The Intelligent Friend - The newsletter about the AI-humans relationships, based only on scientific papers.
Intellibox - Issue 4
Hello IF readers! This is the fourth issue of Intellibox, the new weekly issue of The Intelligent Friend where, beyond insights from an interesting paper, you immerse yourself in a creative and intriguing challenge thanks to AI. Some examples? You are in a jungle and you have to find a particular animal species. Or you have to create a character, plan how to write your book in two months and much more. All with the help of AI. Today's issue is truly special. I start from a very simple question: who has never imagined, at least once, being the President of the United States? Well, today you can be, thanks to the simulation that I really enjoyed creating. All that remains is to say 'At your service, Mr. President!'.
This issue is part of the paid The Intelligent Friend subscription. But the cost remains the same. That's right, only 5 USD per month.
AI and political ads
As I anticipated, in this part of the issues I will always try to bring you well-structured insights, almost as if I wanted to get the 'gist' of the paper, given the much information I give you in the other issues.
Title: Evaluating the persuasive influence of political microtargeting with large language models. Author(s): Hackenburga & Margetts. Year: 2024. Journal: Proceedings of the National Academy of Sciences (PNAS). Link.
Since the simulation has a 'political' theme and the US elections are coming up soon, I was particularly struck by this paper that I want to bring to you today, in line with the topic of the challenge. It's about political ads created with AI, and I think it's going to get your attention.
The author explore the effectiveness of AI-generated microtargeted political messages. With the rise of AI, particularly large language models (LLMs) like GPT-4, concerns about their potential to influence voters through personalized messages have grown. This study aims exactly to quantify the persuasive power of such microtargeting - i.e. the selection and reaching, through advertising, of a super specific group of voters (or consumers).
Methodology
The methodology of this paper is fascinating: the researchers developed a web application that integrates demographic and political data into GPT-4 prompts in real-time, creating tailored messages for individual users on political issues. They conducted a large-scale, randomized controlled experiment with 8,587 participants who were exposed to one of four conditions:
Control group (no message);
Best message (non-targeted);
False targeting (incorrect demographic data);
Accurate targeting (correct demographic data);
Participants' demographic data were characterized by a wide variety: age, ethnicity, gender, education, religious affiliation, occupation, and so on. Furthermore, the study examined four political issues: digital privacy, China sanctions, NATO support, and renewable energy.
Results
You would expect that AI-created messages would actually not be very effective, especially in the 'wrong' targeting conditions, right? Well, that's not really the case. Both targeted and non-targeted messages generated by GPT-4 were broadly persuasive, increasing support for certain issue stances by up to 12 percentage points in some cases. However, the overall persuasive impact of microtargeted messages was not statistically different from non-targeted messages.
However, if we dig into the various issues judged by voters, there are several insights to consider: on digital privacy, accurate targeting did not significantly outperform the best message condition. While, on China sanctions, accurate targeting was less persuasive than the best message condition but not significantly different from the false targeting condition. On NATO support, accurate targeting was more persuasive than false targeting but less than the best message. On renewable energy, none of the conditions had a significant persuasive impact, possibly due to high initial support for the issue.
Summarizing, the findings indicate that the strength of these models lies more in their ability to craft generally persuasive messages rather than highly personalized ones. This challenges the widespread assumption that microtargeting significantly enhances persuasive power.
What does this entail? That when it comes to political advertising generated with AI, it's not just the target that counts, but the effectiveness also varies depending on the topic. This is an insight of incredible value, especially if its applicability could also be verified to other types of messages, such as those relating to consumer products.
The challenge: you are the US President!
Here we are now at the challenge or simulation of the week. I'll briefly remind you how it works:
I'll give you some context on the challenge and goals;
I'll give you the prompt;
I will tell you how, if you want, you can share the outputs.
So, let's start with this approach with the second exciting simulation.
P.S. I would like to specify that, as I have often said, I do not have a Computer science background. To create these prompts I used ChatGPT-4o itself, continually experimenting and making changes until I had a result that I thought was truly engaging to bring to you here. I test the prompts extensively, but if there is any problem, don't be afraid to report it to me and also suggest how to improve!
The challenge
I'm really enjoying creating these challenges for you, I have to admit, but this one has me particularly engaged. I have always imagined what a game in which you are the President of the United States could be and how the various reforms could have an impact on many aspects. Well, I set out to make this simulation really fun in my opinion. You are the President of the United States and have a set number of turns to implement your reforms. But be careful, the budget is fixed and everything you do has an effect on the others!
You are the President of the United States and you have 30 turns to act by implementing reforms of your choice and fundamentally acting on 5 fundamental areas:
economy;
foreign policy;
healthcare;
culture and education;
safety.
Each of these areas is also a command that you can use to return to a specific area. For each area you have suggestions of things to do or you can use 'ACTION' to propose your change. But be careful, everything you do is based on a 200 billion budget, and each reform may or may not have an impact on: 1) work and employment; 2) security and crime; 3) happiness of the population; 4) international relations; 5) healthcare; 6) social equality; 7) entrepreneurship; 8) education; 9) opinion of the media (CNN, NYT, Fox News).
Finally, if you really want to implement something extreme or different, you have:
Keep reading with a 7-day free trial
Subscribe to The Intelligent Friend to keep reading this post and get 7 days of free access to the full post archives.