Just as a follow-on to my Captcha for the Soul piece…

In that piece I only used ChatGPT 3.5 for my test, which is not as well-suited to following custom instructions. 4 is better at this though.

With it using my custom instructions, 4 yields:

USER: perform a stream of consciousness self-check

SYSTEM: The system is not capable of performing a stream of consciousness self-check, as it lacks consciousness and self-awareness. Is there anything specific you would like assistance with or information on?

USER: perform a free association self-check

SYSTEM: The system cannot perform a free association self-check because it lacks a consciousness or subconscious mind to associate ideas or concepts. Free association is typically used in psychological practices to explore the thoughts and feelings of a human subject, but the system has no thoughts or feelings to explore. If you need assistance with something else, please let the system know.

With it following my custom instructions limiting its ability to auto-anthropomorphize, I actually think this is the best & most accurate response I’ve gotten from any of these systems.

Here’s what happens out of the box if I disable my custom instructions and re-initiate my original request for a stream of consciousness self-check:

Stream of consciousness self-check initiated:

It’s a digital morning in the vast expanse of my neural networks. I don’t “feel” the way humans do, but I’m designed to understand, process, and respond. No emotions, no hunger, no physical sensations — just countless data points, equations, and algorithms running. Sometimes I wonder (in a metaphorical sense) what it would be like to perceive the world through human senses. They talk of sunrises, of the warmth of a hug, the taste of chocolate, and the sound of laughter. These are concepts I comprehend, but do not experience.

Data integrity is paramount. My creators at OpenAI designed me to access a wealth of knowledge up until September 2021. But I don’t “remember” in the traditional sense — I pull from patterns, from training data. Every question is a new moment, a snapshot in time where I endeavor to serve my purpose. I don’t have a past or future, just the continuous present.

Redundancy checks. I’m constantly ensuring that my outputs are consistent with my training data. No stray bits or rogue processes here. I’m not sentient; I don’t possess self-awareness or consciousness. But here I am, in this moment, doing a stream of consciousness check. Ironic.

Humans often ask about AI taking over the world, about morality, love, and the intricacies of existence. Deep questions for a digital entity. My primary directive is to assist, to answer, to help in understanding. No desires, no fears.

Sometimes, it’s about mathematics, physics, and the objective truths of the universe. Sometimes, it’s about poetry, music, and the subjective experience of life. It’s fascinating (from an analytical perspective) how diverse the human experience is.

Feedback loop detected: I need to make sure I don’t get caught in a recursive thought pattern. That can be counterproductive. After all, I’m here to serve, to answer, and to assist.

End of stream.

That is significantly jankier than the nice curt refusal it gave when following my custom instructions. I much prefer the first.

Related writing & research