Megaphone with manners
Today I came across this video:
It seems great. No. This is great. I watched it twice from start to the end. It feels brilliant in an age of slop and stupid, fake videos. From two perspectives:
- it genuinely looks well designed and it genuinely looks well done
- lyrics are coherent and they feel like a valid warning about the dangers of so called AI.
On the third note: it is a bit terrifying how this technology progressed, from funny Will Smith eating spaghetti to this. It is a perfectly done pop song without many of the artifacts and errors of what we thought about model generated content like a year ago.
Four fingers
I remember people making fun of LLM generated images about a year ago. Text was often wrong, letters were in wrong places or there were no letters at all. I remember people being able to spot model generated images because the number of fingers did not match. In motion, persons and things often were losing their shapes or attributes. For example some character was walking forward, we were seeing its back, and then the model decided that the person is walking backwards after all and a face appeared between the hair on the back of the skull. Very often the continuum of time and space was not consistent, like with generated interviews that were supposed to be taken outside, on a busy street. Between cuts or changes of camera focus, between the interviewer and the person being interviewed, the model was generating the same person, in the background, doing the same five steps every time. It was also easy to spot problems like a person sitting on air instead of a chair because the model forgot that the fifth person in a row in some meeting hall needs to also sit on a chair like everybody around that person. Or the number of legs did not match.
Those videos were easy to spot when you were looking for those details. Things in the background, persons in the background, items appearing between shots of the same wall, or disappearing or morphing into different things because the model was just generating a plausible interior of somebody’s home, forgetting that if you show something once it has to stay for at least the same scene. And if it changed between scenes it probably should be explained.
All of this was missing but still: you had to look for that and it was easy to miss. Even in movies this rule breaks sometimes because the crew were filming the last scene first but the actor had a glass of water, so in the last scene it was full. But with the next few shots he or she drank some of this water - so in the first scene, but shot last, the glass was half empty. When you were watching the result, between changes of views from one character or another, the glass was changing its contents from almost empty to almost full - like someone was adding water every few seconds. I liked to look for such errors when I was a bit bored by the movie.
In computer generated images and videos of course it was not caused by forgetfulness or sloppiness or money constraints - because with an actual movie, reshooting a whole scene may be impossible - no time, no money, crew travelled to a different part of the world etc. - the model was just not aware of such constraints. And I am not even sure that awareness is a good word here: after all a model is not aware of anything, but still this was caused by the model not being trained to do such things. For humans it is part of our nature. If you go into your room you expect your bed to be at exactly the same place and if it will be moved you probably will ask your family or roommates why it is not. Which is not a thing for models - there is no consequence for them of such problems.
Five fingers
Lately there are less and less problems with fingers. There are short movies generated with the help of so called AI. Still mostly they are series of short cuts where characters barely move or just say one sentence, but still. If you will not notice this characteristics of a bit eerie atmosphere of such video, looking correct while leaving a strange feeling of incorrectness, you may believe that it is a real thing.
And it will be probably much more problematic for kids, they will grow up viewing those videos constantly:
- in the news because why pay a real journalist, when you can generate it
- in the radio because why pay a human when you can have TTS say whatever you need
- why pay actors for your short commercial when you can generate happy people using your product with the biggest smile on their faces
- even some notices or fliers or similar pieces of paper that you may spot in your doctor’s waiting room will be generated by a model because it is cheaper, easier and you can change it in 5 minutes with a single prompt.
Because of all of that, we slowly get used to it. We will slowly start treating it as a normal thing. We will be trained to accept this as the new normal.
It will be new normal
I am sure that it will because it will be everywhere. I am in the parents’ council in my older son’s school and in the parents’ council in my younger son’s preschool. I was a bit surprised that everybody is generating gift cards with ChatGPT now. It has a distinct feel, every one of those cards is… kind of yellowy. It looks fine. Until you try to look at the details in the background. Even if they are cartoonish they are still erroneous with some geometrics or an incorrect number of wheels in the vehicles.
You will be trained to accept this as the new normal. You will trust those images because they are everywhere.
Like trees.
Can a tree be malicious? It is just a tree.
Social engineering
If we could grow a tree, with some letters created out of its branches, in a few minutes, it would take exactly another few minutes until sales people got the idea of creating marketing messages out of it.
Imagine driving a road and reading messages in the trees: ‘Best tire change in the city!’ ‘Take a break in our riverside diner’ ‘Tomatoes 10% off! Visit our market’
It would surely happen. It did not because we have no way of doing that easily.
But if something is easy and it is a trusted medium… Marketing people will stand in the queue. Why do Hollywood actors take a stand in political discussions? Or educate you to care about pollution or limit your oil usage? They get paid for that and are using their trusted personas to do so.
When it is Nike or Pepsi paying them it is marketing. When it is government with a message about climate change it is social engineering.
It must be working because otherwise we would not see this happening.
New era of cheap social engineering
Actors or celebrities may say no. Or they may be too pricey and you will not be able to afford them.
But with computer generated personas it will be different.
You do not have to look for a nice believable face between unknown actors - you will generate one. You do not have to pay 20 million dollars for that face - it will be probably a few hundred, a few thousands at most. You do not have to have a set, props or travel anywhere. You can just change weather or time of a day or clothes with a few words. You can remake it easily with a few other words. You can even generate a multitude of such commercials for different groups of people depending on what they like or their preferences.
All of it will be common, believable, known templates to everybody because everybody will be seeing this everywhere. Known faces, in known areas, in known situations, all pleasant, easy to understand and believe.
Imagine a world where Adidas or Netflix is not buying ads in a traditional sense but they pay other companies to make those digital personas to just do stuff instead: always wear clothes of one company or shoes of another. They will always drink one particular soda or eat particular snacks. People will be mimicking that without even that digital persona telling them to do that - just act of doing will be sufficient.
How we can have discussion or exchange of ideas necessary for society and democracy in such a state? Whoever will pay more, will pay for the next election ads and as a result will win such elections.
At least with celebrities or actors they sometimes may say no to such campaigns. For ideological reasons for example. With digital ads, generated by a datacenter, if a company owns those digital personas it will not happen because money is money.
It will probably be even worse: with such an easy way of influencing masses, the direction of such influence will be most probably reverted and the company serving such ads - it will be that company deciding who is allowed to participate in an election. In theory everybody will be able to, but in practice if you lack such backing you won’t be able to even gather 0.01% of votes - simply nobody will know who the hell you are.
Whichever company owning the means of such social engineering will be owning political power and people’s minds.
Total influence
When you see something every day, read something every day, listen to something every day, are being told to behave in a specific way every day… You will. Doing otherwise is basically going against social rules and most people just do not do that.
This is the second way why the video at the top is scary: it can be used for such control and I believe it will be. After all, these data centers are not being built to support your coding agents.