Facebook Reportedly Developing AI to Summarize News – What Could Go Wrong?

Facebook has been trying to gain a foothold in the news space for years. Last year, the company launched a special section on its site called Facebook News for users in the US. It also wants to extend this program to other countries such as Brazil, Germany and India.

But that’s not the only project that the social network in the news space is working on. According to a report from BuzzFeed News, Facebook is testing an AI-powered tool called TL; DR (Too Long; Didn’t Read) to recap news articles, so you don’t even have to click through to read those articles.

The report noted that the company showed off this tool at an internal meeting last night. It also plans to add features such as voice commentary and an assistant to answer questions about an article.

At first this seems like a great idea – getting a quick summary of an article you don’t have time to read for, right. There are already some similar tools like the AutoTLDR bot on Reddit. However, given Facebook’s sketchy history with news and publishers, there are many ways this can go wrong.

At best, the AI ​​makes silly mistakes when parsing the article content, so you can’t understand the summary it spits out. And last but not least, We’ve seen many incidents where bots pick out problematic parts of the content of their training algorithms and spat racist gibberish.

In the worst case, there is the possibility of creating or spreading misinformation. There are a ton of news sources on Facebook that are not known for their accuracy. If a skewed summary of those articles starts to float, it can cause more trouble.

Facebook will also have to train its algorithm to avoid taking quotes or phrases from articles out of context. An apparently unproblematic summary could contradict the article or topic and vice versa.

In the past, researchers have successfully tricked AI systems designed to detect toxic comments on the Internet with ‘positive’ words. If the folks behind propaganda operations could crack Facebook’s article summary algorithm, they could write stories so that the digests contain the messages they want to spread.

Many reports have pointed to the huge problem of social network misinformation, much of it due to poorly designed software. Although the TL; Facebook’s DR product is not yet public, it already sounds like it could be a disaster.

Published December 16, 2020 – 06:16 UTC