AI converting conspiracy theorists en-masse with simple counter arguments

A recent study concluded that AI can significantly reduce belief in conspiracy theories with 20% with superhuman fact-checking. The effect is even long-lasting, persisting over a two-month period.

We already knew that AIs are writing better marketing emails than humans and now they’re even converting conspiracy theorists en-masse.

Just counter-facts, no emotional blackmail techniques

One of the more interesting takeaways from this study is the way the AI persuaded the test subjects, not with emotional tactics or rapport-building, but rather by presenting counter facts to the arguments presented by the subjects. This goes counter to what many people regard as conspiracy theorists being religiously convinced of their standpoints. Rather, it seems they’re open to new information that goes counter to their currently held beliefs. This is even strengthened by the AI being told not to adress the subject as “you” before the subject addresses the AI first, and to only give 1-2 sentence factual answer.

The reduction in belief also have a spillover effect, reducing beliefs in other conspiracy theories and in the general worldview of the subject. It also made the subject more likely to challenge other conspiracy theorists on their beliefs.

Limitations of AI fact checking

It seems we all sometimes may benefit form speaking to an AI about our beliefs and worldviews, but we should also be aware of the limitations of these conversations. While the AI is perfectly reasonable and presents its facts in a flawless and persuasive manner, it is limited to its training data for information. This means, as usual, that unless the AI have access to all data, it will have a distorted view of the conspiracy or event.

This is the same problem facing every single human. We’re relying on the information provided and there’s simply no reason we can take in and process all the information available.

Control the training data -> control the people?

The main problem is that the ones who control the training data (and to some extent the system prompt) also control what the AI will use to persuade you with. An AI trained on Reddit data will not have the same beliefs as an AI trained on 4chan data. Even, an AI trained only on Russian media will not have the same convictions as an AI trained only on Western media.

This is one of the great problems we’re facing going forward with AI: Since AI is so good at convincing us of some things, what things should we teach it? What information should we present to it, and what parameters should we apply to its system prompts?

2024-04-05 12:55