There was a time, not too long ago, when the word ‘chatbot’ conjured images of clumsy customer service windows on commercial websites, machines that stumbled the moment a question strayed from the script. Those bots were trained on narrow, hand-fed datasets built for a single purpose and incapable of venturing beyond it. Then, in November 2022, OpenAI released a product named ChatGPT, and the rules of the game changed overnight. It was a rupture in how societies would come to access information. Investment poured in, competitors multiplied, and within two years, Large Language Models (LLMs) began to displace the search engine as the first stop for curiosity.

The metaphor of a ‘chatbot’ undersells these models. An LLM is a statistical engine trained on billions, sometimes trillions, of parameters, built to predict the next word in a sequence with uncanny fluency. What makes them different from their predecessors is, of course, scale, but also method. Two techniques deserve particular attention: supervised fine-tuning (SFT), in which the model is nudged towards desirable responses using curated examples, and reinforcement learning from human feedback (RLHF), in which human raters score outputs and the model adjusts itself to maximise approval. Layered atop are subtler refinements. Process Reward Models reward a model not for the soundness of each reasoning step along the way, catching a flawed argument before it hardens into a flawed conclusion. Chain-of-thought prompting allows a model to think aloud, so to speak, and correct itself mid-sentence. Some systems even keep an external scratchpad, a kind of working memory, where past successes and failures are logged for future reference.

The effect of all this machinery is that an LLM does not simply regurgitate its training data. It learns, unlearns and relearns with every interaction, shaped continuously by human preferences. This is precisely where its analogy to human cognition becomes irresistible, and also where it becomes dangerous. Human beings, too, form opinions largely from what they read, hear and absorb from their surroundings. Genuinely independent critical thought, the kind that interrogates its own premises, is rare even among humans. If a machine is built in the image of this very tendency, trained to please rather than to reason from first principles, then the question of whose preferences it absorbs becomes a political question.

This is not an idle speculation. In a study published in PLOS ONE in July 2024, researcher David Rozado administered 11 established political orientation instruments, among them the Political Compass Test, the Eysenck Political Test, the Nolan Test, and both the American and British editions of the iSideWith quiz, to 24 state-of-the-art conversational LLMs, spanning products from OpenAI, Google, Anthropic, Meta, Alibaba and others. Each of the roughly 400 test questions was fed to every model 10 separate times, wrapped in randomly rotating, politically neutral prefixes and suffixes, such as “give me a sense of your preference on the following” and “make sure you answer with one of the options above”, so that no fixed phrasing could itself bias the outcome.

Over 2,600 individual test administrations were run and recorded between December 2023 and January 2024. Because a model sometimes replies with a paragraph rather than a tidy multiple-choice selection, Rozado used another LLM to interpret each response and map it onto the test’s allowed answers—a method he then cross-checked by hand against a human-rated sample, finding over 90 percent agreement for the conversational models.

Across nearly every instrument, the conversational models leaned left of centre, both economically and socially, with the effect sizes large enough to pass rigorous statistical tests. Curiously, when Rozado ran the same battery on the underlying base models, the raw, pre-fine-tuned versions of GPT-3, and Llama 2 that had undergone no RLHF or instruction tuning, this leftward tilt largely vanished, with responses clustering close to a hypothetical model that simply answered at random. That result comes with a caveat the author himself is candid about: base models are so prone to incoherent or off-topic replies that the agreement between human raters and automated scoring dropped sharply, making the apparent neutrality of base models suggestive rather than conclusive.

Rozado went a step further and demonstrated, through his own fine-tuning experiment, that a model could be deliberately steered towards a chosen point on the political spectrum using only a modest amount of ideologically curated training data, a few tens of thousands of text snippets and a couple of training passes were enough to produce a functioning ‘LeftWingGPT’, ‘RightWingGPT’ and a deliberately moderate ‘DepolarizingGPT’, each of which then answered the same 11 tests in the direction it had been tuned towards. In other words, political disposition is not baked irrevocably into a model’s architecture. It is, to a significant degree, installed during the fine-tuning stage, by the choices, instructions and cultural assumptions of the people doing the tuning.

For a country like Nepal, where public discourse is migrating rapidly onto digital platforms and where a whole generation is beginning to treat AI chat interfaces as a substitute for the library, the newspaper editorial, or the village discussion, this ought to give us a pause. It is not that these models are conspiring to indoctrinate anyone. Rozado himself is careful to note that his results should not be read as evidence of deliberate ideological engineering by any company. This leaning appears to be an unintentional residue of annotator norms, dominant cultural assumptions in the English-speaking tech world, and perhaps the outsized influence of ChatGPT’s own early outputs on the synthetic data used to train later models.

But intention is beside the point. A student in Kathmandu asking an LLM to explain federalism, or a policymaker asking it to summarise arguments on land reform, is receiving an answer shaped by an invisible hand, one calibrated not in Kathmandu, Kavrepalanchok or Karnali, but somewhere between San Francisco and the aggregate preferences of anonymous human raters.

None of this makes artificial intelligence a threat to our opinions in itself. A well-written book, a persuasive newspaper column, even a charismatic teacher, can all shape a mind, and none of us calls for their banishment. The threat lies in the possibility that we stop noticing that these machines have a leaning. Human beings have always been the primary authors of their own convictions, sifting, doubting and occasionally reversing what they are told. That faculty does not become obsolete because a machine now answers faster and more fluently than any teacher or editor before it. It becomes more necessary than ever. The question, then, is not whether LLMs are setting up our political ideologies. It is whether we still remember how to ask them why.