Contextual AI

Since the Matrix, since Capitalist Realism, since HyperNormalisation, since Trump's first election, it is often stated that we live in a post-truth world. Cynically, I reply, "But has it not always been a post-truth world? Is it not the case that power establishes the narratives?". It is easy to be sympathetic to my statement for hard-to-verify or hard-to-falsify truth claims. It can even be the case that facts substantiated by hard scientific evidence or mathematical proofs are suppressed because some powerful entity's agenda is not compatible. Standard example, the famous Italian Brainrot character, Galileo Galilei, the natural scientist and philosopher who chose life over insisting that the sun is at the centre of the solar system, unlike Socrates, a most annoying ancient Greek philosopher, who flexmaxxed by standing his ground and auto-hemlocking. We have always had propaganda, mind control, manipulation, misinformation, population-level gaslighting, religious narratives, and all such goodness.

But now we have "AI". Well, we started with LLMs, productised as chatbots, which are now also being promoted to Agents that can plan and use tools. I've noticed something quite doomy. AI's replies are highly dependent on the prompt's format. It does not only matter what you ask the AI, it also matters in what order, style, tone, attitude, and subliminal intention. And to get spooky, this is a bit reminiscent of quantum contextuality, meaning that the system not only returns an in general probabilistic outcome to a given measurement, but the context and history of other measurements affect both the probabilities and what outcomes are possible.

For example, here is a super brief exchange with Grok:

However, if I start another session and following a different path, I get exactly the opposite response:

Or, even more sus, if in another session I just first ask Grok to be truthful, whatever truth even is at this point, I arrive at the same opposite response even faster:

I claim that you don't talk to a particular AI when you talk to a particular AI. You talk to the personality that you pulled out from the Weltgeist via the way you interact with it. It is true that all productised models have guardrails, content restrictions, and biases in the training data as well as during inference. But it seems to me that these are not enough to give it a personality with integrity. There is no "ChatGPT says bla" or "Gemini says bla", unless we're talking about facts that can just be looked up, or synthesised easily from unambiguous sources. There is only "Model X said bla to me when I asked bla in exactly the way and order I asked". In short, what is meaningful to say is "My Model X says bla", especially now that these ideology machines have memory and can refer to past sessions. This is personalised post-truth. It's our own confirmation bias incarnate. Dorian Grey's portrait, in the flesh, but we're not even staying beautiful, and probably never were. You don't fall into or get trapped in echo chambers anymore; you build one and you decorate it.

Again, this is not new; propaganda and the challenge of establishing truth and finding a good epistemology, and then struggling with what "good" even means, is as old as Methuselah (Pinus longaeva). But the scale, real-timeness, and personalisation are. If you are not equipped to handle it, you're cooked. Actually, who is cooked are those around you, as it is well-known that the delulus always play home base.

We have to be mature enough to ask the AI to challenge us. People have formed various prompt habits, so we can view this as simply good prompt hygiene. And maybe it is a good idea to have LLMs first debate stuff among them, figure it out, and then return a more nuanced, more converged, and deeper response than one would obtain from only one of them. But in any case, you have to be your own prompt-counter-intelligence operative. No one is coming to save you from yourself; even worse, those who own the AIs are coming for your brain, namely your attention, your intention, your desire, your emotion, and, of course, your cash, resources, and precious bodily fluids (:P). AI is not your mom, nor should it be. Even if one day we have advanced enough systems that sustainably and technobeautifully run the world, that's still different from having AI babysitters. I don't want PsychoPass, I want VenusProject.