Chequeabot, Chequeado’s fact-checking automation platform, relies on AI and Machine Learning to speed up the process of fact-checking without sacrificing quality to battle the growing amount of misinformation created and circulated at a much faster rate than fact-checks. Chequeabot does many things: it automatically scans over 30 media outlets all over Argentina, as well as all the speeches and conferences given by the president, identifies claims that can be fact-checked thanks to AI and Natural Language Processing (NLP), and indicates which of those claims are related to previous fact-checks. It also provides an open transcription platform and a text analyzer.
Firstly, Chequeabot frees a lot of Chequeado’s editors’ time that can then be used to produce better content, while reducing biases by being sure about its regional and media coverage. Over these two years and a half, Chequeabot has made its way into Chequeado’s newsroom meetings on Mondays, suggesting claims to check and helping our journalists find claims related to specific persons and topics.
The platform runs checkable claims with Chequeado’s database of statements previously checked, which allows to publish quicker in social networks when an already factchecked claim is being repeated, through a second function called “What’s already been checked”. This has already proven extremely useful at key moments, such as debates as it has allowed Chequeado to react faster and publish in social networks relevant content related to what was happening, and freeing the journalists to dedicate themselves to checking new information. We want to take this functionality to other editors as it relieves the work of journalists and help them avoid any omissions that might exist. Chequeado has also developed a tool to extract video transcriptions from YouTube: Chequeado’s Transcriptor (chequeado.com/transcriptor, or chequeado.com/desgrabador in Spanish, is an open source application, based on Chequeabot’s development. It also links every phrase to the exact moment in which is said in the video, to make its verification faster and easier. And it’s free and open for everyone who need to speed up their work.
To complement this, we launched a microsite that integrates several automation tools in one interface, which allows users to submit text to be analysed searching for checkable statements, and relates those claims with previous fact-checks. Although is now restricted to Chequeado’s newsroom, it’s being improved so it can be released and open for every newsroom interested in implementing this tool.
Chequeabot uses scraping techniques to automatically extract information from the selected media outlets, and afterwards, applies Machine Learning and NLP to identify fact-checkable statements and the relevant labels (such as the speaker or the context, or the media where it was reproduced). To do so, it uses Python libraries, such as Scikit.learn, nltk, and spaceit. The information is placed in an MySQL database, which is read by an app accessible and UX friendly to the newsroom, comprised mainly of non-technical professionals.
What was the hardest part of this project?
Artificial Intelligence is a powerful ally, especially in smaller newsrooms with limited resources. However, asking the AI to analyze large amounts of information blindly can become a bigger drawback than the solution it provides. It was hard for us to realize that whatever decision we chose to make could bias the algorithm, even those that seemed to be the ones that could foster and improve our work more efficiently, like asking the bot to prioritize claims from more relevant people. That, in terms of AI and Machine Learning, could have impoverished all subsequent results, neglecting those voices that are also interesting and necessary for the journalistic process. We need our Chequeabot to help us and to be better than us. From Chequeabot, we also need what it offers to be relevant to the newsroom. We need to be sure, for example, that Chequeabot’s results do not hide necessary information, and that it selects claims and fact-checkable with a criterion (that is learned), that reflects the will of the organization, specially when it comes to relevance, plurality and federal coverage. All in all, this kind of experimentation was challenging in several ways, that we could not have foreseen before, being a small newsroom.
What can others learn from this project?
The Chequeabot is an example that it is worth investing in innovation, especially when there are limited resources to be prioritized. Moreover, it shows that the greatest impact is achieved when the problem to be solved is chosen correctly, even if the solution demands too much effort in a first analysis. Failing to diagnose, or misassessing the organization’s priorities, can lead to total failure. If the problem is well chosen, a small breakthrough, like the first demo of the Chequeabot back in 2017, is significant. If the problem is poorly chosen, even major developments can have no impact at all, and that, for a small organization, is critical. For us, Chequeabot was an hours multiplying tool. The investment was big, but the payoff, in the mid and long term, is enormous.