There’s something…synchronistic…about the fact that the last thing I wrote before effectively abandoning this blog back in 2016 was at least partially focused on the ethical questions surrounding AI, and the fact that I resurrected the blog because I wanted somewhere to do something with the countless AI images I was generating for fun nearly a decade later.
And how (some) things have changed in that decade…
The End Is Nigh
Technically at least, it’s not (yet) what we always thought we meant by the term AI. The launch of both Midjourney and ChatGPT in 2022 ushered in the age of generative AI and, potentially, initiated the end of the world as we know it. :D
Countless millions of words have already been written in these two years about the uses, the effects, the impacts, the threats and the opportunities presented by this quantum leap forward in computing technology. Nvidia, manufacturer of the silicon chips (developed for computer gaming to allow multiple simultaneous processes to be carried out very very quickly) that turned out to be ideal for AI-related computation, has become one of the top 3(?) (it changes so fast) most highly valued companies in the world, and “AI” is on the lips and in the minds of everybody now.
There is almost no reason whatsoever for me to spend any effort ruminating on this subject, but that hasn’t stopped me yet.
It’s Not (Really) Artificial Intelligence (Yet)
In the past, artificial intelligence was pretty much assumed to be…well…sentient. Self aware, self-deterministic, basically like a human intelligence, but…artificial.
Now however, this kind of AI, like Heinlein’s Mike (1966), Clarke’s HAL (1968), Scott Card’s Jane (1985), and plenty of other examples have been rebranded as “AGI” or Artificial General Intelligence, to differentiate between the so-called “Narrow AI” of generative models, facial recognition systems, etc. etc. The general public however, almost certainly rarely makes this kind of distinction, and to them, AI is AI.
Although “narrow” AI like the LLM’s (large language models) used for text generation, is not “intelligent” as we use the term in relation to people, humans are largely unaware of the difference. The fact that LLM’s are literally mathematical language models which select the next word in a sentence based on the probability of that word being the statistically most likely subsequent word in its corpus of training data is not only irrelevant to people, it’s largely not understood either. (I recommend this excellent short paper, Talking About Large Language Models, for an easy-to-understand overview of how a LLM like ChatGPT works.)
But It Might As Well Be…
The thing is, LLM’s are “programs” for mimicking human speech. That’s what they do. Whether the speech was accurate, factual, etc. etc. was never the point. It’s becoming part of the point now, but they are just so good at mimicking human speech that it’s almost impossible not to anthropomorphise them when you’re interacting with one.
The average user has no problem assuming (and acting as though) the LLM is a conscious entity. The fact that LLM’s don’t know anything is not even on their radar. Anthropomorphisation is a common trait in humans. To the extent that a 2013 study found soldiers were forming attachments to their bomb disposal robots, even to the point of mourning them, and sometimes risking themselves to save robots deployed in military situations. And these were just remote control “robots” guided by a human operator.
Our natural tendency to project thoughts, feelings, motivations, etc. onto the things we interact with has a long-standing history. And as those things become more responsive, the tendency is only stronger. And it is (of course) leading to all sorts of interesting social (and ethical) dilemma’s in both the use and the creation of these tools.
But Is It A Revolution?
Every now and then, some form of technology is developed that literally changes the world. It reshapes our societies, our behaviour, our interactions, etc. Agriculture. The wheel. Metalwork. Gunpowder. The printing press. Internal combustion engines. Electricity. Transistors. Microchips. The internet. Smart phones. And now…AI.
I do not think it an exaggeration to say that the impact of AI on human society may be as significant, if not more so, than the development of the internet itself. Once we figure out what the hell to do with it. Right now, we’re caught in the “Hype Cycle” for AI. We may already have passed the “Peak Inflated Expectations” phase technologically, but in this instance, public perception may still be hovering around there, and unusually, we may potentially either experience only a fairly shallow “Trough of Disillusionment,” or we may skip straight into the “Slope of Enlightenment.”
Certainly the proliferation of “AI” apps etc. suggest that developers have no intention of backing down any time soon, and while we might be building an AI bubble, indications are that it is going to be a massive one.
More AI Ethics
When last I spoke wrote about the ethical questions involved in AI, it was from the perspective of the ethics of AI itself…how could we make it ethical, whose ethics would it have, etc. etc. But, without knowing at the time, what I was talking about were the ethics of AGI, not the “narrow” AI that we’re seeing proliferate now.
Turns out, there was a whole other question of AI ethics that we didn’t think about until it was effectively too late already. (Because rushing in is what humans do best.)
Generative AI like Large Language Models depend on being trained. And what they’re trained on is data. When you want your AI to be able to perfectly mimic human conversation, what you need to do is give it a huge amount of actual conversation to “learn” from. When you want it to produce images, you have to give it a huge amount of images to “learn” from.
Problem is, those conversations, those images, were all created by other people. And in some, perhaps many, cases, that data is owned by those people. And so, the great copyright question has arisen. Generative AI models have effectively been trained by scraping vast amounts of information from the web. And since the web contains (in some respects) almost the sum total of human knowledge somewhere and in some form, (publicly accessible or not), that is a LOT of information.
As a result, the pretty reasonable question of what kind of permission was granted, what kind of remuneration was offered, and to what extent the use of that data constitutes plagiarism on the part of the AI.
Training, Stealing, Copying – The Generative AI Dilemma
The reality is the genie is already out of the bottle. Alea iacta est. The generative models already in widespread use have been trained. They’ve scraped the vast body of human-produced data that is, for the most part, freely available on the web, it’s all been annotated, tagged and tokenised (largely by human beings working for slave wages in developing economies with a glut of cheap labour) and fed into the deep learning systems of countless digital neural networks to create systems that can spit out 1,000 (or more) words on any topic you care to mention, or an image based on whatever description you care to give it.
Is this plagiarism? Well…it’s complicated. The LLM does not (usually) disgorge that training material verbatim on request. It does not pretend that its output is the product of original thought (or computation). It (now) tries to cite sources for the answers it gives. Although it has learned facts this way, what it has really learned is how to talk. Similarly to a child learning to talk by spending a few years listening to what adults say and how they say it.
As somebody who is fundamentally a writer, I’m a bit ambivalent about the whole thing…in one sense, call for my craft is diminishing as people tend to reply more and more on the output of generative text “AI.” In another, AI writing is still AI writing, and as with human content, it usually needs editing (by an actual human being) (like me) to make it really suitable for whatever purpose it’s intended for.
Should the content creators of Reddit and Stack Overflow and Wikipedia be somehow compensated for the use of their output in training LLMs? Maybe they should. I’m not necessarily opposed to that. But I do think it falls short of “stealing” their (my) work.
I’m a bit less on board with the idea that text-to-image models are stealing the work of artists though. If, as an aspiring art student, I study the works of Van Gogh and practice his technique until I can paint in the same style, have I stolen his work? Was Picasso, who was taught to paint by his father, stealing his fathers work every time he produced his own painting in a style recognisably similar to what he had learned from? Not so much.
Should artists whose work has been used to train image models likewise be compensated? To the same extent that artists are compensated when their work appears in art textbooks at least? I would say yes again. But generating an image in the style of Van Gogh does not involve taking bits from existing Van Gogh paintings and compositing them together to create a new painting, and it certainly does not involve claiming that it was painted by Van Gogh.
Gatekeeping Skill Sets
To an extent, I understand the…frustration…of artists on this level. They have spent years, perhaps decades, perfecting their skill, and now some upstart with a computer comes along and can produce an image in mere seconds that would take them days or weeks or months to do. Of course they’re upset about it. But as I say in the intro to my new AI Galleries page here, I’m finding this democratisation of art incredible.
For somebody with basically “full spectrum” Aphantasia, the ability to convert words in my head into an actual image that I can see is nothing short of amazing. I plug in passages from books, verses from poems, the lyrics from songs, purely to see what it comes up with. I’ve downloaded literally hundreds of GB of checkpoint models for image generation, and I’ll just run the same prompt on 20 different models, one after the other, to see how their training producing a different image from the same input. It is, frankly, incredible. And I’m having a lot of fun with it.
Of course, I’m not a struggling artist trying to make a living with my skills and afraid of becoming obsolete either, so points off for that probably. (The GF, who is an artist, (the paints and brushes type), thinks it’s great too. She never uses it herself, but she follows a lot of AI image communities, just to see what people are producing. In general, I find artists either love it or hate it.)
Of course, the other other coin is the “AI actors” being digitally inserted into movies and stuff like that…I lean a bit more toward the actors view point there, especially in the sense of established actors being replaced by digital copies of themselves. That one, I’m not such a fan of.
Amelioration
One very interesting thing that offers us humans a potential ray of hope though is that in order to get better, models need to be trained on more human content. And human content that they haven’t already trained on is running out.
Models trained on AI output rapidly undergo what has been termed “model collapse” and become effectively useless.
So, the robot overlords will probably need to keep us around for a while yet. ;)
The Human Way
As I’ve mentioned before, the human way is to first do it, then figure out how and if it should actually be done. It’s been done. The “AI” revolution is upon us. Now we have to try and figure out how to manage it, regulate it, deal with the fallout and the questions and the casualties.
If I had to give any advice to the writers and artists out there trying to decide how to navigate this new world, it would be that when a tool comes along that threatens to disrupt how you do things, your first priority should not be to fight against that tool. The time for that fight was before we even knew it needed to happen. No, the priority needs to be becoming an expert at using that tool. Leverage it. Turn it to your advantage. Take its strengths for yourself.