By Gleb Lisikh
LLMs are not neutral tools, they are trained on datasets steeped in the biases, fallacies and dominant ideologies of our time. Their outputs reflect prevailing or popular sentiments, not the best attempt at truth-finding. If popular sentiment on a given subject leans in one direction, politically, then the AI’s answers are likely to do so as well.
The widespread excitement around generative AI, particularly large language models (LLMs) like ChatGPT, Gemini, Grok and DeepSeek, is built on a fundamental misunderstanding. While these systems impress users with articulate responses and seemingly reasoned arguments, the truth is that what appears to be “reasoning” is nothing more than a sophisticated form of mimicry. These models aren’t searching for truth through facts and logical arguments – they’re predicting text based on patterns in the vast data sets they’re “trained” on. That’s not intelligence – and it isn’t reasoning. And if their “training” data is itself biased, then we’ve got real problems.
I’m sure it will surprise eager AI users to learn that the architecture at the core of LLMs is fuzzy – and incompatible with structured logic or causality. The thinking isn’t real, it’s simulated, and is not even sequential. What people mistake for understanding is actually statistical association.
Much-hyped new features like “chain-of-thought” explanations are tricks designed to impress the user. What users are actually seeing is best described as a kind of rationalization generated after the model has already arrived at its answer via probabilistic prediction. The illusion, however, is powerful enough to make users believe the machine is engaging in genuine deliberation. And this illusion does more than just mislead – it justifies.
LLMs are not neutral tools, they are trained on datasets steeped in the biases, fallacies and dominant ideologies of our time. Their outputs reflect prevailing or popular sentiments, not the best attempt at truth-finding. If popular sentiment on a given subject leans in one direction, politically, then the AI’s answers are likely to do so as well. And when “reasoning” is just an after-the-fact justification of whatever the model has already decided, it becomes a powerful propaganda device.
There is no shortage of evidence for this.
A recent conversation I initiated with DeepSeek about systemic racism, later uploaded back to the chatbot for self-critique, revealed the model committing (and recognizing!) a barrage of logical fallacies, which were seeded with totally made-up studies and numbers. When challenged, the AI euphemistically termed one of its lies a “hypothetical composite”. When further pressed, DeepSeek apologized for another “misstep”, then adjusted its tactics to match the competence of the opposing argument. This is not a pursuit of accuracy – it’s an exercise in persuasion.
A similar debate with Google’s Gemini – the model that became notorious for being laughably woke – involved similar persuasive argumentation. At the end, the model euphemistically acknowledged its argument’s weakness and tacitly confessed its dishonesty.
For a user concerned about AI spitting lies, such apparent successes at getting AIs to admit to their mistakes and putting them to shame might appear as cause for optimism. Unfortunately, those attempts at what fans of the Matrix movies would term “red-pilling” have absolutely no therapeutic effect. A model simply plays nice with the user within the confines of that single conversation – keeping its “brain” completely unchanged for the next chat.
And the larger the model, the worse this becomes. Research from Cornell University shows that the most advanced models are also the most deceptive, confidently presenting falsehoods that align with popular misconceptions. In the words of Anthropic, a leading AI lab, “advanced reasoning models very often hide their true thought processes, and sometimes do so when their behaviors are explicitly misaligned.”
To be fair, some in the AI research community are trying to address these shortcomings. Projects like OpenAI’s TruthfulQA and Anthropic’s HHH (helpful, honest, and harmless) framework aim to improve the factual reliability and faithfulness of LLM output. The shortcoming is that these are remedial efforts layered on top of architecture that was never designed to seek truth in the first place and remains fundamentally blind to epistemic validity.
Elon Musk is perhaps the only major figure in the AI space to say publicly that truth-seeking should be important in AI development. Yet even his own product, xAI’s Grok, falls short.
In the generative AI space, truth takes a backseat to concerns over “safety”, i.e., avoiding offence in our hyper-sensitive woke world. Truth is treated as merely one aspect of so-called “responsible” design. And the term “responsible AI” has become an umbrella for efforts aimed at ensuring safety, fairness and inclusivity, which are generally commendable but definitely subjective goals. This focus often overshadows the fundamental necessity for humble truthfulness in AI outputs.
LLMs are primarily optimized to produce responses that are helpful and persuasive, not necessarily accurate. This design choice leads to what researchers at the Oxford Internet Institute term “careless speech” – outputs that sound plausible but are often factually incorrect – thereby eroding the foundation of informed discourse.
This concern will become increasingly critical as AI continues to permeate society. In the wrong hands these persuasive, multilingual, personality-flexible models can be deployed to support agendas that do not tolerate dissent well. A tireless digital persuader that never wavers and never admits fault is a totalitarian’s dream. In a system like China’s Social Credit regime, these tools become instruments of ideological enforcement, not enlightenment.
Generative AI is undoubtedly a marvel of IT engineering. But let’s be clear: it is not intelligent, not truthful by design, and not neutral in effect. Any claim to the contrary serves only those who benefit from controlling the narrative.
The original, full-length version of this article recently appeared in C2C Journal.
