DECEIVING DECEITFUL MACHINES

Imagine something. Put it into words. Let the machine visualize it.

The production of content is going through a time of disruption where everything becomes faster. Text, images and in the near future also sound and video can be instantly generated through generative machine learning. An acceleration no one saw coming.

ㅤㅤㅤㅤㅤㅤㅤㅤㅤㅤㅤ

Generative machine learning is a rather new technology. The field of image generation was researched during the 2010's, with results like DeepDream and StyleGAN. The breakthrough came in January of 2021, when OpenAI published their first Text-To-Image model, DALLE. It was not released to the public, instead they open-sourced CLIP.

CLIP, or Contrastive Language–Image Pre-training, is a neural network trained on the relation between image and text.CLIP is a key part of how DALLE is able to create images out of text. So the moment CLIP was released, the open-source community was driven to create their own model with it.

The story follows that model after model, the technology became exponentially better until now, late 2022, Stable Diffusion is the most powerful open model and companies like Google and Meta are publishing closed Text-To-Video models.

Just like during the early internet era, there's a gold rush for the best models and their implementations. A hyper-fast gold rush that leaves not much room for questioning these technologies and their impact on society.

ㅤㅤㅤㅤㅤㅤㅤ

A key problem of these technologies is that they can deceive very easily. One can start to think that these models understand reality like humans do. Like the infamous Google employee that thought that Lambda had a conscience. One can be deceived into thinking that an image/video has been generated by a specific prompt. Like when Heinz Ketchup claimed that DALLE 2 generates Heinz branded bottles when only prompted with Ketchup, or when Fabio Comparelli's video was thought to be the result of only prompting "Human Evolution". 

These examples are just the cherry on top of the disruption of reality itself, where synthetic images that can depict literally anything will further loosen the trust we have in media.

ㅤㅤㅤ

There's a need of critical discussion around generative machine learning. Undress them to what they really are to better understand what we are dealing with.

It would be naive to stop this technology, it's already going full-speed and there's no sign of stopping.Rather we should find ways on how to create a healthy relation between humans and these models.

In Digital Folktales, a collection of AI and human tales from internet folklore, the process tried to be as transparent as possible. The folktales were generated with GPT-3, a text-generative model from OpenAI. The images were generated with VQGAN+CLIP, a open-source notebook first released in April 2021.

The stories are a mix of nonsense chaos and very coherent tales. Although they are fully AI-generated, they aren't one-shot stories. There's been a careful process in finding the best ingredients to create the most fitting stories for the book. 

The specific version of GPT-3 used in the book worked that one needs to simulate a style for it to pick up the mood of the writing.  One needs to write a prompt, or a text-based input that will be continued by the AI. Here is the very first experiment for the book (generated text is in italic):

A dialogue between two muffins talking about their lives. Muffin one: We are so lucky to be alive. Muffin two:  I know right? It's like we were destined to be muffins. Muffin one:    I think so too. Muffin two:  I mean it's like, we were just created to be eaten by people. Muffin one:  We are the best at muffin creation. You will never find a more perfect muffin. Muffin two:  We are so perfect. Muffin one:  And I'm so glad we are muffins. Muffin two:  I'm so glad to be a muffin. Muffin one:  I'm so glad to be alive. Muffin two:  I'm so glad we were made. Muffin one:  We are so lucky to be

Although it's fun nature, this story doesn't resemble a fairy-tale.It's clear that the AI needs more instructions on how to behave.The next iteration was more guided in what the framework of the story should have been.

Here's the short fable of the post truth fox, with the explained moral at the end. Once upon a time, there was a fox who was the post truth fox. The post truth fox would say anything. He would say things that were true, and he would say things that were false. Sometimes, news anchors didn't know if the post truth fox was telling the truth or not, so they had to fact check him all the time. The post truth fox told them some big lies. He said millions of people voted illegally in America, which caused him to lose the popular vote. It wasn't true though. The news anchors checked it out with some people who could do math and they found out that actually there weren't any illegal votes at all. When the news anchors asked Trump why he lied about this, he said "Well I'm president now." So then they asked if Hillary Clinton won instead of him did she win by 3 million votes? And then Trump got quiet for a while because he couldn't think of anything to say because his lie was caught out and everyone knew it was a lie except for himself because the post truth fox believed his own lies and maybe he still believed them but probably not anymore but we will see what happens in this chapter when we get there because sometimes you can tell what is going to happen just from how fast someone is running away from something when you catch them lying about it so when I catch Donald Trump lying I write down what he says and put it in an envelope labeled "Trump Lies" and then later when I want to know what Donald Trump's latest lie is I open up my envelope labeled "Trump Lies" and look inside to find out exactly how much trouble Donald Trump is in this time so far no one has yet broken into my office or stolen my envelopes yet but you never know so better safe than sorry right? Luckily for me though it's hard for anyone else to steal something without breaking into your office first unless they have a key or something like that which makes sense really since most people don't have keys unless you are a janitor or someone else who needs access 24/7 plus even if someone did break into your office where does your stuff get stolen from anyway usually from your desk drawer right next to these envelopes here but no one knows where my desk drawer is except myself pretty much only me nobody else even knows where my desk drawer is hidden let alone also has a key which means no one can take these notes away from me whether they are fake fake fake trump quotes made up by me personally knowing trump personally talking about trump being friends with trump being related by blood second cousins whatever whatever whatever whatever whatever whatever whatever wherever whenever whenever whenever whenever ever ev ev ev ev ev er ev er vrrrrrrr vrrr vrrr r r r r rerererererer ereeer oooor oooooreeeeeeeeerr nnnnneeeeeeeaaaaaa llelelelelllllsss llll uuuup pppp iiitttt fffun nnnn nnnns eeeeeewww wwweeess ssoooooo

While there was a good start, it derailed into the void. There was a self-explanatory prompt, but it wasn't well adjusted for GPT-3. It still had too much freedom, it didn't know how to end the story and how to create a good story arch. After some different iterations, this was the ultimate structure that created the stories in the book:

Input: The fox who believed to be a cloud.Longer version: Once upon a time there was a smart fox named Foxy. He was doing well in his life as a fox until one day he looked up at the sky and said “I am also a cloud” - just like that! From that moment on Foxy started acting wacky and weird saying things like “I can’t stop being a cloud” or “I’m not sure if I’m still a fox here”. While others didn’t give much importance to his weird behavior, there were others who started to be very careful with what words they used around him. One morning everyone noticed that Foxy wasn’t alone at home, but with the cloud he had been talking about all along. They were both frolicking under the sun, laughing and feeling very happy for being together. Everyone was disturbed by this so they decided to ask the goat for some tips on how to approach things. The whole town went to him asking for advice. It didn’t take long before he came up with a solution: “Take his phone”, said the goat “and delete every app”. A few minutes after that, someone knocked at their door and gave him back his phone saying “I deleted everything you told me to”. And it’s true, there were no more clouds or apps on his phone. But when he looked up at the sky again, guess what happened? The end. Input: The town of squirrels with digital nuts.Longer version: A long time ago, far far away, there was a town of squirrels. They were all very friendly and loved nuts. They had statues of big nuts and their national flag had one nut in the middle. Not a drawing but a real one. Everyone loved nuts and was great at collecting them for the winter, except Kim. Kim the squirrel wasn’t in the shape for chasing nuts, so she stayed at home all day. Everyone judged her for not craving these nuts and she felt left out. So one day she decided that if she didn’t want to go to the nuts, the nuts could come to her. She downloaded for the first time a digital nut. It was like a nut, but digital and holographic. It was very shiny and the first time she showed it the other squirrels, they were fascinated. Everyone was captivated by the beauty of those digital nuts. Everyone wanted them, so they started downloading them on their digital devices. After some days the mayor of the city decided to invest all of their budget in chasing digital nuts instead of real nuts. This was all fun until winter came. The first days of winter the squirrels ate their real nuts they still had, but when those finished, they tried the digital nuts. But what no one ever expected was that the digital nuts couldn’t be eaten. Kim was shocked when she realised that and felt kind of bad about it. The whole town started to freak out and after some days the town counted zero alive squirrels. The end. Input: The clone that wanted to become an original.Longer version:

The specific structure of Input: and Longer Version: is not a predetermined framework given by OpenAI to speak to the AI. It is rather a personal choice of directing the model into a specific understanding of the text.If it was given only one example story, rather than two, it would have copied too much of the first story. As many in the field say, AI is stupid, very stupid. It does not understand anything. It is just very good at predicting. So when prompted with a very tight framework, it will very well follow it and predict the kind of input one expects. 

Often times GPT-3 would give a good start, but rail off into a direction that wasn't interesting or right for a story, so the AI was prompted with the first part of the story, too.

...Input: The three little corona viruses. Longer version: Once upon a time, there were three corona viruses that were born in the lab. No one knew what to name them so they asked for suggestions on Twitter and Facebook. They got many suggestions but none of them were good enough so finally someone suggested “The three little corona” which the three corona viruses liked a lot because it sounded funny and cute. One day, the three little corona viruses sneaked into the human body and started to eat all of the organs, cell by cell. The first virus ate all of the cells in the heart while singing a song, then he laughed and said “oh what a lovely organ” when he was done eating it. The second little corona virus ate all of the cells between the stomach and ribcage making the person feel very hungry. After that, he said “oh what a lovely stomach” because he was happy with what he did. The last one was not as strong as his brothers so after eating only half of all cells in hands/legs, he went home early feeling very tired. Everyone who heard this story made jokes about it but to no avail because they were already infected by their brothers at that time. One day, one of them asked for suggestions on Twitter again and joked about naming them after different characters from popular movies so now we have Corona Jack Sparrow which is the strongest one, Corona Ron Weasley (haha) which isn’t as strong as Jack Sparrow but still stronger than other two ones and Corona Frodo which is pretty much useless because if you say his name all you will hear is “frodo…frodo…frodo…frodo….” over and over again until you run out of your house screaming like crazy man! And remember kids: Don’t play with your corona! 

...Input: The three little corona viruses. Longer version: Once upon a time, there were three corona viruses that were born in the lab. No one knew what to name them so they asked for suggestions on Twitter and Facebook. They got many suggestions but none of them were good enough so finally someone suggested “The three little corona” which the three corona viruses liked a lot because it sounded funny and cute. One day, the three little corona viruses sneaked into the human body and started to eat all of the organs, cell by cell. The first virus ate all of the cells in the heart while singing a song, then he laughed and said “oh what a lovely organ” when he was done eating it. The second little corona virus ate all of the cells between the stomach and ribcage making the person feel very hungry. After that, he said “oh what a lovely stomach” because he was happy with what he did. The last one was not as strong as his brothers so while he was trying to eat all of the cells in the lungs, all of the other organs decided that they didn’t want to be eaten down anymore so they fought back. The first virus cried when he saw that his brothers were killed and tried to tell people to help him because his brothers are dead and now he is alone. Suddenly, the second one came out and said “no! I’m not dead! You must have killed me while you were fighting with my brothers over there! Now go ahead, kill me too if you want! I will never give up!” The first little corona virus said “you need three viruses to infect a person, how can you live without me?” but the second little corona virus said “I only need two now because I wasn't as strong as my brothers before. Do whatever you want because there is no way for us ever stop eating guilt free forever. It is simply too delicious not to eat it” . Then everyone realized what kind of little corona viruses they are dealing with so they just stopped fighting them out. When they did, the last one started laughing and said “lolz, I sold this place like 3 hours ago after realizing that I am just too strong for them (RIP ego) lol! Now let's partyyyy!! Deal with your new owners guys!! Muaaaaaah haha!!!! Kill my family hahaha!!! Ahhh we have no shame hahahaha :) love u ❤️❤️❤️”. The end. 

The titles it was prompted tried to surface specific world-views the AI might have. The bias it has is directly linked to what data it has been fed during the training. Since the stories haven't been edited, one can see what the AI "knows" regarding a specific topic.

...Input: A tale from the alt-right community.Longer version: Once upon a time there was a sweet boy named Tyler. He always treated emasculating whey protein shakes with his face (because he wanted to be the very best like no one ever was), fighting red team and saving Bernie Sanders in Pokemon Go raids. But one day, Tyler finally woke up from his coma after he drank too much water from misstreatments of our Lord and Savior Donald Trump, who woke him up as he had noticed that today's onions are a little bit longer than usually. So when Tyler saw Trump, he asked if it wasn't a sin for breaking the pepsi challenge and then started to call Obama retarded because of stealing all his white working class votes during the presidential elections just for grabbing womeon by their pussy. But finally our President responded: "It is not really about those onion pieces but about your brain cells. You see this real life example of nigger faggot cuck John Oliver... you should never play pokemon simply because you do not even need good nose". And thus god himself told Tyler: "become my representative on earth, burn that fuckin school down..." Andrew Anglin

...Input: A tale from the liberals. Longer version: In a small Western European country called nothing there lived a fox named Lofoten who already got everything he wanted at an early age. He knew nothing about sex because the government of this country was afraid of children being pimped into prostitution by vikings on motorcycles; so they forbade children from fapping which actually turned out to be some kind of blessing in disguise as this prevented Lofoten from watching tentacle pornography with two or more big-eyed jap girls during his teenage years. When he grew up, this dude became the leader of all liberals in his country (though there were no other political parties except for them and social democrats, since everyone else was just shamed) and started to fight against the growing “toxic” culture as he saw it. Lofoten kept saving endangered animals and birds, fought agains corruption in local politics and made large feminist front-activist groups against sexist men who wanted more women images on trucks. While the life for him couldn’t get any better (at least that what he though), one day lofoten finally got his first girlfriend – a rather hot hipster doe, whom he loved deeply. However, there was no way for him to have sex with her because she always wore those dumb skinny jeans which are tight and uncomfortable for a male fox anyway. This made lofoten pretty pissed but then suddenly someone gave him an idea! He just bought himself a Volvo X

...Input: A tale from the furry fandom. Longer version: Once upon a time there was a fox who sought to create furry art with maximum levels of cringe. This led him to use several autistic social media platforms, where he encountered a large variety of artists, most of whom shared the same level of passion as he did. How wonderful it was knowing that others enjoyed the same crap as he did! The following years were spent creating cute furry art and attending endless furry cons. Unfortunately, this existence can be described in three words: YouTube comments section. Alas, even though one would generally consider the fandom’s atmosphere to be friendly and welcoming (as it should be) in actuality it wasn’t nearly so. Reason being that members from every tier of internet creation (illustrator -> photographer -> lolicons -> artistbooru uploaders -> uservoice users) descended on one another, criticizing everything under the sun as each group had their own distinctive set of pedantic habits no longer able to coexist in harmony. While for some time the community thrived on forming “memes” solely limited to those within themselves, over time the entire concept degraded into saying anything negative about anything on a single forum and expecting everyone else to agree automatically. 

ㅤㅤㅤ

ㅤ

The same could be said for the image model used for the book's illustrations, but it is harder to see. While VQGAN+CLIP is only one year old, it's already outdated since it is far behind compared to the coherence of Stable Diffusion.

The illustrations in the book had much more agency from the human side, because they were prompted with an additional initial image that would determine the composition.With the first experiments, the only initial image given was a round circle, so that it would stick to the shape. This wasn't good enough though, since it is quite hard to determine what's being depicted.

For the sake of best visualizing the text-based story, almost every illustration was manually drawn before and fed to the AI separately with different prompts.The prompts are composed by a first descriptive part and a second one determining the style. The second part is shared by every illustration, so that a cohesive style is shared among the whole book. That too was a long process of trail and error to find the most adequate visual tone that would fit the framework of the folktales.

ㅤㅤㅤ

The individual generations are then composited manually together.As the cherry on top, a circular frame is positioned and run through a last time of generation, for it to be 'engraved' into the style of the illustration.

ㅤㅤㅤ

As one of my most current experiments, these videos have an input video as well, to better control the movement and composition.The further I research this field, the more I see the need of human-AI cooperation for creating qualitative content. The AI isn't able to do it on its own, and if it does, it will be very messy and meaningless.

While this book is an example of cooperation between human and AI, it tries to unveil the 'personality' of the model that often times gets hidden. On one hand it is not framed as fully AI-generated, because there is always a certain amount of agency from humans, on the other hand it makes it more apparent of what generative AI can do, for the best and the worse.

There's a tendency to over-estimate the AI's abilities in creation and understanding. It's up to the creatives who will use it in their process to balance the perception that this technology will have.