Machine learning is pretty dumb when it comes to sifting fact from fiction, according to new research.
Current machine learning models just aren't up to the task of spotting ‘fake news’, even if they can convincingly generate false news reports.
Two new papers produced by MIT researchers, and reported by Axios, show that current machine learning will disappoint those who hope it can be trained to counter the plague of fake news spreading across the internet.
MIT doctoral student Tal Schuster’s studies show that, while machines are excellent at spotting text that’s been generated by machine, they can’t make a true or false distinction on facts. At least, not yet.
Many automated fact-checking systems are trained using a database of true statements called Fact Extraction and Verification (FEVER).
In one study, Schuster and a team showed that machine learning struggled to handle complex negative statements, such as “Greg never said his car wasn’t blue”; even though they had no trouble deciding a statement expressed as a simple positive was true – “Greg says his car is blue”.
Researchers say the problem is that the database is full of human biases: those who created FEVER had a tendency to write their false entries as negative statements and their true ones as positive, so the computers learn that negative statements are more likely to be false.
The systems, then, were deciding about something other than fake news.
“If you create for yourself an easy target, you can win at that target,” said MIT professor Regina Barzilay. “But it still doesn’t bring you any closer to separating fake news from real news.”
The second study, however, did show that machine-learning systems can do well at detecting when stories had been written by one of their own.
But while detecting which news stories have been written by a machine can be useful, it doesn’t follow that all automated stories are false. Text bots can be created to change true stories for different audiences, or pull together statistics into true news articles.
Sourced from Technology Review, The Verge; additional content by WARC staff
