AI Watermarking, Algorithmic Conditioning, and Superego Capture
Page Rank
Inventions don’t spring out of nothing, they are the result of a problem and of a mind with tools. When the two founders of Google invented Page Rank (their contribution to the improvement of search algorithms), they had a very clear inspiration from their everyday life. Back in the day, just as today, Internet searches had grown increasingly irrelevant, because web sites wanted to direct as much traffic as possible to themselves, regardless of whether users were going to find anything relevant on them. Since web searches in the day relied on what web sites claimed about themselves, adding a number of interesting keywords made them relevant to a lot of searches.
This was similar to a problem academia had had to solve: how do you decide if a researcher is any good? Until no too long ago, the number of articles published by the scientist had been taken as proof of the professional’s quality, but that just forced a lot of irrelevant publication, under the heading “Publish or Perish.”
Academia being full of smart people, they figured out something: it mattered less how many publications you had published and more how many people had cited your publication in their own. Since you needed other people to do the citing, your own opinion about the quality of your work mattered little, and the direct opinion of other researchers (who might be jealous and hence biased) was not a factor.
The founders of Google realized that they could adapt the solution to the problem of web searches. Instead of taking citations in scholarly articles, they would take web links. It didn’t matter if your site said it was about, say, knitting - it mattered that other sites about knitting linked back to it.
This connection between an innovation and the creators’ knowledge is well-documented throughout history. Even the most innovative people take inspiration from their immediate environment, and sometimes we can in fact reverse the relationship and read the inventors’ state of mind by the type of invention made.
AI Watermarks
You may have read that AI companies have standardized on a way to create watermarks in text produced using their tools. These watermarks are completely invisible to the naked human eye: the text produced doesn’t look in any way unusual and might as well be unmarked for as much as we can tell. Only an AI company would be able to tell the difference, which is very much unlike other watermark types, that are much more explicit and visible.
The fundamental idea of AI watermarks is very ingenious and sneaky. The exact implementation is irrelevant here, but the generally idea is simple: Large language models (LLMs) generate text one word (token) at a time. For each word generated, the next best word is guessed. Words get ranked by likelihood of being the best next match. Normally, the LLM would pick the word with the highest probability and put it in the text.
Imagine now if the algorithm decided to put the second-best word next, for every so many words, especially if the scores between the best match and the second-best are close. While the best match would approximate what a very skilled human would produce, a text with predictably second-best matches carries its status encoded. In fact, you wouldn’t just be able to tell it was produced using AI, you could probably even tell exactly which model generated the text.
There is one visible flaw with this approach, and that is that it takes a lot of text to be able to tell if something is generated using predictably imperfect matches or simply unskilled. If you have thousands of tokens, it becomes pretty apparent that something is regularly replacing words, while a short text would not even have anything regular.
(This is as good a moment as any other to point out that there is really no good reason to find out if a text was generated using AI, and by which model precisely, unless it’s because AI companies plan on charging for their services a pretty penny and the knowledge is a pointer to a bill that is going to be due.)
Algorithmic Conditioning
Why does this style of watermarking say anything about the creators’ state of mind? Because the idea of AI watermarking did not randomly appear in the minds of the people responsible for these algorithms. Instead, it’s exactly how algorithms work on social media. While you think the feed on your app is randomly tailored to you and your specific tastes, it’s actually interweaved with a “digital watermark” that pushes content onto you.
If you are an advertiser and you have the opportunity to push your content onto people without them knowing, that gives you a very powerful platform on which to stand. There is no outward way to find out that your content shows up in people’s feed - unless you are as heavy-handed as a certain site owner, known to push his own posts and those he approves of very blatantly on users of his media empire.
You may have heard of the Great Avocado Toast War: for a while, it was reported that GenZ-ers complained they could not afford anything, and their Boomer parents/grandparents complained that they could not afford anything because they spent all their money on frivolous expenses, like pumpkin spice lattes and avocado toasts. The Great Avocado Toast War saw more skirmishes, with GenZ proving they could in fact not afford anything because housing in particular had gotten too expensive. There was a well-publicized counteroffensive, pointing out that GenZ-ers relied on parents for housing, but a devastating blow was dealt when tuition costs were revealed.
In any case, the problem with the Great Avocado Toast War was that both sides had a point. GenZ correctly pointed out that “things”, meaning more or less everything, had gotten extremely expensive and unaffordable. But Boomers probably correctly pointed out that people who could not afford anything should probably not afford frivolous expenses.
Avocado toasts were probably not the best thing to point to. Instead, I would focus on things like eating out and especially food delivery. I know GenZ-ers that would think nothing of getting groceries delivered by app, despite living an honest block from a supermarket. Asked why they would do that, the reply was usually some shade of “I am worth it,” or “It doesn’t make that much of a difference.”
Which is precisely the kind of thing you’d expect an advertising agency to say. I had a strong suspicion that I wasn’t getting my friends’ answer, but that coming out of their favorite social media scrolling app.
Social Conditioning
In my previous article on this blog, I noted that humans must have developed a way to ensure our society is protected from individual selfishness. That is a requirement of all social species, but has particular needs in an intelligent social species, since we can’t hard-code all behaviors like other animals might.
To summarize the point, there is constant tension between the needs of the individual and those of the species, at all levels of society. The simplest level is procreation: a human infant cannot survive, but is also very time-consuming. So the parents must care for it, despite the fact it gets constantly in the way.
In more complex terms, you could look at a shared resource like a fishery: if everyone takes out as much fish as they want, there is soon no fish left and everybody starves. Society must create limits on the amount of fishing and must enforce them drastically, or it will disappear.
In modern times, you could think of the side of the road we drive on as an example. Clearly, you can drive on either side of the road with no issue. In fact, there are lots of countries where people drive on the right side, and lots others where people drive on the left. But each country chooses one side and sticks with it, because you can’t have traffic flows if people decide on a whim which side to drive on.
The mechanism that forces the benefit of society on the individual is called social sanction. It is an innate mechanism that is extremely powerful in humans and is located in the substructure of the mind Freud called the superego.
Social Sanction, the Superego, and Social Programming
Humans must have a system that forces them (us) to put the needs of society above their own, and this system must be powerful enough to resist selfishness and rationality in a critical amount of us. As Freud pointed out, the superego controls the ego, which is where logic and selfishness exist.
Social sanction needs to be programmed, somehow, as it consists of rules that are not innate. Nothing in us tells us we should have fallow periods in fisheries, and we would not be as versatile a species if our genes had to know that singular fact. Our reprogrammable intelligence must go down to the superego level to make us successful, and it clearly does.
But how exactly do we get these rules into us? If there is a mechanism that controls us for the benefit of our society, and if that mechanism is programmable - how do you program it?
It is probably not the only way to do it, and other intelligent species might have figured out a different way of doing the same (still needing something that overrides selfishness for the benefit of the species). For humans, Freud already pointed out that we have an animalistic substrate, the id, and a rational substrate, the ego. The superego is the third substrate, that we want to program. We can’t use the ego, because that’s what the superego is supposed to control, so we have to use the id.
The id is our animal self. It responds to basal stimuli, and those are the ones we have available. We can use physical sensation, emotions, and repetition.
Social Sanction Nonsense
This is where the circle closes. Because if we find something that people do repeatedly (repetition) and gives them pleasure (sensation), we can reprogram them over time. Take an app that shows you short clips of people, make it give you the pleasure of entertainment, give it an underlying message you want the user to import. Done.
The fundamental problem (and advantage) of this approach is that this kind of conditioning goes beyond the ability of the rational self to resist, because it uses a mechanism specifically designed to circumvent the rational self. The person that gets themselves reprogrammed can be convinced of absolutely anything, because the rules the programming needs you to follow frequently seem irrational by design.
In fact, some of the rules don’t just seem, but are absolutely irrational - meaning they are not even useful by themselves. They are just there to ensure that an individual can be programmed successfully, because some of the more important rules might not be directly enforceable. Checking an individuals willingness to stick with patently useless rules serves as a proxy for the important ones that cannot be monitored.
We see that commonly with many religious rules, that seem to come out of nowhere. Proscriptions on specific kinds of food, for instance, or fashion standards (dress codes or hair styling), are easy to monitor and of dubious value in isolation. They are also easy to follow, so lack of conformity serves as sign of lack of programmability.
It is easy to see why societies insist on punishments of non-conformity that remove the offender from the gene pool: humans that do not respond to social sanction are a threat to society as a whole, as their selfishness might actually result in a survival advantage for the individual that collectively might doom the group.
Social Reprogramming for Profit
We now knowa mechanism that controls the voluntary expression of a human and that is programmed using sub-rational strata. It was designed for the survival of the species and is, in fact, the reason the species survived.
We know how this mechanism is programmed, and that it only takes repetition and emotional charge to do the programming. Religions and cultures and states have used the mechanism for millennia for their own furtherance, always assuming that they were “society” at large.
What if now this mechanism is captured by corporations trying to sell you something? All the corporations need is access to the mechanism, to you, and a way to push their message.
Social Sanction and Allergies
Have you ever heard of anaphylactic shock? That happens when someone has an allergic reaction strong enough to (potentially) kill the person.
Have you ever wondered why humans can be killed by their own immune system? No matter what the thing is that the immune system is trying to protect you from, it can’t do worse than kill you. So why is the immune system allowed by nature to kill you, at all?
The reason is the human species as a whole. Your immune system may kill you, because your death might protect the species as a whole. Your peanut allergy is just an unfortunate side effect, the real enemy is the fever that kills you so you can’t infect more people with the plague. The fever is not generated by the pathogen, it is generated by the immune response.
That is precisely what is so dangerous about social sanction and its corporate takeover. It is powerful enough to kill you, and you’d be enthusiastic about it. Because hundreds of generations of humans only survived because they knew how to follow what they were told, no matter how stupid it sounded, as long as it was handed in a pleasant and repeated form.
Now you have a $1.000 device in your hand that gives you little dopamine hits while it convinces you to spend more money on things you don’t need. You will not just heed its advice, you will die on a hill for it, swearing that’s what you always wanted.