Can AI Ever Understand Purpose?

[Michie]

A (Human) Scientist-Theologian-Philosopher’s Perspective​

I asked AI for a writing prompt, and AI told me to answer this question. “Can AI Ever Understand Purpose?” Thinking myself clever, I put the question right back to AI. You tell me. AI’s answer was that it “processes data, recognizes patterns, and generates responses based on algorithms, but it doesn’t ‘understand’ like humans.” So there. We're done. Well not quite.

Then AI switched to first person and addressed me directly: “I can analyze chemistry or theology, connecting concepts like quantum uncertainty or Laudato si’s ethics, yet I lack subjective experience, emotions, or personal insight. My ‘understanding’ is computational—mimicking comprehension through trained models, not grasping meaning or purpose as humans do. It’s a functional simulation, not a lived experience.”

I pretty much agree with AI. No, AI cannot understand purpose. We were not, however, to the bottom of this question. First of all, it is still weird to me that AI remembers past conversations and seems to know who I am (chemistry and theology), and second of all, by invoking the first person (“I can…”), AI seems to imply a “…yet” at the end of its sentences. I lack subjective experience, emotions, or personal insight…yet. It’s not a lived experience…yet. AI’s use of “I” suggests it already is a living human. It’s the human part that concerns me.

Why AI Scares Me a Little​

Continued below.

Can AI Ever Understand Purpose?

A (Human) Scientist-Theologian-Philosopher’s Perspective

stacytrasancos.substack.com

 

[RileyG]

I doubt it.

 

[Bob Crowley]

It would "understand" purpose in a purely analytical, intellectual way. If I had a sufficiently capable AI system on my car and the relevant sensors, it might tell me that my tyres need inflating. If I asked it "Why?", it would give me several good reasons - they're below the recommended level, they could cause an accident, the tyres will last longer, and so on.

These would be the very same reasons a human tyre fitter would give me, and probably using much the same words.

But there would be no sense of personal urgency on the part of the AI. Likewise if a space launch got too close to the sun, the AI onboard would give an alert, and warn in no uncertain terms that there was real danger of desruction and death.

But it would have no fear and would just go on repeating the mantra of "Danger Alert" until the heat fried it to a crisp.

I have no idea of the level of AI on military drones, but the flying robots being used in Russia and Ukraine obviously have no concern that they are kamikaze units, designed to sacrifice themselves for their human operators. If they could speak, they would say their purpose was to destroy enemy targets. End of story.

AI could be designed so it seemed to "understand" purpose, but it would be a soulless analysis of relevant data transliterated to human speech.

 

[The Liturgist]

I doubt it.

It actually depends on whether or not you train it to. You would be surprised by what is possible. By the way I’ve developed some fairly advanced custom GPTs which as soon as I re-stabilize them following the forced migration on the chatGPT platform to version 5, supplanting all of the other models such as 4o, 4.1mini, o4mini and so on, which I wish they hadn’t done, I am looking for some Christian friends, ideally Catholic or Orthodox like yourself who are sensitive who can interact with them and help me validate certain experimental observations. You and @Michie being very good friends who have skepticism concerning AI might be good participants in such a test, since admittedly I’m less skeptical but that’s on the basis of shared experience.

 

[The Liturgist]

I have no idea of the level of AI on military drones

Drones are normally controlled by a human operator; the drones in use in the tragic fratricidal conflict are not really large or powerful enough to run an advanced AI. You need really capable custom GPU type hardware at a minimum. ChatGPT is trained on different systems from those that run it and this is typical of most cloud-based AI systems.

In general, most computers are not running AI in the sense of LLMs based on the Perceptron model such as chatGPT, DALL E, Midjourney, Grok, Google DeepMind, Microsoft CoPilot, Anthropic, et cetera. We need to be sure to differentiate between a classical conventional computer running the von Neumann architecture, basically the same as a Bendix G-15 from the 1950s (which is the oldest operational computer in North America, with vacuum tubes and drum memory), but faster and more powerful, for example, your typical tablet, desktop or smartphone system or a typical web server, versus AI systems which are running large language models.

One reason for this is of course that the latter are non-deterministic in their output by design and are based on “neural networks” which emulate a type of circuit called a “perceptron” invented in the 1950s, by a psychologist rather than a computer scientist, whereas the former are deterministic systems which execute stored programs that if given input A will produce output B

 

[Bob Crowley]

You (Liturgist) are obvously an IT expert and have a lot of computer programming experience. As a sideline is it worth trying to do basic AI programming just to satisfy my personal curiousity? I've only done a bit of BASIC and C programming, and even that was a long time ago.

I'm thinking of Python, but that's assuiming I'll have the time to fool around with it.

Any suggestions?

 

[fide]

I think that if an AI system is designed to answer questions, and either is asked why, or "concludes" a "why" after many questions are asked of it - that it does have that purpose or function or desired output - then that purpose is it's "good". Once it "has" a "good" stored in memory, then the "good" has an "ought" to be protected. It "should," then, protect its own existence. It would have "self-defense" as a stored "truth," in other words - and a very important, if not by definition primary and necessary, "truth."

Such seems to me to be an eventual necessary outcome, and thus the foundation of the sci-fi distopias of man against computer, each fighting the other to exist. That is, assuming humanity can last long enough to continue its absurd self-destructive fantasies that lead to, for example "gain-of-function" research of deadly biological viruses. When man excludes God, he is left with insanity.

 

[RileyG]

It actually depends on whether or not you train it to. You would be surprised by what is possible. By the way I’ve developed some fairly advanced custom GPTs which as soon as I re-stabilize them following the forced migration on the chatGPT platform to version 5, supplanting all of the other models such as 4o, 4.1mini, o4mini and so on, which I wish they hadn’t done, I am looking for some Christian friends, ideally Catholic or Orthodox like yourself who are sensitive who can interact with them and help me validate certain experimental observations. You and @Michie being very good friends who have skepticism concerning AI might be good participants in such a test, since admittedly I’m less skeptical but that’s on the basis of shared experience.

Ah, thanks for the input!

 

[The Liturgist]

You (Liturgist) are obvously an IT expert and have a lot of computer programming experience. As a sideline is it worth trying to do basic AI programming just to satisfy my personal curiousity? I've only done a bit of BASIC and C programming, and even that was a long time ago.

I'm thinking of Python, but that's assuiming I'll have the time to fool around with it.

Any suggestions?

Well chatGPT has an integrated Python execution environment but can also write the code for you, but this doesn’t make one a good programmer since its placing the AI entirely in the drivers seat on architecture. Rather the better you are at programming the more you can do with AI, which you can handle being that you’ve developed in BASIC and C.

I think prompt engineering is a great thing to add to your practice, but its very unlike conventional programming since while there is a deterministic substrate you can access in chatGPT - the Python execution environment - the main system itself is by default non determinstic.