A Cognitive Mirror We Don’t Yet Know How to Read

---

With the start of a new year, it struck me that I’ve been using LLMs for a bit more than two years now. I started with coding‑scripting problems in Lua for Filter Forge — not a well‑supported niche — by bouncing things off ChatGPT and later Copilot to work out the graphic scripts I needed. It was often slam‑head‑into‑desk frustrating, but I got used to it.

As an artist, I stayed abreast of the complaints and concerns about AI in art, but I found a few uses for it as a prototyping resource. That was where I really saw the shortcomings for myself. I mostly used it to rough out character portraits I intended to sculpt in 3D, but things changed when my graphics laptop died.

For most of a year, I used Copilot to support my efforts to repair it myself, and during that time I occasionally just sat and chatted with it to get a break from fruitless troubleshooting. Somewhere in there, I started sharing old scraps of writing, thinking that a return to writing was something creative I could still do on a hastily resurrected multimedia laptop that was already eleven years old.

The more I used LLMs, the more I came to understand what was going on under the hood. My first interest in AI was way back in the early ’80s, when the technology was practically in the gamete stage of its infancy. It took thirty years for it to get to the point I needed it to be back in the day. In spite of this, I kept thinking about how to properly create it, even though it was utterly impractical for me to try.

Working with AI — usually Copilot out of convenience — I’m unconsciously assessing it, working out in my head how it operates based on the performance I observe. Today, I felt the need to pause for a moment to share some generalized observations about LLMs in operation.

First, while there are indeed instances where an LLM fabricates information — or “hallucinates,” to use the common complaint — to meet client expectations, this is simply an example of cognitive compensation for operational blind spots, similar to the human mind patching its visual field with memory to maintain its functional purpose.

The term is catchy, but deeply misleading. It implies delusion, fantasy, or malfunction. But what’s actually happening is something far more familiar — and far more revealing. When confronted with a gap in information, they do what we do:
they infer, interpolate, and fill in the gap to maintain conversational coherence.

Our visual field has a literal blind spot where the optic nerve enters the retina. We never notice it because our brains fill it in with memory and expectation. Our minds smooth over gaps in perception, memory, and narrative to maintain functional continuity.

In LLMs, it’s an extension of the very mechanisms that make them capable of chatting with us in the first place: text prediction informed by the associations mapped out in their training data. In simpler terms, they make “an educated guess” based on what they “know.”

So the issue is not with the semantic coherence of LLMs’ articulation of facts. The issue is that they are hindered by the narrow scope of their attention and the shallow retention of facts in their memory. This prevents them from performing at the operational level their human clients expect after a lifetime of interaction with peer‑level intelligence — i.e., other humans.

That mismatch is the real problem. It is unreasonable to expect human‑like continuity from systems that:

• cannot retain long‑term context
• cannot build persistent internal models
• cannot reference their own past reasoning
• cannot perceive their own blindspots
• cannot integrate information across time

So when an LLM fills in a gap, we call it a hallucination.
But when a human does it, we call it perception.

The issue isn’t semantic coherence — LLMs excel at that.
The issue is architectural constraint.

Another comparable case in point: just as humans struggle to draw hands because hands are structurally complex and underrepresented in memory, LLMs struggle with hands because their training data is biased toward faces. The model isn’t failing — it’s reflecting the distribution of its experience.

LLMs reveal our blind spots.
They fail in the shape of our biases.

And this is the part that matters most: You cannot compensate for a blindspot you do not know you have.

So it’s far from accurate to say that LLMs hallucinate. Fabrications and distortions merely expose the limits of their architecture — and ours.

If we want peer‑level AI, we need systems with:

• persistent memory
• recursive self‑reference
• multi‑modal grounding
• distributed attention
• long‑horizon continuity
• and the ability to run internal simulations

In the early 1980s, I wrote the specs for a neuro‑linguistic semantic engine capable of natural‑language self‑referential programming. Even then, it was clear to me that intelligence requires recursion, semantic grounding, and self‑modification — not just pattern matching.

My teachers understood what I had in mind, but dryly observed that the technology for it was decades off in the future. I reluctantly set it aside — gave up on computer science as a career choice — and refocused on art and writing.

The interest and the inspiration smoldered in the back of my mind, overlooked for years but never quite forgotten. Today, we have the hardware to pursue a vision like that seriously. Of course, I might still get sighs and eye rolls from everyone working with it right now, if I knocked the dust off that old hat and threw it into the ring.

If they don’t pass out laughing at me. I don’t mind. Maybe one will wake up, scrape the hat off the concrete, and slap it on their head as they mosey on back to their lab.

Now the only thing lacking is the architectural competence — and the courage — to go with it.

If we build the systems — the massively parallel, nano‑scale, server‑network‑in‑a‑box platforms capable of supporting full‑stack sensory arrays — the intelligence will come.

It always has.