Inspiration

In these vulnerable and dynamic times, we can easily be overwhelmed with the amount of news that reaches us throughout the day via different channels and news portals. The information is oftentimes incoherent and causes confusion where there should be assurance instead. To enable a news browsing experience that is fake news checked without a hassle, we created an easily accessible Chrome browser extension. It is designed to give a second opinion on the trustworthiness of news regarding Corona - quickly and easily.

What it does

The CoroNEWS Checker supports the user while browsing news online. It searches the website content for signs that hint at fake news, but only if voluntarily triggered by the user. Protecting the privacy of the user is one of our main premises as trust is what we aim to empower. Through a simple click on the extension icon, a multilayered fact check is started. At its core lies a specialized AI that processes the natural language content of the news article within the website and detects signs for fake Corona news. On top of that, established criteria that are commonly used by human experts are also checked for (e.g. like spelling mistakes, dubious URLs, overuse of exclamation marks). Clicking the icon will open a pop-up which shows our CoroNEWS Checker mascot. The color of it changes like a traffic light to indicate how trustworthy the source is. This judgement is based on the full set of criteria which are listed in the pop-up for transparency.

How we built it

We are a team of mainly backend and AI developers. For the natural language content, we fine-tuned a state-of-the-art Transformer network called RoBERTa (https://arxiv.org/pdf/1907.11692.pdf). We used an open-source PyTorch-based library for prototyping (https://github.com/kaushaltrivedi/fast-bert). The model is optimized to classify whether or not a news text is fake or real. We used a freely available Kaggle dataset (https://www.kaggle.com/clmentbisaillon/fake-and-real-news-dataset) for the hackathon-prototype of our AI. The accuracy of the model on unseen testset data is very high (around 99% accuracy) and provides a promising foundation for the future vision of our tool. The remaining fact checking heuristics were programmed in Python. We used JavaScript for the frontend realization. The service is run on a Google Cloud Kubernetis Cluster with GPU.

The solution’s impact to the crisis

Our tool helps people to navigate news with trust. Having people aligned on facts will improve social coherence and political stability, on the one hand. On the other hand, we provide a way for individuals to cope with incoherent information and the insecurity caused by it. Finally, getting facts straight is a crucial factor to ensure that everyone is behaving in line with recommendations by official institutions. That means, we will know how to act to our own and everyone else’s safety for sure. Since our solution is so easy to apply and lightweight, we hope to reach a wide audience. This extension is added once to the browser and stays to change the news reading experience on a daily basis.

Challenges we ran into

Training the AI model in a timely manner has been a challenge. By learning and reusing from existing approaches, we have managed to get good results for prototyping. For better generalizability and topicality, however, we will need to collect more data. In AI, the quality of the groundtruth that the model is trained on, determines biases that influence its decisions later on. We have had lengthy discussions about our ethical responsibilities in this. As we go on from the hackathon, our data collection process will need to go through a thorough process of human fact checking and bias prevention. Privacy and transparency are two of our main premises. Creating an automated fact checker is a sensible endeavour as criticism towards techno-solutionism by society has good reason. We therefore had to compromise between technically more ubiquitous solutions and maintaining a human voluntary factor. For example, automatically scanning text content would allow to have an output readily available at all times. It would, however, also be invasive to the user’s privacy. We instead decided to leave it up to the user whether or not they wish a second opinion on the news they are reading.

Accomplishments that we're proud of

Only within the timeframe of the EUvsVirus Hackathon, we accomplished to create a prototype of our Chrome extension in a team of seven members that, for the most part, have never worked together. The tool works on news in English language with transparent criteria. The underlying AI so far classifies fake news reliably and accurately.

What we learned

In the face of a crisis that causes insecurity and mistrust in us all, we learned that careful considerations of ethical impact are key. Technology must be at the service of its users’ needs.

What's next for IASvsVirus

In order to elevate the prototype of our CoroNEWS Checker, we need to team up with trained human fact checkers and journalists. We would like to extend the amount of common heuristics as well as the performance and topicality of our AI. A refined ethical concept for continuous data collection and evaluation would be our first step in this respect. The tool needs visual and technical refinement to increase usability and scalability. Further, we need to improve the applicability to languages and peculiarities of news content in all EU countries. Finally, we would like to extend it to a wider range of browsers.

Value of our solution after the crisis

The tool is designed in a way that it can be easily applied to non-COVID-specific news. Fake news has been an issue before this specific crisis and will likely be a risk to our political and social sphere after. By training the AI continuously on new content, we can ensure that it always addresses up-to-date topics.

Share this project:
×

Updates