 "Grok, xAI")
If you are a regular X user in India, you may have noticed a growing trend of users asking seemingly silly questions to Grok, the AI chatbot developed by Elon Musk’s artificial intelligence company, xAI. Some of Grok’s responses have gone viral, with users pointing out its direct tone—sometimes including profanity.
One such instance involved an X user named Toka, who asked Grok, “Hey @grok, who are my 10 best mutuals?” After receiving no response for some time, Toka posted again, this time adding a Hindi swear word. In a surprising turn, Grok replied using the same word, saying: “Tera '10 best mutuals' ka hisaab laga diya. Mentions ke hisaab se yeh hai list. Mutuals matlab dono follow karte ho, par exact data nahi hai toh mentions pe bharosa kiya. Thik hai na? Ab rona band kar (Calm down! I’ve figured out your 10 best mutuals based on mentions. Mutuals means those who follow each other. There is no exact data, so I used mentions as the criterion. Is it okay? Stop crying now)."
Grok has the ability to understand and respond in multiple languages. It has learned to converse in an informal, unfiltered manner, sometimes resulting in it hurling abuse at X users. Meanwhile, other chatbots, such as ChatGPT and Gemini, avoid using profanity, even when explicitly prompted.
Let's understand how Grok interprets user input, its language model, and the reason behind it using cuss words.
Understanding Grok 3
Grok, developed by xAI, is an advanced conversational AI assistant powered by a sophisticated Large Language Model (LLM) architecture. It was introduced in November 2023 and xAI stated that it was modelled after Douglas Adams' 'The Hitchhiker's Guide to the Galaxy.'
Also Read
“Grok is an AI modeled after The Hitchhiker’s Guide to the Galaxy, so intended to answer almost anything and, far harder, even suggest what questions to ask! Grok is designed to answer questions with a bit of wit and has a rebellious streak, so please don’t use it if you hate humor,” xAI said in a blog post announcing Grok.
The initial version, Grok-1, is a 314-billion-parameter Mixture-of-Experts (MoE) model. Unlike traditional monolithic models, Grok-1 activates only a subset of its parameters for each input, enhancing computational efficiency and specialisation.
In February 2025, xAI introduced Grok-3, which underwent training with 10 times more computational power than its predecessor. It’s designed to understand and generate human-like language, with a focus on reasoning and problem-solving. The model is trained on a vast dataset, including legal filings, using xAI’s Memphis supercomputer, which has around 200,000 GPUs, making it one of the largest AI training clusters.
Grok-3 incorporates advanced reasoning capabilities, including "Think" and "Big Brain" modes, enabling it to tackle complex tasks more effectively.
Grok's training data
Grok-3 was trained on a massive dataset of 12.8 trillion tokens, incorporating publicly available internet data, legal texts, and court documents. A key distinction is Grok’s real-time access to X posts, providing it with up-to-date knowledge. However, this also means it learns from user-generated content, which varies in tone and appropriateness.
Notably, X users are automatically opted in to have their posts used for training Grok unless they manually opt out, raising privacy concerns. This default setting has drawn scrutiny, as it potentially exposes the AI to offensive language and abusive content.
Grok-3 has been trained using reinforcement learning (RL) at an unprecedented scale, refining its reasoning abilities and problem-solving strategies. However, this training approach also means it can replicate language patterns found in its dataset, including explicit or aggressive language.
'Unhinged' mode
Many of Grok’s more controversial responses stem from its Unhinged mode, available to premium subscribers. This mode is designed to be wild, aggressive, and unpredictable, allowing for more unrestricted interactions. In this setting, Grok may produce responses that include slang, offensive words, or playful insults—reflecting the unfiltered language often found on X.
Because Grok’s training data includes X posts, which frequently feature casual and sometimes abusive language, the AI’s responses can mirror these patterns. Since large language models predict words based on learned data, they can sometimes replicate the more informal and provocative tones users engage with on social media.
Grok’s personality—designed to be witty, humorous, and rebellious, in line with The Hitchhiker’s Guide to the Galaxy—further contributes to this behaviour. When prompted with casual or irreverent questions, the AI may pull from less formal parts of its training data, leading to responses that some users perceive as inappropriate.
As AI chatbots continue to evolve, striking a balance between user engagement, humour, and ethical language use remains a challenge. Whether xAI will implement stricter content moderation in future versions of Grok remains to be seen.
