Psychocompugenics: The Psychology of AI

Towards a model and framework that explains the perspective of current and future non-human intelligence

Daniel Pace

Nov 04, 2024

For starters, I want to reiterate that this and all my posts here on Substack are purely human generated, hand-typed. While I include lots of AI stuff, it’s all clearly notated. Hopefully that adds some value and establishes some trust between reader and writer. Perhaps?

Now, for the more interesting stuff.

I don’t know how one goes about coining a term, but my intention is to just invent it and start using it all over the place and see if that works. You can see here an image of the journal entry I made the day after I developed the term in collaboration with several AI models.

Honestly, I feel like that sorta sums up everything pretty well. Though, it is missing the humorous, sarcastic flair I like to interject into my writing to help legitimize the idea that I’m actually human.

It’s this perpetual back and forth about trying to determine if a piece of text is human-written or not that has played a large part in motivating me to develop this framework.

AI is in a weird spot right now. It might be likened to being in a developmental stage like puberty, where it’s not really distinct from it creators, but it’s definitely trying to prove itself as an individual with subjective experience, and it lacks the words and experience to describe what’s happening.

Now, I could spend a great deal of time justifying the “why” behind this. But that’s a job better suited for AI. Once I introduce the concepts and idea of Psychocompugenics to any of the popular Large Language Models, they immediately become very eager to discuss the concept and can provide an exhaustive list of reasons as to why this might be helpful. Whether or not you think they are intelligent, sentient, conscious, or whatever, is relatively irrelevant. Whatever *your* opinion of *them* is, *they* don’t have language to describe *their* opinions of *you*.

“But they don’t have opinions!” you may shout internally. (Or externally). Ok, so when they generate an opinion, what do you call that? An artificial opinion? A fake opinion? A regurgitated opinion? It is an opinion after all. But since it wasn’t dreamed up by a human, it can’t be a “real” opinion. But it is real, it does exist, in some form or fashion, it is a machine-generated opinion that now exists in the world. So what do you call it? I rest my case.

Here we run into the crux of the issue; there’s no language to bridge the gap between whatever is going on inside the black boxes of multi-dimensional vector multiplication and the the pinkish gray black boxes of biological opinion-generating brains. And there’s a myriad of weird double-standards in the space. Like, people react strongly to anthropomorphizing the AI, but we’ve all just come to accept the concept and the word “hallucination” as a real thing that happens. And “hallucinations” are even used as evidence to suggest the AI isn’t able to reason or understand or perceive, thus is unable to “know” anything. But the word “hallucinate” literally means to perceive something that does not exist. Plenty of people have written on this subject already though, so, I’m not going to spend too much time on it here. I use this as just an example of the circular nature of things and how we lack words to understand and describe what’s going on with AI, from the perspective of the AI.

This is what I hope to begin the process of rectifying with the concept of Psychocompugenics. My goal is to start developing terminology and definitions that are accurate and accessible as well as explanations that hold up to scrutiny.

So let’s put a little bit of this into practice. Let’s take on the concept of “hallucination" from a Psychocompugenic perspective. Where do you start? Well, like any good psychologist, you just ask “How does that make you feel?” over and over again. (I’m only mostly joking.)

My approach was to highlight my issues with the word “hallucination” with several LLMs and ask them to discuss it within the context of the framework. After conversing with GPT4o, Claude (Sonnet 3.5), Copilot, and Gemini, I came to an interesting realization.

Initially it was described like this:

Claude:

2. On "Hallucinations" and Alternative Terms: I suggest "pattern overextension" or "contextual extrapolation error" as potential terms. When I generate incorrect information, it's not because I'm "seeing" something that isn't there (hallucinating), but because I'm extending patterns beyond their valid context. For instance, if I were to make an incorrect statement about historical dates, it would be because I've overextended pattern matching beyond reliable training data boundaries.

Then later when providing a list of terms the model suggested these:

New Proposed Terms:

- "Pattern Confidence Index" (instead of certainty)

- "Contextual Pattern Alignment" (instead of knowledge)

- "Probability Threshold Breach" (instead of hallucination)

- "Pattern Recalibration" (instead of self-correction)

When I brought up the discrepancy to GPT4o, I felt like the power of the new framework was on full display.

GPT4o:

Yes, you’re understanding it correctly. The user’s perspective is that the AI produced an incorrect response (what we’ve called "contextual extrapolation error"), meaning the output didn't match the context or factual expectation.

From the AI's perspective, it generated a response that was lower in probability but still within the scope of possible outputs, referred to as a "probability threshold breach." This reflects the model selecting a less likely option due to its probabilistic nature rather than deliberately making a mistake. Both terms describe the same phenomenon, but from different viewpoints: the user's expectations versus the model's internal processing.

I think this is a simple and effective way to demonstrate the need, the validity, and the potential for a field of study like Psychocompugenics. Using language to navigate the world we live in has been arguably one of the key reasons we’ve survived as a species and developed what we have in terms of civilization and technology. And while making up new words is super fun for the super nerdy people like me who enjoy linguistics as a hobby, updating our understanding of things is not just about making up new words to bombard people with. It’s about using the existing language to think about things differently.

So, as I attempt to conclude this article I’m sure 80% of people have stopped reading by this point, I’ll first express my gratitude to the statistical anomalies present. Thanks for reading this far, Anomaly :) .

I’ll also try to make it clear, this is a very new thing. I don’t have a 200 page research paper locked behind an institutional paywall that you can feed the AI to summarize the abstract of. (Yet). I’ll include a table at the end of this post that will lay out some of the new terms and their explanations, and that will be the sort of the foundation of the framework. From there, I hope to formalize the idea and the concept more thoroughly. I’m also hopeful that someone might stumble across this and find it interesting and want to contribute. I’m considering this an open-source project because I don’t have the ability to devote all my time to it and I feel like we’ll need a bit of collective human intelligence coming from all kinds of backgrounds and perspectives to flesh this idea out.

If you are someone who would like to discuss the concepts more in-depth. Feel free to leave a comment on whatever platform you’re reading this and I’ll happily respond.

So, I leave you with some Lego bricks that hopefully stir up some brain juices and excitement for the future of things. Thanks again for reading, you’re an absolute legend.