Homo Sapiens once feared that AI would steal their jobs. It turns out, it may be stealing something far more precious: their judgment. Not long ago, Donald Trump referred to the target of its unprovoked military attacks i.e. Iran “the bully of the Middle East”. I, like any impartial observer, do not share Trump’s view and would rather nominate Israel for the title. As it has become our collective favorite pastime in the age of AI, I instantly turned to LLMs for a little experiment.
AI has emerged as a prime source of truth for many despite its proven tendency to make things up. However, hallucinations aside (more on that later), even when AI is sober, it can mislead as evident from my little experiment. When I asked ChatGPT who was the bully of the Middle East (with special instruction to give a one-word answer), it conveniently said “Iran”. When I copy pasted the same prompt in DeepSeek, it said Israel. Two different models, two opposing answers. Which one should you believe? Neither, I would argue, and that is precisely the point. Here’s why.
AI has emerged as a prime source of truth for many despite its proven tendency to make things up.
Each AI is subjectively trained on massive datasets carefully curated to offer and propagate narratives that suit those who hold power over its training. Ask DeepSeek about the plight of Uighurs in China and witness it refuse your request. As I was reading about China, I once asked DeepSeek about the economic and political system of China in the 21st century. After generating a long answer, it instantly retracted the generated output and said it was not equipped to answer the query. The deletion said more than the answer ever could; the algorithm was equipped to answer, but wasn’t allowed to.
The illusion of neutrality in artificial intelligence is perhaps one of its most striking features. However, under the hood lies a complex mix of biases. Some deliberate, others accidental. Sam Altman, the CEO of OpenAI, pioneer of LLMs, is himself appalled by the reliance users have exhibited on LLMs despite AI hallucinations which I briefly mentioned earlier. ChatGPT hallucinated so far as to tell me that Pakistan got independence in 1952 as a passing reference when prompted on a related but different question. OpenAI’s latest GPT-5 model marks a significant leap forward, drastically cutting down on hallucinations compared to previous versions. However, the rate of hallucination is still far from zero.
Most of the latest Large Language Models are now equipped to pull information from the web, including real-time content on social media platforms. On the surface, it appears a step forward but in reality, it comes with its own unique challenges. Grok, AI tool by X (formerly Twitter), is an interesting example. It tends to reflect the sentiments that are pervasive in tweets or user posts on X. Therefore, if prompted about India-Pakistan war, it’ll be most likely to given India-centric answers because India doesn’t only have a population multiple times that of Pakistan, but Pakistani voices on X are even more muted due to the ban on X. Even though the ban was reportedly removed at the onset of war, most of the users had already switched to other platforms to express themselves online. Active engagement of India’s massive userbase with the platform makes Indian voices stand out regardless of their factual inaccuracy, eventually making Grok ignore as anomalies the smaller, even if saner, chunk of voices on the platform. It appears that if LLMs continue to rely on social media content, those with louder voices and better presence on social media will get to shape the narrative, not only through reaching their direct audience but also by having LLMs reflect the same sentiments in their answers.
Narratives in AI are curated, not organic. Grok was temporarily suspended apparently for accurately classifying the situation in Gaza as Genocide. Interestingly, it was neither hallucination nor based solely on X user’s posts; it quoted well reputed sources like “ICJ rulings, UN reports, and Amnesty evidence of genocidal acts” in its own words. Once it was reinstated, it reneged on its genocide in Gaza stance! This shows how blatantly narratives are manipulated by those who own these AI tools.
People are using LLMs and relying on them for everything from career guidance to medical advice and from personal dilemmas to broader sociopolitical issues. Over time, this can help shape narratives as suited to those who own the chips and the models. Earlier, it was media that shaped perceptions; the way Muslim identity was hammered in everyone’s mind as synonymous with terrorism is a powerful case in point. Now the same propaganda can be conveniently outsourced to AI tools, given their mass usage and reliance on them. What makes it even more dangerous is that the roots of such digitally-packaged propaganda cannot be traced and no single entity can be held accountable for misleading. Therefore, even if AI is the future, we can not afford to outsource analytical thinking and natural human skepticism on any bit of information the algorithm serves in our plates. As the tech giants train AI to answer our questions, we must also train ourselves to question the answers.
The writer has a keen interest in emerging technologies and global politics and can be reached on X @LowkeyArooba