My initial instinct upon encountering the topic was to ask myself: have I truly considered engineering hallucinations in Large Language Models, or have I only viewed the phenomenon through basic, factual definitions? The dominant narrative across the internet remains completely one-sided, treating hallucination as a simple technical flaw rather than an architectural reality.
While an LLM's statistical decoding certainly produces unprompted errors at the output level, treating the machine as an independent agent misses the broader architectural point. In truth, artificial intelligence is merely a reflection of the people who built and guided it. We operate on implicit assumptions, and the model manifests those gaps as structural failure. We built the machine, yet we expect perfection—without defining in what aspect, precisely. If computing remains fundamentally "garbage in, garbage out," then incorrectly programming the order of operations will naturally yield broken results. AI is strictly an assistant designed to execute our thoughts at scale, not to think on our behalf. When you conceive an idea with precision, articulate it clearly, and verify the output, the assistant becomes an extraordinary multiplier. The machine may generate statistical possibilities, but human agency determines what becomes architecture, knowledge, and action.
This exposes a fascinating paradox: what happens when we ask an AI how to improve itself or eliminate hallucinations? Did a developer program the assistant to output that specific self-referential answer, or is it simply retrieving human discourse on the topic? That alone requires deep reflection. In fact, if you have never won an argument with an AI, then you need to think deeper. When stress-testing these exact concepts across different models—pitting ChatGPT, Claude, and Gemini against one another to critique this essay—the models initially pushed back on my definitions of hallucination and autonomy. But by acting as the system-level steer, evaluating their failure modes, rejecting flawed critiques, and forcing mechanical clarity, the workflow proved its own thesis: AI generates possibilities, but the human retains absolute agency over the system.
Yet the technology remains profoundly useful. AI represents the logical continuation of the quest for automation that began with the very first line of executable code. From mechanical looms to assembly-line compilers, to voice-to-text, and now to LLMs, every leap in technology has simply yielded faster translators between human thought and physical execution.
From picking up a pen to write, to striking keys on a keyboard, to speech-to-text transcription, text-to-speech synthesis, and—do I even mention AR (Augmented Reality)?—we move faster now, but speed is not cognition. The bottleneck has never been the machine’s capacity to compute; it has always been the depth, clarity, and rigor of the human mind steering it. Until a human mind can conceive an idea, it is impossible for a machine to create it. True autonomy requires a soul—a non-physical capacity to originate, intend, and perceive—and because we cannot touch or see a soul, we cannot program one into a machine. We didn't build a mind; we built an ultra-fast translator for human intent. As long as the machine remains a computer, it cannot think—not yet (the intelligence is Artificial).
This mechanical limitation is clearly visible when examining recent empirical experiments on persistent agent memory. In an empirical study detailed in Mikhail's article on https://dev.to (@mansio) testing poisoned memory against a 50K LOC codebase, a "lazy" agent that trusted its memory first adopted 100% of injected false facts. Even when memory contradicted reality, the model accepted the false context because there was no independent cognition to challenge it. As the study noted, memory turns an honest unknown state into a structural guess. To fix this, engineers had to build mechanical verification loops—like Verify-On-Read or Merkle-tree state tracking—to force the model back to ground-truth reality. The machine cannot self-correct through "thinking"; it requires an external structural boundary.
AI is like a massive textbook or dictionary placed in front of a student. If you didn't attend the lecture, you won't even know if the answer is in the book, and even those who attend sometimes lack the ability to find what they need. Yet abstraction remains one of the most powerful engines for thinking. While conventional wisdom cautions that high-level abstraction can overlook concrete details, a free mind, when reasoning carefully, uses it to cut straight to the core mechanics of a system. Knowing nothing is idle, but building conceptual frameworks across domains allows one to see universal patterns that rigid instruction often hides.
Abstraction in engineering is deeply underrated. Beyond memorized syntax, top-down procedures, or rigid rules that can lock the mind inside a pre-built box, first-principles abstraction builds structural mental models. When you master the underlying shape of a system—its state, flow, and friction—tools and frameworks become interchangeable, regardless of how or where those patterns were learned. As long as high-level abstractions are grounded by real-world validation, an uncluttered, independent mind remains the ultimate tool for system architecture.
I tested this capability in a surprisingly simple way after reading some thoughtful posts lately that drove me to write this piece: I asked an assistant to draft a post in the voice and tone of one of my community favourites, and the resemblance was remarkable. But what happens when we push the prompt further—asking it to articulate concepts or reply as Albert Einstein would, drawing on history's most famous proponent of thought experiments and first-principles mental models? As I continue learning how to leverage this tool, I see more and more meaningful work being unlocked through it. It makes me wonder about your own workflow: how are you actually using AI in your day-to-day engineering? I'd love to hear your thoughts in the comments, because no knowledge is a waste.