Only Grok can judge you. It’s scary, and not so smart.
Summary
- Where we stand on matters of history and ideology is increasingly judged by what we post and share online. Now, AI tools of social media platforms are complicating an already delicate situation.
We may want to live by the rather Biblical adage of “only God can judge me", but the truth is that our online persona—curated in the opinions we write online, the posts we reshare, and the people and accounts we follow on social media—is shaping how we are publicly perceived.
When controversy breaks out online or in the real world, especially relating to sensitive matters such as religion or politics, we are quick to judge others based on what they post or reshare on social media. Now, social media’s new-fangled AI tools are making these ideological bubbles much worse.
The past two weeks, Indian users of Elon Musk’s X have been playing around with the social media platform’s artificial intelligence chatbot Grok, prompting it to answer uncomfortable questions about controversial topics in history and contemporary politics, gauging its opinion on the country’s biggest leaders, and even persuading it to respond to taunts and questions in Hindi with aggressive North Indian slang peppered with the choicest abuse.
Messing around with AI chatbots to test their limits isn’t new. When OpenAI’s ChatGPT model was first released to the public in November 2022, early adopters tested its guardrails that prevented it from sharing abusive or sexually explicit information. Grok has been around since November 2023, and its latest version, Grok 3, was launched in February, with Musk declaring it to be “scary smart".
Grok’s users may have taken Musk’s words too literally.
Also read | Online paradox: In need of a digital detox? AI assistants could be of help
Hey, Grok, who…
Users on X are prompting the bot to draw up lists of accounts based on their ideological bent, apart from asking difficult questions. Grok’s answers are going viral. For example, this question to Grok, asking it to name the “most communal politician of India", received more than 700,000 views as on 20 March. Asked “who personally spreads the most misinformation on X", Grok named its owner Elon Musk. This post received 1.4 million views.
Opinions and snap judgements are the lifeblood of all social media, and X in particular thrives on it. In 2022, shortly before he bought Twitter (now X), Musk described the social media platform as the “de facto public town square" where complex questions about politics and society are often simplified into majority opinions.
Chatter on X has indeed become crucial to shaping public opinion on routine things as well, from ‘bot’ accounts that promote movie stars and their projects to meme accounts engaging in ‘stealth marketing’.
Naturally, it wasn’t long before X users turned to Grok to get its opinions too and brandishing its answers to buttress their own ideological point of view. For instance, in this post, Grok lists India-based X accounts that it says spread “fake news" in India. Most of these accounts tend to lean to the right and post pro-government views. Soon enough, Grok was prompted to generate a list of left-wing accounts as well.
Also read | Why Elon Musk’s Grok is the internet’s latest fad
Scary, not smart
There’s no good end to this. Grok is an AI model, and just like its peers it is trained on publicly available data and can merely analyse and reproduce the average of public opinions and thoughts it was trained on.
What’s more, any carefully crafted prompt with detailed instructions and caveats can prompt even a “scary smart" AI engine to produce a desired answer. Using Grok’s responses to such questions as ‘proof’ of some objective truth can have real-world consequences.
Consider this post, where comedian Abhishek Upmanyu found himself on a Grok-generated list of X users who allegedly “continuously post anti-Hindu propaganda". In another post, social media-famous forest service officer Parveen Kaswan found Grok listing him as a “top twitter handle for UPSC", the centrally administered exam for recruitment to India’s civil services. Kaswan says he has never posted anything remotely related to the exam or its latest syllabus.
It’s easy enough for our social media followers or even strangers to judge us for what they can see us posting online. But AI models like Grok can quickly draw conclusions about our beliefs based on years and years of our posts online, often with little or no context.
Already, some people are taking such AI bot-generated conclusions seriously, although for most Grok’s answers are little more than an amusement or perhaps fodder for an online fight. How soon before its answers get real people in real-world trouble?
topics
