Judging the quality of (fake) news on the internet

Serious news articles give clear information on the writer, whose identity is verifiable and whose background knowledge on the topic is documented somewhere. It is the authors response to assure the validity and have the underlying facts undergo a quality check, since the actual source behind is usually not disclosed (even protected in good journalism). For web pages, an imprint, date of last update and host are the minimum requirements. Absence of such information is not necessarily a sign of fake news, but an indication to be at least cautious. For images, it pays to use the “image reverse search” that many contemporary online search engines offer, to see where the picture occurred elsewhere in the web, and ultimately to see where it came from first.

Many people get their news and information from search engines like Google. But where does Google get its information from? Roughly speaking, Google ranks pages according to the lot of links leading there. Similarly, science nowadays judges research quality based on how often an article is cited. Alas, the number weblinks, “likes” (in social media) or citations is not an indicator of correctness, truth or soundness, since these numbers do not tell the context of the citation (a link can call the reference valuable, false, or even dangerous to believe, but irrespectively of this would count as a +1 into the citations). The study of citation graphs and the internet topology uses random graph theory, and has made interesting discoveries: for example, some citation graphs exhibit small-world phenomena (Li-Chun et al. 2006), i.e., the short average distance in the graph between any two nodes. Similarly, the Internet is what we call a scale-free topology (Barabási et al. 2000; Zhang et al. 2011). Such networks have a few nodes with very many connections, called “hubs”, and very many nodes with only a few connections. In social networks, we would call the hubs “influencers”, while nodes with only a few connections are “followers”. For the question of where information comes from, a small-world graph topology means that an information can quickly spread and reach a wide audience taking only a few steps forward from the originator. From the consumer’s perspective, if the topology is scale-free (as the Internet appears to be), then it can most likely be traced back to one of the few hubs to be the originator. We find instances of this effect in many cases of news on the web, since newspapers often buy their information from larger news companies, so that articles, even if they appear in different media, may nonetheless originate from the same news agency. Tracing back the source of information is thus a crucial matter for judging the quality of information, as it can only be as good as the original source, and most likely, there are not too many independent such sources (and every forward can—even unintentionally—modify, degenerate or otherwise blur the information).

With all this known, it is tempting to ask if artificial intelligence could help us distinguish right from wrong information, but the challenge remains tricky. Certainly, a machine has—unlike a human—no self-interest or unethical intentions, but since AI draws its power from learning, it can be trained to serve unethical needs of a human designer. The point is that AI can only be as good as the data that has “shaped” it, e.g., recognizing or reproducing patterns that were in the training data. Consequently, AI will only be good in recognizing fake news if it has been trained on good set of reputable vs. fake cases. Likewise, it will only produce high quality and objective articles, if its training author provided such articles beforehand. The bottom line is that AI can at best be a supporting technology, not a substitute for human intelligence. The recently coined term “robot journalism” (Dörr 2016) uses the power of AI to machine-generate news articles. Can a human intelligence recognize this, say, if the AI were to publish biased (and hence in a way fake) news? Essentially yes: humans can be critical, reflect and challenge their own opinion and that of others. AI cannot do the same trick as good as a human, and as long as no AI has ever passed Turing’s famous test (which as of today has not happened), humans can recognize a machine talking to them with a chance of more than 50% in general.

So, is AI at all more a danger than useful to recognize fake news? The answer is clearly no, since the power of AI to recognize even the smallest inconsistencies in possibly huge lots of facts, or the ability to classify images based on features that are invisible for the human eye make it invaluable weapons for the recognition of fake news. “Stylometry” applies pattern recognition to linguistic data to assign authorship of an unknown text to a specific person (e.g., Ma et al. 2009; Holmes and Kardos 2003). Statistical methods like the Benford test (Durtschi et al. 2004) can recognize forgeries in numbers (in a weak form but enough to raise awareness to take a second look). The real power, as always, lies in a clever combination of such techniques.

There is neither a silver bullet nor a one-size-fits-all algorithm to recognize fake news. Essentially it boils down to a battle of wits between those generating false facts, and those who shall buy it. The position made here is to raise awareness that the best practice to protect against fake information is to follow a scientific approach that, perhaps not surprisingly, also kids adopt by nature: question everything and always ask “who said that?”. Try to use more than one source of information, say, not only use Google, but use a meta-searcher instead, which can query Google but also independent other sources. Likewise, do not buy a picture “as is”, but rather check if it appears elsewhere too, perhaps in a different context. Numbers can appear more convincing that qualitative facts, but even a perfectly ethical and careful intention behind a statistic is no protection against human error when the numbers are computed. Statistics should always open the underlying data to let others judge and reproduce the results; an issue called the “scientific reproducibility crisis” (Stoddart 2016) today. The same applies for news vs. fake news: can the claimed facts be verified from independent sources, i.e., not only by getting many copies but getting the same data using a different origin? The question has no ultimate answer, but awareness is already a key step towards tackling the risk of fake news.