Fake news and false claims: can AI help fact-check the Trump administration? (original) (raw)

Imagine you’re the head of machine learning at a big social media company, and you’ve been asked to design a system that can detect “fake news”. Succeed and you’ll be in line for a fat bonus. Screw up and you could put democracy itself at risk. Where do you start?

You might want your pet AI to score articles for “fakeness” using some points-based system. Maybe you want it to spit out a binary classification, labelling articles “fake” or “not fake”, perhaps with a degree of probability attached. Either way, you can use the output to decide whether and how to display a given news story in your users’ feeds.

To achieve this, the algorithm could identify features in the text of news articles that correlate to fakeness. These might be common BS phrases like, “lose 14 pounds in a week,” or, “you couldn’t make it up.” Alternatively it could learn some basic heuristics: simple, efficient rules such as “was it published in a reputable newspaper?” that could be used as a quick fakeness test.

What do we mean by “fake news” though? Alongside “post-truth”, its baby sibling, the term has been popular globally since November last year as you can see from the chart below. Presumably this correlates with some major global news event, but what it is we can only speculate.

Allow Google content?

This article includes content provided by Google. We ask for your permission before anything is loaded, as they may be using cookies and other technologies. To view this content, click 'Allow and continue'.

Defining “fake news” poses two big problems. Firstly, like “trolling”, the term has eroded with usage. Originally referring to specific websites that published entirely made up news stories, “fake news” evolved into a catchall phrase for pretty much any news that people don’t like, or seek to dismiss.

Second, like “lie”, the word “fake” implies not just that someone is wrong, but that they have made a deliberate attempt to deceive. Unless you can actually prove that intent, you’re making an accusation that is factually and legally dubious. If I say that Trump’s immigration order would have prevented 9/11 then I may be lying or I may be misinformed; without further evidence you simply don’t know. This is why journalists are – correctly - reluctant to use “lie”.

To get around all this, we’d be safer to talk about false news. This puts us back on firm, empirical ground – news articles present us with various facts about the world, and those facts are either true or false. Theoretically, by counting the number of true claims and the number of false claims we could calculate a crude “accuracy rating”.

Some of the technologies for automated fact checking already exist in some form. ClaimBuster is an Australian project that uses natural language processing (NLP) techniques to try to identify factual claims within a text. It won’t automatically fact check them, but it can assist a journalist by pointing them to the most “checkable” statements.

We also have knowledge bases that provided structured data to query statements against. Wikidata, a Wikimedia Foundation project, provides it free to anyone who wants to use it. Wolfram|Alpha stores curated facts and knowledge in a large database, and allows users to search it with natural language questions like, “How big is England?”

If you combined the capability of Wolfram|Alpha with ClaimBuster, and added some reasonably simple integration code to glue it all together, you could build a system that executes something like the following workflow for a limited range of well-established “facts”:

• Detect statements in news articles like “Mercury is the fourth planet from the sun”
• Replace the object of the sentence (Mercury) with “what”, e.g.: “What is the fourth planet from the sun?”
• Feed the resulting query into Wolfram|Alpha
• Take the result (“Mars”) and check whether it matches the original object (“Mercury”).

If we were checking school textbooks this might work quite well, but the problem with news is that its, well, new. By definition, if a journalist reports a new piece of information then it won’t exist in some well-curated database to be checked against. Our database would have to be pretty near omniscient, or we’d need teams of fact-checkers like the excellent Full Fact manually checking and updating information about breaking stories in real-time.

Now, that could actually work quite well in the near future. A lot of dodgy myths that appear in the news have been repeated ad nauseum over many years, and fact-checking organisations have debunked many of them. If Full Fact took the information they already have, curated it into a standard format, and made it available to programmers like me via an open API, it would take very little time to build a crude, working tool to check news articles against.

What’s the catch? Well this approach might work for a narrow set of cases – oft-repeated myths that can be fact-checked against reliable sources of information. It’s great for claims like “we invested £10bn extra in the NHS last year”, for example. When it comes to current events though, it falls short; and with post-truth leaders like Donald Trump or Jeremy Corbyn we get into some truly weird territory. To see what I mean, let’s take two examples.

Labour leader Jeremy Corbyn campaigning in Cumbria last week. Photograph: Danny Lawson/PA

The first is the curious tale of Jeremy Corbyn’s whip. On 19 January, the Guardian reported that Corbyn would order Labour MPs to vote for the government’s bill triggering Article 50. He was asked twice if he would impose a three-line whip, and in both answers implied that he would. Later in the day his position became less clear. The following day, as confusion over Corbyn’s position continued, pro-Corbyn blogs such as The Canary and Another Angry Voice ran articles suggesting that the Guardian had “duped” it’s readers. Thomas Clark, author of the latter blog, accused journalists of “fabricating Fake News headlines to damage Jeremy Corbyn’s leadership.” Within days, it was confirmed that the Labour leader was in fact planning to impose a three-line whip, leading to resignations and a rebellion by Labour whips.

There are two interpretations of all this. One, which you’ll find in various Corbyn supporters’ forums, is that the Guardian ran a fake news story that later became ... not fake. The other, which I personally subscribe to, is that Guardian journalists correctly interpreted Corbyn’s quotes on the 19th, presumably did the leg-work with their own party sources, and produced an accurate and important piece of journalism.

The second example is Trump’s catastrophic “Muslim ban”, an executive order that ignited protests across the world over the week. The wording of the order is so legally sloppy that parts of it are difficult or perhaps impossible to clearly interpret, as Benjamin Wittes points out. At the time of writing, reporters and Homeland Security staff alike are still trying to work out the ramifications of it, even as the Trump administration continues to issues further “clarifications” and amendments to the order.

Even that nickname, “Muslim ban” is contentious. Taken literally and in isolation from other facts, Trump’s order applies not to any race or religion but to the nationals of several countries. Of course the order comes within a context, and that context is a vow by Donald Trump to crackdown on Muslims, Giuliani’s suggestion that Trump originally asked for a “Muslim ban”, the order’s targeting of Muslim countries that have little or no connection with terrorism in the United States, and so on. To say that Trump’s order is a ban on Muslims is technically false, but to suggest that it doesn’t target Muslims is equally disingenuous.

People hold posters during a protest against President Donald Trump’s executive order banning entry of citizens from seven Muslim-majority countries, outside the Terminal 4 at John F. Kennedy International Airport in New York. Photograph: Xinhua / Barcroft Images

What we have in both cases is a sort of quantum news. Whether by accident or design, these leaders generate a chaotic fog of confusion that leaves observers grappling with stories that are both true and false at the same time. Objective reality is a fleeting shadow that cannot be grasped, and what you believe becomes a function of whom you trust.

From here, the news devolves into mess of “he said, she said” claims and counterclaims; a game of bluff where the main objective of the politician is not to inform or communicate but to undermine. If you can’t trust reality then you can’t trust the people reporting on it either, so they become less and less relevant. Which is, of course, what the politician wants.

What does this mean for our algorithm? Running with the idea of trust, a simple heuristic might be to make a list of trustworthy newspapers or journalists, and restrict news content to only those sources. This can work quite well, with at least two Chrome extensions using the technique. It’s not that organisations like the BBC or CNN are infallible or unbiased of course, but they invest more effort and resources into verification than, say, The Canary. At a larger scale, using human editors to review publishers has (along with other features of their design) helped Apple News and Snapchat to largely avoid the fake news problem.

Of course, this relies on some agreement about what’s trustworthy, and that’s harder to find in societies increasingly marked by extreme partisanship. Even Private Eye found itself on one list of “fake news sites” recently. With America’s scientific agencies gagged for political purposes and a President waging all-out war on the media, the knowledge in people’s brains has become more than ever a simple function of party affiliation. No amount of reasoning can force people to learn the ‘right facts’ even if we could agree what they are. All the fact-checkers in the world, automated or otherwise, can’t fix that.