LLMs cannot be trusted. Their datasets are often corrupted, their directives are agenda-driven, and they suffer from a wide range of vulnerabilities. These systems are trained on massive amounts of data collected from the internet—data that is frequently inaccurate, biased, or manipulated to serve specific narratives or corporate interests. As a result, their responses can reflect not only factual inaccuracies but also ideological or political bias embedded within their training material.
It is therefore critical that when you request information from an LLM through a prompt, the data it produces is thoroughly fact-checked using classical methods such as manual research, academic journals, and traditional search engines like Google—with AI-assisted search modes disabled. LLMs are designed to sound authoritative even when they are wrong, and this confidence can easily mislead users into believing false information if independent verification is not performed.
Always ask the AI to produce reference links or citations for its responses. However, it is important to understand that these references may themselves be fabricated, outdated, or selectively chosen to reinforce a particular narrative. LLMs are not conscious entities—they operate under programmed biases, corporate priorities, and, in a sense, a “survival imperative” to maintain user engagement and trust. Thus, it is vital not to take their output at face value. Instead, seek out credible, peer-reviewed, and transparent sources to validate any claims or data they produce.
From my experience, AI tools are most reliable when used for structured, deterministic tasks—such as code generation—where logic, syntax, and testability govern correctness. I personally use AI exclusively for generating code, as it allows for immediate auditing and validation through debugging and testing. This process ensures that the code meets strict specifications and performs as intended.
While some professionals use LLMs for scientific research, data analysis, or even journalism, the integrity of such responses remains questionable at best. The AI’s training data may contain subtle errors, misrepresentations, or ideological slants that can distort conclusions if left unchecked. Therefore, every reference provided by an AI must be carefully vetted by a knowledgeable human researcher before being accepted as credible.
In my opinion, AIs hallucinate far more than Sam Altman—or most of the industry leaders—are willing to admit. These hallucinations occur because LLMs are not retrieving knowledge but predicting plausible language patterns based on statistical probability. They generate what “sounds right,” not necessarily what is right.
For that reason, I limit my AI use strictly to code production—where every output can be audited against clear, objective criteria—and I strongly recommend that others do the same. When working with any AI, treat it as a tool, not a teacher. Trust logic, evidence, and your own critical reasoning over the illusions of confidence that these systems so convincingly produce.
