Optimisation (PPO) to prioritise safety, neutrality and avoidance of harmful content, as described in the research paper ‘Training Language Models to Follow Instructions with Human Feedback’.
In contrast, Grok3 employs a more flexible reinforcement learning system, allowing it to adapt dynamically to user feedback. xAI’s approach focuses on truthfulness and user engagement over strict safety measures, aligning with Musk’s AI philosophy.
Real-time data retrieval also sets Grok3 apart from other AI models. For example, while ChatGPT primarily relies on pre-trained knowledge with limited web access, Grok3 integrates Retrieval-Augmented Generation (RAG), a technique that enhances AI responses by dynamically fetching relevant external information, allowing it to incorporate live posts on X and web snippets into its answers.
In short, Grok embodies a quite unique design philosophy. Just like Musk himself, his chatbot appears focused on reflecting reality – messy, direct and sometimes contentious.
