Training a Dcipher Research Bot to answer our questions about the EU AI Act
In this blog post, we take a look at how Dcipher Research Bots can be used to quickly find information hidden in long and complex texts, using the EU Artificial Intelligence Act which has recently been approved by the European Parliament as an example.
As we all know, new applications of AI—artificial intelligence—are being introduced with breathtaking speed these days. Not least have chatbots built on generative AI become widely adopted by individuals and organizations alike during the last one to two years. Here at Dcipher Analytics, we have contributed to that trend by releasing Dcipher Research Bots—chatbots we have developed specifically for research and insight work—as an addition to our toolbox of AI-powered text analytics tools. Other organizations are developing AI applications in virtually all fields of society including healthcare, military, education, recruitment, and energy systems, to just mention a few.
As new technologies have done again and again before, AI has the potential to transform society in a myriad of good and bad ways. With the aim of steering developments in ways that boost the benefits of AI while controlling its risks, governments around the world are introducing new regulations of the field as we speak. One significant regulatory project is the European Union’s “EU Artificial Intelligence Act,” aimed to be the world’s first comprehensive AI law.
What should we do if we want to learn more about this historic piece of legislation? If you want to know more than what is mentioned in the summaries that have been published, you can of course turn to the act itself: 108 pages of dense legal EU prose. Let’s say that you are also interested in how this historical piece of legislation has come about; then you are lucky, as the official drafts, proposals and opinions are publicly available online. Your only problem is that you now have 6,685 pages to go through.
Isn’t this exactly the kind of situation where it would be nice to have an AI read the texts for you? Of course it is! And you are in luck, as Dcipher Research Bots will read whichever texts you throw at them and answer all of your questions through a familiar chat interface.
To see how a Research Bot can help us understand the EU AI Act, we first download the official documents published by the EU institutions and made available as PDFs on the union’s website dedicated to the act. We then log into dcipheranalytics.com and choose to create a new Research Bot based on PDF reports. We click “Select input data” and upload our PDFs. After selecting to do an “immediate single run” and providing a name for our new bot, we can hit the “Create” button.
After a short while, we will get an email notifying us that our Dcipher Research Bot is ready. Clicking the “Start chat” button then takes us to a chat window where we can ask all of our questions to the bot. A Dcipher Research Bot excels both at summarizing long and complex texts and at finding specific pieces of information hidden within them. What we do from here is up to us and depends on what we are curious to find out from the thousands of pages the bot has read through. We may, for example, start by asking a very general question: “What is the EU AI Act about?”. Typing this into the chat window will give us a summary of what the act claims to be intended to bring about: harmonized rules, AI systems in line with European values, and social trust in AI.
To understand more about what is behind the lofty words, we can continue by asking our bot more specific questions, but we also have the option of going to the sources ourselves. Dcipher Research Bots make this very easy by providing source references for every statement they make. Mousing over a number within brackets in a bot’s answers will show us an excerpt from the original text that a part of the answer builds on, and scrolling down to the bottom of such an excerpt we can see the name of the document that it belongs to as well as on which page in that document we can find it.
Let’s say we now want to take a step back before digging deeper, and ask the bot how AI is defined by the act. Again, getting an answer to our question is as simple as typing the question:
We may then go on to for example ask about what different actors will need to consider and comply with under the act. Things we can learn after a few prompts include that the act is intended to create a permissive framework for small and medium-sized enterprises that are using AI, while some specific practices are prohibited (e.g. using AI for student assessments) and others are classified into different risk levels for which different requirements are imposed. Whenever we come across something we want to know more about, we type a follow-up question. For example:
Taking advantage of the fact that the bot is trained not only on the legal act itself, but also on other documents produced throughout the legislative process, we can also ask questions about that process. For instance, we can ask which actors came up with suggestions for different parts of the act’s content. After coming across a mention of a ban on using AI for establishing social credit scoring systems, we ask who it was that called for that to be included:
Why not try it out yourself? Train a Dcipher Research Bot using any set of PDF documents, or set it up to collect and read any number of news articles or social media posts from an area you are interested in. Following the link you will be able to set up a Dcipher account if you do not already have one and get a welcome gift in the form of Dcipher credits worth US $50 which you can use on Research Bots and Dcipher Workflows. Before training your own bots, you can also start with trying out the Dcipher Research Bot experience using our trend bot demo.