DebunkBot

Gave it a shot. I tried to use actual beliefs of mine. The AI responds very diplomatically, with a hint of ass-kissing in one of my trials. That makes sense, I suppose, since most people with conspiracy beliefs have to be handled with kiddy-gloves, but it was a bit irritating if you're open to an actual dialogue.

For an AI trying to help break down conspiracy beliefs, I found its arguments quite weak and almost tautological. For example, one of my beliefs was "The CDC was pressured by corporations into hastening a return to 'normalcy' during the pandemic", and I noted that the studies used to support a return to unmasked-schooling felt cherry-picked, based on studies in higher-income schools with smaller class sizes and better ventilation than the average school. The AI just replied along the lines of "the scientific peer-review process helps mitigate factors like cherry-picking", which (A) doesn't actually prevent cherry-picking, it just means you need to pick more cherries, and (B) just leads us to the next issue, which is that the 'scientific peer review process' is already proven to have issues with replicability and the like. So, I questioned that, and then it just kept reiterating the same points, pretty much.

In my second trial, I chose a belief I felt less strongly about, but still felt reasonable: "John Barnett was assassinated for whistle-blowing on Boeing." This time, the AI's defenses basically boiled down to 'assassination is a risk that businesses and the US government would not undertake, because it would be illegal and breach the public trust'. When I noted that corporations and the US government both routinely engage in illegal actions that breach the public trust, it basically had no actual defense for that. It just kept reiterating that it's "good to be skeptical" but that we can't be skeptical about everything. It seemed to presume that all actors are rational and good-faith in its arguments. While conspiracy theorists often do presume too much competence and intelligence of people, you absolutely cannot trot out "They wouldn't break the law because it would be illegal" as an argument without it immediately devolving into tautology.

Basically, any amount of actual questioning of its initial defenses just lead to pretty much the same responses being regurgitated. Much like an actual debate that I'd imagine between anyone believing in a conspiracy theory, and anyone trying to refute it, I felt the arguments got circular pretty quickly.

I ended up talking to the bot about the unliklihood of it convincing someone who was truly invested in a conspiracy theory using the strategy it was employing, especially over the course of such a brief exchange.

It acknowledged this and offered legitimate strategies for dealing with an invested, potentially defensive person, but it didn't really take it's own advice and ultimately ended up repeating itself (as others noticed).

I think the experiment is sort of broken in that, even for a well trained human, a three message exchange wouldn't even begin to be enough time to establish trust and common ground before carefully starting to introduce evidence or lead them to think critically about their own assumptions.

I wonder what the data from such a test could be reasonably used for. It certainly won't tell us anything about how effective an AI would be at changing minds since the exchange isn't long enough to have really any impact on strongly held beliefs.

I tried discussing a relatively serious conspiracy theory with this bot, and it quickly became rather unfun when I realized it seems to be programmed to offer any alternative explanation for the theory you explained at the beginning. I wish I could go back and discuss something more fun with it. I'm going to try again with something less serious. Next time I try this, I'm going to discuss the secret life of the Thane of Whiterun with this bot. I'm pretty sure that guy works for the Dark Brotherhood and the Riften Thieves Guild...

To assume that all these incidents point to the Thane being a member of the Dark Brotherhood or the Riften Thieves Guild requires a certain degree of speculation that goes beyond the evidence at hand.

Talking to this machine about my suspicions of the Thane has really illuminated how it's programmed to cast doubt on literally anything you talk about with it. I'm going to have to think of something with harder evidence than "this guy walks around in Dark Brotherhood attire talking to the guards about Sithis all the time"

Unsurprisingly, it wasn't able to convince me that Geo. W Bush & the gang didn't knowingly mislead Congress and the US in order to drum up support for the Iraq War. What's interesting is that the arguments it used seemed directed more toward establishing doubt rather than presenting refutations, almost as though it were approaching it as a legal defense rather than a compelling argument. It really had nothing to say when I acknowledged that the internal motivations of others are impossible to know with certainty, but that the effects of their actions suggested greed as a motivating force.

This comes down to the basic problem with LLMs: they are incapable of anything but statistical analysis of the linguistic tokens they work with. In other words, they can't evaluate the strength of individual arguments, they can only work with the frequency with which those arguments get used in their training sets. It's related to the Strawberry Problem: many models are incapable of recognizing that the word 'strawberry' has three R's because they aren't evaluating the word 'strawberry' in itself, but rather the frequencies associated with its token. It's the Chinese Room Paradox writ large (I think I'm using that correctly for you, @updawg, though I expect you to correct me if not).

I suspect that the people driving the development of these models are hoping that improving the statistical analysis of these tokens will somehow bridge the Chinese Room gap, but I think it'll take some kind of conceptual leap forward before we get to a point where the true evaluation of semantic meaning can happen. It's funny, because I both look forward to that development, because AI will only have very limited utility until that happens, and I fear it, because that's the point where true AGI development will begin, and once it does, things will proceed very quickly, I think. Too quickly for many aspects of human life to keep pace, I suspect.

I'd be interested in trying again with a conspiracy theory that's established to have been true. One that springs to mind is MK Ultra, in which certain groups within the US intelligence and military performed experiments in mind control on unsuspecting subjects using psychedelics and other psychoactive drugs. It's still a popular conspiracy theory even though the bare bones of the conspiracy are confirmed, it's just a matter of the scale and success of the project rather than the fact of its existence that still gets debated. I can't put my finger on just why, but I think this distinction might venture close to that evaluation problem I've been talking about. I suspect that it will acknowledge many points of fact I might raise, but make a vague "one must remain skeptical" argument that doesn't really accomplish anything. I don't want to complicate the MIT team's dataset by performing experiments on their experiment though.

I tried to give it a shot but it got boring. Plus using it in mobile wasn’t good. I wanted to go back one step by clicking the back button and it restarted the whole thing. Frustrating.

Heads up, this doesn't work on Firefox for some dumb reason, so you'd need to use a Chromium-based browser. I'll check it out later, but hit that snag getting into the chat and lost short-term interest.

It's an interesting idea to engage with an AI, and without giving too much away, they also seem okay with you testing the AI if you don't believe something, but I would encourage going in with anything you believe to see how it goes.