Abstract:

Civic society is threatened when major issues like climate change and biological evolution have a high degree of consensus among scientists, but public debate is tainted by misinformation. An anonymous Google search of 100 pseudoscience terms yielded 1.3 billion unique web pages; the Internet is awash with scientific misinformation, often propagated and amplified by social media. This project aims to combat misinformation with a technology approach using neural networks and machine learning. The training phase involves undergraduate science majors selecting several hundred legitimate and fake articles on each topic. Climate change and evolution were the testbeds. Various language models and algorithms were used, and the trained neural network achieved 95% accuracy in distinguishing the real from the fake. This success was maintained in scaling up to tens of thousands of articles, assisting by using misinformation articles for “hard negative” training of the neural network. The main deliverable from this project will be a website where users enter the URL for a non-technical science article and the neural network will return a Bayesian probability that the article is legitimate. If the article is suspicious, the system will return suggestions of correct content. Another application of the technology will be in a smartphone app to gamify the process of classifying unattributed articles as real or fake.

Keywords:

artificial intelligence, machine learning, technology, science, misinformation