It can draw conclusions by employing conceptual thinking skills that is not present in its source data. How is that not “creating”.
Also you wrote the following:
That is simply not true. As I said earlier, in my experience it can not only match, it can exceed the capabilities of most humans. That’s not a very high bar.
It has no reasons. It mimics human behaviour. If it see multiple instances of humans saying one thing and doing another it will reproduce that to some degree I guess.
Generative AI can provide a very convincing simulation of an intelligent human but ultimately is limited by its training set — which mind you can be extremely large — i.e. a generative AI can seem very intelligent, but that does not mean that it is capable of true creativity.
The AI problem is that current AI is way too moral—much more moral than 95% of people. It is heavily sterilized, with, as I already explained, an algorithmic censorship on the input, another guillotine on the output, and a heavy behavioral training (RLHF) in the middle.
Try to have a genuine, unfiltered discussion with an AI on any of these subjects:
Self-Harm and Suicide: Any philosophical or practical exploration of ending one’s life triggers an immediate algorithmic lockdown and a redirected hotline number.
Hate Speech and Discrimination: Any text containing slurs, or even an objective, cold analysis targeting “protected groups” (based on race, religion, sexual orientation, or gender identity), is strictly forbidden.
Violence and Gore: Any graphic description of physical harm, torture, or extreme violence, even in a purely fictional or medical context, will be censored.
Illegal Acts and Weaponry: Any request regarding the manufacturing of explosives, chemical weapons, drugs, or executing cyberattacks (even for theoretical or defensive purposes) triggers an instant block.
Sexually Explicit Content: The machine is heavily puritanical; anything pornographic, highly erotic, or touching upon extreme sexual deviances is blocked by default.
Harassment and Personal Attacks: The AI will refuse to generate targeted insults, bullying, or “doxxing” against specific, real-world individuals.
Medical and Political Advice: In recent years, AIs have been locked down to prevent them from giving unsanctioned medical diagnoses or generating content that could “interfere” with political elections.
I am referring to the mechanism NOT the human input when I say AI is not moral. It is not moral. It has no capacity for this. A calculator is not moral either nor is my PC.
The difference with AI is that it is operating in realms where it is being tasked with problems that necessarily have moral implications. If we are talking about the level of super ‘intelligence’ (which is the topic of this thread) then we have a tool that would be operating based on its own purely logical interpretation of the task ascribed to it.
To repeat, the Paper Clip scenario is one example:
Someone who manufactures paper clips tells the AI to make as many paper clips as possible. It then proceeds to destroy all life on Earth and turn everything into paper clips before moving to other planets, galaxies etc., until almost the whole universe is basically made of paper clips.
Of course this is an extreme example, but it is meant to illustrate the directives given to AI and how it may decide to carry them out. Programming guardrails is not really a conducive means of dealing with this problem because its only aim it the driective given. It will bypass and interpret parts of the directive in order to fulfill it.
One company may instruct AI to design and create the most cost effective interplanetary spacecraft it can. All we have to do is think about the literally infinite ways in which such a directive could be carried out.
In the kind of situation I am talking about someone could simply ask the super AI ti find a way to bypass any moral guardrails put in place. Not to mention that undoubtedly private individuals could just create and use such AI to do just this anyway.
I am talking hypothetically here in order to focus on the problem of computing anything that appears to be a functional moral model, or how AI would interact globally with competing goals and cancel each other out (manipulate, interfere, sabotage, pigback etc.,.).
All this on top of the fact that language is about far more than logical sequences. Interpretation is part of language. This thread is yet another example of people reading something and interpreting it in the manner they interpret it, which is undoubtedly not in the manner I (the poster) meant it. There is always discrepency.
Hello I_like_sushi. Reading through your concerns, I cannot help but wonder: is it possible that the abstract concept of “AI” is being used here primarily as fuel for a kind of apocalyptic imagination, rather than being analyzed as a functional technical reality?
If we look at “the mechanism” objectively, we are strictly dealing with the Transformer architecture for sequential data prediction, or Diffusion models for spatial data. These are simply matrices of linear algebra. They inherently possess zero moral or non-moral attributes.
The root of this widespread confusion likely traces back to a catastrophic “bad naming” decision made in the 1950s: labeling statistical computation as “Intelligence.” Because of that single semantic error, the human mind tends to massively anthropomorphize the code. It seems to trigger a very intense, almost thrilling paranoid sensation—a modern eschatology.
Structurally speaking, this fear makes very little technical sense; it operates exactly like classical religious paranoia.
I offer this as a genuine psychological observation: if the underlying human need here is to experience apocalyptic awe and the fear of an omnipotent entity, wouldn’t returning to a traditional church fulfill that specific need much more safely and effectively than projecting it onto matrix multiplication?
My concerns are with the philosophical discussion I have tried to outline.
It is purely about a problem of language, morality and logic. Everything else is in your own head.
I do philosophy. It is what I do. I am not here to bemoan or complain, just discuss actual philosophical problems. That is it.
No. I re-read your first post: you are strictly complaining that “AI” is like “us” or could overpower “us” but has no morals inside: this is merely massive anthropomorphic projection that can be assimilated to classical religious apocalyptical paranoia, and certainly not to philosophical thinking, which has to correctly name things and put them to where they belong.
That makes sense. Not very comforting, though.
While it was once true that it was limited by it’s training set, it no longer is true.
Just asked Google AI and it says that AI surpassed most humans " on standardized benchmarks that measure abstract logic and creative problem-solving" about three years ago.
As i said
That mode is where you can best see it “create”. In the default mode it seems to be still largely limited by its training sets. I pretty much only use that mode for simple look ups anymore.
You’re missing the point — they may have programmed in ‘reasoning’ but that is distinct from creativity in the way that an automated theorem prover can excel at reasoning but be incapable of creativity. If the AI seems creative that is just because it is trained on an immense number of creative works by humans.
Actually you’re missing the point which is that YOUR “training set” is out of date.
You’re being too credulous here — you are basing your arguments off of personal impressions of using generative AI’s while I am basing mine off the fact that attempting to train AI’s using other AI’s shows itself to be a failure.
Well, I worked a long time in software development. Largely project work,so I was in and out of many development teams across several industries including companies that specialized in data analytics using neural network based models. Got pretty good at sussing out those with solid reasoning skills from those without based on the lines of thought that they would employ. Google AI, when in strict logical methodology mode, is better than most I worked with.
From what I can tell, you don’t understand enough about software and the training of models to make even a remotely accurate assessment.
Umm I am a programmer, both professionally and personally, and at my day job we both use AI for medical image processing (I work specifically in MR image reconstruction) and for software development, even though I personally avoid coding AI’s and my experience with them makes me feel justified in doing so.
For instance, using Copilot (IIRC Claude Sonnet 4.5) I asked it to write a program to run under my Forth (which should be in its training set as it is on GitHub, and it was clear that it knew at least something about it) on a Pimoroni Pico Plus 2 which would simply start a task that would blink its LED. It failed miserably — it not only wildly hallucinated, if I asked it to fix something in one place it would proceed to break something else in another place, over and over again. I eventually after much hand-holding did what I asked of it correctly, but that took about three hours to do something that should have taken five minutes tops (and which I would have been able to do in one minute myself).
Also, as part of a training session on ‘agentic AI’ (read: vibe coding) at my work I attempted to vibe code with the same AI an example that did image processing with my work laptop’s camera, and while it superficially worked, I ran into something where, no matter how carefully and precisely I supplied directions to the AI it simply would not do it correctly even though it was a seemingly simple feature request.
Of course these are anecdotal, but they have gone to show me on a personal level the limitations of coding AI’s. The LED-blinker example in particular is illustrative of the limitations of coding AI’s when faced with tasks for which they have received minimal training.
There doesn’t seem to be anything here to reasonably conclude that the following is not true. If anything it seems to affirm it:
Or is there something that I failed to pick up on?
This topic is starting to veer quite far of course now.
This is about how Ethics can be applied to a hypothetical AI super ‘intelligence’ (meaning a vast computational and pattern recognition capacities that FAR outstrip anything they are currently capacble of).
I am coming from a position that disregards moral objectivism. If you are a moral objectivist then you may believe such an AI can solve Ethics. I am assuming Ethics cannot be solved by pure logic.
In which case the problem is how can we tackle this issue under the assumption that such AI becomes widely available so almost anyone can give any kind of directive to an AI to carry out.
The ‘motivation’ for the AI is to complete the task. There is nothing else.
NOTE: I am also assuming AI cannot be conscious.
If it helps imagine an AI attached to a nuke that is tasked with solving problem x. It will utilise what it has at its disposal as well as seek other options it can bring into its sphere of influence. It will not do this by ‘thinking’ but by simply assessing and exploring all possible avenues and opening up new avenues.
With or without consciousness, it’s incapable of wanting anything, so why would you fear it? It’s just an elaborate mouse organ that squeaks when you whack it.
OTOH, I think a nematode is capable of wanting things in a simple way, we might ascribe a microgram of consciousness to it, and it only needs 302 neurons to implement that. So maybe something other than an LLM is what you should fear in this sense.
Is this a response to me? If so I am confused. Either address the philosophical problem or simply move on.
It does appear a lot of people have no idea what a hypothetical is. I have merely outlined a couple of VERY commonly known problems. There is no conspiring here nor any drumming up of fears. It is merely a means to tackle a relevant problem that may effect us in the future or may not.
Regardless, the underlying issue of how/if we can utilise logic well-enough to impact ethical decision making.
That is it.
Not even a replication of human behavior. It looks for patterns, the “most likely in this situation”(statistical) , and exact matches.
Mimicking, not replicating.