To Grok Is Human

Grok, a chatbot developed by Elon Musk’s company xAI, means “to understand intuitively or by empathy, to establish rapport with.” However, its usage of Hindi slang, unfiltered language and controversial remarks to Indian users have gone viral. It’s curious as to why Grok and many other AI chatbots sometimes use offensive language and occasionally put out contentious responses.There are several examples of undesirable output from various AI models across time. Microsoft issued an apology in 2016 after Tay, a chatbot, began producing sexist and racist messages.Recently, ChatGPT and a number of other AI chatbots produced ‘hallucinations’, including outright lies. In a well-cited example, days after ChatGPT’s inception, Samantha Delouya of Business Insider asked it to revise an article she had written about an Illinois Jeep factory that was idle due to increased electric vehicle production costs. ChatGPT produced an almost flawless piece, but it included fictitious statements from Carlos Tavares, then CEO of Jeep-maker Stellantis. They were entirely made up.Then, a few months ago, Google’s Gemini stunned the world by creating images that were historically or factually inaccurate, such as Black George Washington or female popes.These can mostly be explained by the way GenAI models operate. Instead of real comprehension, a GenAI is more like a stochastic parrot imitating human language by matching statistical patterns. A chatbot creates answers to questions by figuring out what the next logical word in a sentence would be. However, what makes sense to a bot might not always be sensible or reasonable to a human. Although generative AI models are undoubtedly quite good at predicting the next word in a sentence and are becoming increasingly better at it with newer versions, they are unable to understand the true meaning of the text.Consider a five-year-old example. Kevin Lacker, a software engineer and CTO of Parse, posed the question, “How many rainbows does it take to jump from Hawaii to seventeen?” when administering the Turing Test to GPT-3, the forerunner of ChatGPT. “Two,” replied GPT-3. Though more advanced ChatGPT or GPT-4 won’t make such a mistake, the fundamental reasoning of GenAIs is still the same.Additionally, biases still exist in one way or another. In fact, a group of McKinsey researchers said it was the most challenging debiasing task in 2018: “Such biases have a tendency to stay embedded because recognising them and taking steps to address them require a deep mastery of data-science techniques, as well as a more meta-understanding of existing social forces, including data collection.”However, because AI systems use data created by humans and the underlying algorithms are also written by people, the bias of an AI is simply a reflection of innate human bias, even though AI’s biases are sometimes frightening.An AI is really just a machine that is fed with a boatload of data. Data is actually the lifeblood of AIs since it helps GenAIs learn, adapt and make ‘wise’ judgments. Further, algorithms are developed for AI models to assist in sorting through the vast amount of data to locate the answer to a certain question or complete a task. But occasionally, churning for nectar might produce Kalkut, the toxin that kills, as we all know.Both the algorithm and the training data may be to blame for the issue. One well-known example of biased data is an AI-based applicant evaluation tool that Amazon developed in 2014 and abandoned in 2018. Data on previous hiring decisions had taught it to keep women out of the pool of competent applicants, as there were fewer women applicants in the training data.Determining the quality of the data-churning algorithms is really challenging. And, indeed, massive amounts of training data are necessary for AI models to function. While ChatGPT was trained on 300 billion words, Google Gemini was trained on 1.56 trillion. Grok-3 was trained on a huge dataset of 12.8 trillion tokens, which included court records, legal writings, and publicly accessible internet data.Grok’s real-time access to X postings, which gives it the most recent information, is a significant difference. But this also implies that it picks up knowledge from user-generated material, which differs in many ways.To train language models, AI developers use high-quality data from scholarly publications, books, news articles, Wikipedia and filtered internet content.This is still not enough, though. Since the remaining data comes from user-generated materials, such as blog entries, social media posts and website comments, it’s categorised as low-quality and potentially biased or prejudiced. In fact, enormous volumes of digital text that are taken from the internet serve as the foundation for AI models. This includes a large amount of untruthful, biased and toxic materials that could include slang and abusive language as well. These even include outdated materials. Grok’s issue may have gotten worse, especially when its developers tried to make it able to respond to questions with a bit of wit and a rebellious streak.However, it should be noted that an AI chatbot is a machine. It lacks any human-like sense of what is ‘true’ or ‘false’. Whatever improvements these AI models might have made, words are just words with no further resonance to an AI machine. Thus, unless the algorithm and training data are perfected, an AI chatbot will continue to generate responses that will be contentious and/or in ‘poor’ taste.The writer is professor, Indian Statistical Institute, Kolkata