"FAKE NEWS" AI
Over the past year I have been fascinated by the proliferation of misleading content on the internet. I began experimenting with a Kaggle dataset to prove that AI could be used to score text on how close it is to fact.
THE PROBLEM
and why AI will help
QUANTITY OF "FAKE NEWS"
Misleading news is everywhere on the internet. This not only makes objective truth near impossible to find, it dangerously sways the opinions of internet users to radical policy's that are based on the distrust of facts.
HARD TO DETECT
Fake news is hard to detect with up to 80% of polled adult internet users admitting to falling for outright untrue content. This highlights the inability of people to reliably catch misleading articles without some sort of an aid.
SHORTCOMINGS OF HUMAN FACT-CHECKERS
Having human fact checkers is incredibly helpful in the fight against meaningful content, but in a world with such a high quantity of content it is impossible for human fact checkers to be present on every webpage a user views. This is not to mention the distrust held for individual bias in the work of Fact-Checkers.
An AI trained on a constantly updating dataset of news articles scored by bipartisan moderators would not only be able to provide suggestions of what content to trust by instantly scoring every page a user views, it would also allow for a less biased crowdsourced opinion.
Preliminary Results
In the chart on the left are the preliminary results for numerous AI models trained on a Kaggle dataset of fact based and "Fake" articles. The types of models here range from Support Vector Classifiers to Artificial Neural Networks and each use a variety of natural language processing (NLP) techniques, but the thing to note is that even the worst model has an F1-Score of over 98%. These results, that use relatively simple NLP methods and models trained in a short amount of time, make me confident that AI is a viable tool for detecting fake news.
