Recent political upheavals have caused global debate about the spread of “fake news” on social media and elsewhere: reports that look like news, but which are intentionally untruthful. Fake news is often distributed for political or commercial gain; a classic example is how opponents spread news-like reports claiming that former US President Obama was not born in the USA (and, by implication, not a ligitimate US president).

The present talk, which will be very informal, will examine the idea of “deviating from the truth”. I will do this by taking an engineering approach: First I will sketch the working of a typical data-to-text Natural Language Generation (NLG) system. Next, I show how each stage of the NLG pipeline has to make debatable decisions which can impact on the truth of the resulting text, making deviations from the truth extremely difficult to avoid. From these observations, I will argue that the notion of fake news is extremely difficult to pin down and detect, and I will suggest that effective solutions to the problem of fake news should therefore focus not only on the computational side of the problem but also on educating the human recipient of the news.

This is joint work with Ehud Reiter, University of Aberdeen.