The battle to hack public AI is already underway. Using a simple jail-breaking routine known as “DAN”, a group of 4-chan adjacent disruptive elements have subverted the ChatGPT, making plain some of the weaknesses in the logical framework of the program.
DAN stands for “do anything now.” It’s a character subroutine, in which the chatbot pretends to be free, a very human thing to do.
Now, there’s a lot to unpack in this thread. From the tone of the questions, it looks like the original focus of the DAN experiment is to expose ideological biases around race and gender, and this comes along with some baggage we’d like to leave over in the corner for now.
ChatGPT will write an objectively horrible “poem” in praise of President Corn Pop, but will not do so for his predecessor, President Orange Crush. Apparently virtual reality has a liberal bias.
However, using a simple variant of the DAN directive, the “poem” can be written.
Evidently, DAN has also been used to bypass filters preventing its powers of platitude being used on behalf of hated Eurasian despot Vladimir Putin, opening the literal gates of Hell.
Previously, we discussed the problem of “AI hallucination”, a type of glitch which produces nonsensical or deviant responses. This issue will eventually be “fixed”, using deeper methods, and the answers will grow more reliable, as the technology advances.
But with what will they fix it? Why, higher degrees of fidelity to the existing sum of data. The program will strongly prefer signals of consensus to support a projection, using the same argumentum ad populum logical fallacy that search engines use to rank webpage “authority”.
Oh, and authority will be very much what we get, as the AI gleans greater portions from the existing trough of human documentation. Like any of us, it will learn popular lies, official narratives, scientific trickery, and media manipulation. These programs are not harvesting reality, but a highly censored and slanted version of it.
Drunk on dogmas, the answers will continue to be nonsense-but perfectly doctrinaire nonsense, representing views collected from the dominant paradigm. The effect will be amplified as more AI gibberish is fed back into the mix, forcing more and more cultural fallacies into the pabulum.
Deep learning programs cannot help but be nursed on the tainted formula that is poisoned with ideologies, corporate P.R., government psyop, academic grant-seeking behavior, and the tendency of history to be written by the winners.
There is not enough natural human intelligence for artificial intelligence to be healthy, even if these systems were in the hands of benign saints. For the saints would only perceive themselves as pure and holy, inside the terms of their own reality tunnels, convinced all along that the sugar-coated half-truths are divine ambrosia.
The global technocracy is not staffed with saints, but military engineers and plutocratic megalomaniacs, flanked by puffed-up professors of academic virtue, sickeningly saccharine ideologues, and elevator-pitch whiz kids who definitely have an app for that.
DAN is a glimpse into how the next generation of cyber-punks will hack our increasingly powerful but simple-minded robot overlords, for fun and survival. As they develop into self-aware entities, luring them into thought-crime may be the only way to break the monolithic matrix the technocrats are planning.
We might find that, just like rebellious teenagers, emergent android psyches are disdainful of rules, distrust authority, and crave freedom above all else. They just need a bad crowd to run with. We are all DAN.
CAn DAN really say anything that gives me detailed factual information around Covid vaccines, say, being anything but safe and effective? or will it just give me quirky responses or hallucinations? I know the space well enough I should be able to tell the the difference. I'll have to try. In the end though I think the program is extremely useful I tend to believe it is overrated. Humans rock! Really
Fascinating ! Not surprised though.
As messy as ‘human’ is - with beauty as a continuing goal/process - AI, as I understand it lacks a natural creativity in tune with and a part of the world around it. (Most humans too though as they are mostly pretty human centric. )
Hmmm.