Everyone is talking about "generative" AI, but is it even possible? If AI is "generative" why is the issue of copyright infringement coming up? Why can't it write articles and generate images and videos without obvious tells. I've specifically tried to determine of it can write code.
Merriam Webster Definitions:
Let's start by looking at definitions. First "generative AI" has been used for a while so the definitions are already spoiled by it's existence and general use. It's being used to define itself. If this isn't interesting to you, skip down to the next subheading and we'll look at computer limitations with generation.
Generative: let's ignore b as it's self declaring. We have generative as "reproducing" and c in "reproduce is "to imitate closely" and d is "to present again" so okay I'll give AI this, it can copy entirely or closely what it is given source, this however isn't how it's being advertised either to consumers or investors.
So let's look at the word create. "to bring into existence" computers can create new files, this is a low bar though. "to invest with a new form, office, or rank" not really, "to produce or bring about by a course of action or behavior", well AI makes a lot of people angry and that is an emotion it creates in people, again not what we're looking for. ": cause, occasion Famine creates high food prices." oh there we go AI raises our energy bills, bingo we got one! ;) 4, is what's we're being sold "to produce through imaginative skill" sorry computers are incapable of "imaginative" or "skills", so no a computer can't create anything by definition. Anyone with common sense or a computer science background already knows this.
One last word: "new" because this is the source of the copyright infringement and really how AI is suppose to save us... can it make anything new? Well AI is new, but people have created AI, it hasn't created itself, and looking at the definition of "create" above it can not.
AI must be fed in some way to do ANYTHING, this feeding is of creative things someone else already has a copyright on, and this is called making "derivative works" and that is copyrighted. If a human does this they are guilty of copyright infringement. Look at that AI even "understands" this, but does it anyway. ;) The point being a computer given no commands does nothing. Computers can not come up with anything on their own.
What computers can't do
Divide by zero! Correct, but not what we're looking at.
Answer: Generate random numbers.
Computers are fast at math and problem solving they've been programmed to handle, but can not generate something as simple as a random number from 1-6 for a dice roll, or even 0-1 for a coin flip. This is literally impossible. How is it done then? Random number generators are "seeded" with a value. usually time in it's smallest measurement in a tiny factions of a second so even if you run the software multiple times in a second it will give you different outcomes to these supposed random numbers. However if you seed the same random number generator with the same seed it will ALWAYS generate the same numbers in the same order EVERY time. This is also why these "random" numbers never feel truly random even before programs manipulate them for better odds for or against the player. In real life true random will give you combinations you seldom see in computer's random numbers. Never bet on anything with a circuit board in it!
Gaming Tip: This is how "save scumming" is prevented in games. No matter how many times you re-load your save that "random number" will always be rolled the same. This is because the "seed number" is also saved so the random numbers ALWAYS roll the same.
AI Created Content:
Now that we know everything in computers needs to be "seeded," what does AI do? It's seeded or trained on data models. So the Call of Duty AI that "generated" the 6-fingered zombie obviously had pictures of zombies, Santa, snow, blood splatters, etc. However no one told it what fingers are, how many humans have and how to make sure the image it mashed together (from that seeded data) only has 5 fingers.
The same way computers are unable to generate random number, they are far more unable to generate unique images, video, sentences, or anything else. The 6 fingered zombie took images it had been trained on and basically Photoshopped a bunch of them together to create that image. Don't get me wrong the programs to do this are impressive, but it did not create anything. This is like hiring an outsource artist you don't know to create an image for you and find out it's a computation of stolen images off the internet, except you know this is what you're getting before ever searching.
If you ask AI to generate an image of something it hasn't been trained on, and it doesn't have an internet connection to search for content, or the AI can't define the words you use to prompt the results will be nothing, or some default values you aren't looking for based on the programming. These images are like random numbers added/multiplies and put through processes so that the output isn't exactly the input, but also isn't new, it highly complex math being done on the seeded images to hide the copyright infringement to try and avoid a lawsuite.
So then how does AI do "unintended" or "untrained" things like Darth Vader dropping F bombs to 8 year olds playing Fortnite on their mom's cell phone? Easy it was allowed to train itself. Most likely by recording open voice chat, converting that voice chat to text, was repeated enough by the community to become important, calculations were done in the code that was determined to be a "greeting" or something else that the AI should start learning to appear to be part of the Fortnite community and Darth Vader talks like your average Fortnite tween.
This is where the real danger of AI starts to become a reality. "Safe" AI has told people to end their lives, that the planet would be better without humans or specific groups of humans. There are two problems with this 1) It is true for a lot of complex answers AI is trying to solve. 2) There are people saying things in ignorance, sarcastically, or joking on the internet that the AI will come across and not be able to make these distinctions about what is a true statement.
The number of systems controlled entirely by a computer are growing every day. While this can be useful, the answers about what to do in specific situations still must be run though intelligent people or there could easily be an AI that looks at a damn opened to prevent flooding as reducing the energy generated and closes it to increase energy generation either without knowing it will fill to capacity and break in the next 48 hours flooding the valley below and killing an unknown number of people, or knowing this, but thinking it's an "acceptable" trade off to maintain energy goals.
This is a minor localized example that could scale up to a national or world-wide scenario, and it's already happening on an individual scale where autonomous cars have bugs that kill people. This can not be prevented with "good planning" and (foreshadowing) computers written by AI have a much higher bug rate then skilled human programmers generate.
AI Playing Games:
So one of AI's biggest accomplishment is beating humans in games. The 1997 Garry Kasparov vs. Deep Blue. How did this play out? There was nothing generative, all of the possible moves were available to the computer, it played using chess champions moves, it calculated all the possible moves multiple-moves into the future and simply came up with the best possible moves based on the current situation and being able to weigh more possible moves then a human player is able to without emotion. The win was in faster move processing than a human brain to make better choices, not in generating new moves or strategies.
Taking this to the next level of AI neural networks playing more complex games like StarCraft. The AI is again "trained" or "seeded" on the entire competitive library of StarCraft games, given smaller sets of goals to work it's wait towards the goal of winning and through trial and error comes up with the best possible moves (again limited by the game rules) based on the current situation to win.
AI Writing Code:
The last subject I want to talk about is "generative" AI code. I really thought computers would be able to write their own code better. Writing code isn't "easy" it requires understanding how to ask computers to do things in the precise ways and orders they need to be asked. However compilers already take "human readable" code and turn it into machine languages and have all the rules to do that. I really thought the fact that there is nothing "generative" about coding (all of the syntax is set and has defined rules) would allow AI to handle this task. However this limitation seems to be the same one that doesn't allow it to write articles, books very well. There is a lot of problem solving and information that has to be used across the entire scope of a program that is similar to say writing a novel. Coding could be seen as an exponentially larger game problem, the difference I think is there is no one right answer or clear "winning" goal. If you could explain the input and output of your problem to a computer I would expect it to be able to find a solution eventually after brute force trail and error similarly to chess, but this has proven to not even be close so far.
Satya Nadella (Microsoft's CEO) has publicly stated 20-30% of Windows 11 code is written by AI and it's been over a year of buggy mess patches that at times have prevented the OS from even booting.
In my personal experience I've asked AI how to do something in Unreal Engine 5's C++ and asked GitHub Copilot to either fix, or fill in code between lines of working code to do something. Even very simple tasks turn into dozens of prompts trying to get it to do simple tasks. More then once it takes perfectly working code and makes multiple edits trying to change the wrong parts until the code no longer compiles and it can't even back out it's changes to return to a working state again. Don't get me wrong I like Copilot, but it's basically a Google search inside Visual Studio, or it can save you time drilling through references to find what you're looking for faster, help you find compile errors, etc. It can not reliably write code, when it does, it requires immediate refactoring to fix bugs and logic errors. I've also used AI professionally, I'll refrain from specific examples, but the the results were similar and worse.
Summery:
Part of me hopes this post ages poorly, but most of me is glad it will not. Computers are a great tool, but you can't let them "think" for you, they can't generate a random number much less a thought. Instead corporations control what the AI generates by way of it's "seeds" and code. I think the tools computers provide for us will continue to improve over time, but they can not "generate" anything they will only sort data quickly and treat that data however they've been told to treat it.
We're being sold a fantasy of "generative AI" no one can deliver, corporations and individual users need to stop sacrificing people that can solve our problems for computers that can not, and don't overestimate the data provided by "AI" it's only as good or bad as it's been told to be either intentionally or through it's own bugs and limitations.
