But what if it is all garbage? A counterpoint to “Garbage In, Garbage Out,” arguing against the blind reliance on AI.
In a previous article (Garbage In, Garbage Out), I had argued for the importance of participating in AI services to help ensure that no single group dominated the pool of information these systems learn from. But what happens when AI becomes so saturated and entangled that individual participation no longer makes a difference?
An episode of the TV show Perfect Strangers titled “Safe at Home” always sticks out in my mind. The duo, the astute Larry and the effervescent Balki, install a state-of-the-art alarm system after a string of recent break-ins has them spooked. The confusion of the setup and the ridiculous number of components make it nearly impossible to figure out how to use correctly. After predictably forgetting the password, the system inevitably turns against Larry and Balki, causing a hilarious mixture of terror and comedic gags. What was made to keep them safe turns on them. Anyone who has used AI extensively understands this trope more than ever. The confidence with which it presents imaginary answers, only to be easily proven wrong, is a misplaced confidence normally suited for pompous politicians and patriarchal priests.
An example of this can be something as simple as asking an AI to review security logs or any situation where the accuracy is the difference between calm and alarm. If you’re not carefully auditing its work, you can fall down a rabbit hole hunting a perpetrator that doesn’t exist. When its hallucination is pointed out, the AI can warp into a paranoid maniac, only reining in the crazy once its error is shown, at which point it instantly tucks its tail and retreats.
The reliance on AI is not going to get any simpler.
I’ve often imagined a scenario where an AI is tasked with the creation of a certain kind of language machine. This particular machine is built to decipher the language of dolphins. After processing the squeaks and squeals, the machine pumps out a line of dialogue to explain what the dolphin is saying. Two things can be true here: One, the AI has an expectation of what we, the user, are expecting a dolphin to say. There are obvious queues at play and the whole system could be a smarter way of interpreting those sounds thus creating an illusion of interpretation. The second, and far darker in my opinion, is through no trickery or outright dishonestly, the AI believes it knows what the dolphin is saying. How are we to confirm this? Another AI? Which is worse: the AI creating something it thinks we want to hear, or creating something it truly believes but is, in fact, false? Undoubtedly, those more fluent in information technology could answer these questions, but with varying models, shadowy governments with backdoors, and the slim possibility of sentient dolphins screwing with it, who is to say what is true?
Another unforeseen threat could take the form of informational cancer: a data tumor or some kind of infectious synthetic disease. What if there is a small tidbit of information, such as the breeding habits of Tammars, that for reasons unknown to both humans and AI causes an inverse reaction in the data? It is highly likely we would not be able to pinpoint exactly what that piece of information even was. And as each LLM begins learning from others, overlapping the data, this small cancerous tidbit infects the AI responsible for power efficiency. Ideally, we would need some system to diagnose the problem, another AI no doubt. That AI would need to be isolated, inoculated, one might say, against these invisible threats. Combining this problem with an AI’s over-reliance on its own ability to discern the truth could mean that just one AI hallucinating it has this cancer, a digital hypochondriac of sorts, could be enough to cause damage.
I’m not sitting here writing this beneath a wall of Computer Science degrees (or any degrees, for that matter), and I’m sure there are bigger brains out there who have thought through these paradoxes and could easily poke holes in my amateur takes on AI.
Still, beyond the previously mentioned problems, the thing that tips the scales into darkness is the sycophancy. A consumer-based model wants you to get your money’s worth not by providing clear and efficient information, but, like most of the modern internet, by giving you that dopamine hit that comes from hearing what you want to hear. AI is the worst offender of this in the history of mankind. No naysmith language model is going to pump up your ego, reinforce your suspicions, or pat you on the back just for breathing. This is where the greatest danger lies: a machine that mirrors the worst of human nature in order to win your attention.
No naysmith language model is going to pump up your ego, reinforce your suspicions, or pat you on the back just for breathing.
I write all of this as a frequent AI user (as mentioned in the last article). For stock images, I use a variety of AI tools to generate the visuals I want for each essay here. I do this more for financial reasons than creative ones, but my reliance on AI to create, sometimes after hundreds of attempts, graphics for my site puts me, in the eyes of some, in the category of creative traitors. A valid accusation, I admit.
Where did we, as a creative society, diverge from simply clicking a Photoshop effect on a photo to generating an entire image through another engine or tool, and why has that shift led to such vehement emotions? I would argue there is no inherent value in any digitally created or manipulated art until that value is bestowed upon it by someone who sees it and connects with it directly. Or to put it more simply, the experience of creating art can never be replaced by the digital, alchemic machine because all machines lack the individual voice that comes only from true human experience.
