It's a bit like mixing nitric acid and glycerin. I know.
I am on a minefield. I know.
I don't want to make any political promotion but just talk about a technical possibility. Be aware.
After discussing the difficulty of reading political programs, my two neurons connected and started dancing. Now, AI Language Models (LLMs) can work with many words, and there is no word shortage in politics. So, I thought... why not try to process some party programs ahead of the upcoming European elections and see how AI handles such long and complex topics?
This activity would have been impossible a year ago; the 'temporary memory' of LLMs was very short, and after a political program, usually composed of about twenty pages, the AI would have started to forget the beginning of the message.
But now we're ready. If you want a technical deep dive, read the paragraphs with the red pill.
🔴 Why is it technically feasible now? The most advanced AIs now have very large context windows, around 128,000 tokens equal to about 60-80,000 words (roughly two short novels), and they can perform 'inference,' meaning they can make articulated reasoning on all points of a text accurately, considering all aspects. (There would be a long, even more technical explanation behind this sentence, which I will spare you).
The Prompt
Without overthinking, I uploaded the programs of four random parties from their official websites and started a conversation with one of these AIs using this prompt:
I am a voter in the upcoming European elections and want clarity. I would like to use the old approach of reading and comparing the party programs, but I don't have the time. I know that you [AI NAME] and I may have biases, as do the parties themselves. I would like you to help me get as impartial an analysis as possible based solely on the content provided in the electoral programs.
If you want a more complete prompt, use this part too and maybe add your own touch. The important thing is to clearly explain what you want to achieve.
Goal: Provide a comparative analysis of the electoral programs of the main parties for the upcoming European elections. Context: Compare the electoral programs of parties 1, 2, 3, etc., on various key issues (economy, environment, social policies, etc.). Specific Instructions: Focus exclusively on the content of the official electoral programs. Avoid personal interpretations and opinions. Use neutral and objective language. Report only facts and direct statements from the programs. Output Format: Summarize each program in a separate section. Highlight the main points of each theme covered (economy, environment, social policies, etc.) Provide a comparative table highlighting the differences and similarities between the programs. Reducing Bias: Do not include information not present in the official programs. Avoid making deductions or assumptions. Specify when information is unclear or not present in the programs.
Here's an example output:
Party 1 Economy: [Summary of main points] Environment: [Summary of main points] Social Policies: [Summary of main points] ... Party 2 Economy: [Summary of main points] Environment: [Summary of main points] Social Policies: [Summary of main points] ... Comparative Table Theme Party 1 Party 2 Economy [Main points] [Main points] Environment [Main points] [Main points] Social Policies [Main points] [Main points] ... Ask me for the programs of the various parties before doing the analysis.
After listening to its 'well-meaning sermon' suggesting I read the programs myself, I started uploading the programs, more or less like this:
Here is the program of party 1: ... Don't give me answers until I ask you to. Now I will pass you the program of Party 2. ...
And so on.
At this point, you can do many things. If you use the extended version of the prompt, each program will still be summarized upon upload.
Direct Questions
Here is where you, your ideas, and your ambitions come into play. You can ask anything that interests you about one or more programs, compare them, and have the passages cited on which the response is based.
Remember to ask the AI to stick to the text inserted and make no assumptions. In any case, double-check.
Some example questions. Replace the text in [ ] with what you want and get creative.
Which party is in favor of [issue]?
Compare [Party 1] and [Party 2] on the theme [your theme] and give me evidence from their programs on how they intend to address [this issue].
"What are the main differences in the economic policies proposed by the various parties?"
"How do the various parties intend to address environmental issues and climate change?"
"What proposals do the parties have to improve social policies and reduce inequalities?"
"How do the parties intend to improve European cooperation on defense and security?"
"What measures do the parties propose to promote digitization and technological innovation in Europe?"
I will spare you the answers to avoid burdening you with my numerous biases and to avoid writing political content, but you will see that they will be careful and probably exhaustive.
Charts
Reducing a program to a number is more of an exercise in style but can help to have a macro orientation.
Ok, now you can create a table with the 4 parties in columns, the points analyzed in rows, with sub-points for each to highlight nuances, and indicate with a score how each party favors that point. -1 if unfavorable, 0 if not declared or neutral, 1 if favorable. Use two decimals for nuances. Then make a chart.
Here is the response because it is certainly interesting.
But here, too, you can get creative.
Notes:
These examples work if you use AI versions with long context windows. To be clear, if you are paying for your AI in the premium version. Do not do this with GPT 3.5 or Copilot or free versions offering less than 100,000 tokens of the context window.
Do not exceed 3-4 programs, especially if they are very long, as they may cause you to exceed the context window limit and lead to hallucinations, which, especially in this case, are problematic. 🔴 In my example, I used only 24,000 tokens out of 128,000 of the model I used (copy the entire conversation into Word and count the number of words, then multiply by 0.5 or put it here to get an exact token count). This means I can still expand the conversation without likely encountering the risk of hallucinations.
Double fact-check at the end: verify by reading the program as in the old days to ensure that the party that inspires you the most is indeed in line with what the AI proposed.
If you are, as you should be, concerned about your privacy and use Chat GPT, go to https://privacy.openai.comand follow the instructions to request that your data not be used for training. For others (Claude and Gemini), make sure to disable all tracking options.
On Biases and Their Impacts
The topic of biases is complex. But, in short, for those who have never heard of it, it works like this:
Imagine having a consultant who has read only one or two newspapers supporting a certain point of view all their life. Their view of reality will inevitably be partial. Similarly, AIs learn from the data they are given and, consequently, inherit their biases.
So, in this AI analysis, since AI received information from a party's program, there could be 'unintentional bias' de facto.
How to deal with it?
We would need AI tools that explicitly declare their biases (utopia). As mentioned, you can get a general idea, but it is essential to reread the original texts to ensure they align with what you have understood.
I added a graph from a site that tries to test the political orientations of AIs daily: https://www.trackingai.org. Keep it in mind because it is a real issue.
So?
I have already said everything, but will recap the main points:
Awareness of Biases: it is crucial to be aware of and try to recognize them.
Choice of Tools: this analysis cannot be done by all AI models but only by the most advanced ones.
Human Review: The choices and decisions are always yours. In workshops (and in the upcoming book), I often describe Generative AI as an intern who knows everything but cannot take responsibility.
Transparency is one of the pillars of AI ethics. Who knows if we will ever achieve it?
Moral: I don't know if the answer is positive or negative for you, but I hope I have given you some ideas on how to use your AI model.
I am very interested in your opinion. Do your tests and suggest how the approach could be improved; I certainly have not considered all the necessary considerations, and the prompt can be improved.
But only by trying can you understand. (FAFO: for those who know what I'm talking about 🙂)
📢 Se il contenuto ti è piaciuto, aiutami diffondendolo.
📝 Iscriviti al Blog Glimpse per non perdere nessun aggiornamento:
📚 Dai un'occhiata a 'Glimpse', il mio romanzo sull'intelligenza artificiale 🇮🇹https://www.glimpse.blog/glimpse 🇬🇧https://www.glimpse.blog/en/glimpse
🗓️ Contattami se vuoi organizzare un Workshop sull'IA o per qualsiasi idea.
A presto! Massimiliano
コメント