Fake News, Conspiracy Theories, and Facebook

It is likely that fake news sites played a role in the recent US election. Web companies, particularly Facebook, have come under tremendous pressure to curb fake news. It seems disingenuous for Facebook to claim that it is not a media company and hence cannot take any responsibility for content posted on its site.

Yet, the solutions proposed by many are draconian: we do not want Facebook, or other private sector titans to decide for us what news we are allowed to see. Furthermore, many proposed solutions are impractical: there is no chance that we will be able to build a perfect fake news filter in the foreseeable future. To see this intuitively, consider the much easier problem of filtering email spam: such a filter has to separate spam, fake email messages, from real email intended for you. Such filters are in wide use today, and do a pretty good job. But, even after years of tuning, they are far from perfect. Some spam mail makes its way into my inbox and some real emails get diverted to the spam folder. Similar technology can be used to identify fake news, but the problem is harder, with fewer clues since there isn’t an explicit recipient and there usually isn’t a desired action (e.g. click on phishing link or purchase a male enhancement product). Indeed, people have already built such classifiers today. I haven’t had a chance to test any yet, but they couldn’t possibly approach the quality of commercial spam filters today.

One promising approach is not to classify news items directly, but rather to classify sources. You may have no way of judging on its own merits whether a statement I make is the truth. But if you have known me for a long time, and you know from experience that I am usually truthful, then you will believe my statement because I said it. We often have much more information about a source, and therefore can classify them more accurately than individual stories. New sources will have to build up credibility before they are believed, and we will have to create mechanisms for them to do so. If we make building credibility too onerous, then new sources can never gain credibility and become important; but if we make it too easy, then fraudsters will just repeatedly create new sources, do their damage, and move on. Devising well-balanced mechanisms is a great challenge for the tech giants to take on, and they have experience doing this in other spheres, as with user-submitted product and travel reviews.

Unfortunately, even the most honest among us will fall prey to falsehoods from time to time. Even if you correctly believe me to be trustworthy, I could be genuinely mistaken about some fact. For example, some years ago the entire country fell prey to fake news about non-existent weapons of mass destruction in Iraq. You could read about it in the most reliable of sources, like the New York Times. It is unlikely that a smart algorithm would have fixed this problem.

Fake bag looks like the real thing       zizzie fizz       CC-BY-ND-2.0

Even worse, we have errors of omission, biased reporting, unjustified extrapolation from an incident reported truthfully, and so on. Each of these errors can be just as troubling as straight up fake news. Reputable news sources usually try hard to avoid such errors. If they fail, they have a reputation to lose. Less reputed news sources have less to lose. It may be possible to algorithmically detect lack of balance in reporting or unsupported conclusions. But now we are going into algorithmic labeling far beyond identifying outright falsehoods. It will be even more difficult to do this right, and potentially more consequences for getting it wrong.

The bottom line is that conspiracy theories have been around forever. The web did not invent fake news or rumors. Their outsized impact today is because the web accelerated their spread. One man in Texas, with only 40 followers on Twitter, was able to start a rumor that protesters were being bussed to Trump rallies by the Clinton/Democratic organizaion. (He had actually seen busses for an unrelated event organized by Tableau software, which happened to be not far from the Trump rally). Enough people wanted to believe this story that it was seen by several hundreds of thousands (perhaps even millions) before it was debunked. Without technological acceleration, this one misinformed person could not have done nearly as much damage.

What this suggests is that technology can be used to avoid the accelerated spread of such stories. If we can slow down the spread, we can provide enough time for verification and for response. There are many technical options for how to do this. There is also room for innovation. Speed bumps are what we need to combat fake news and not censorship.

You need privacy to get good data: The Lubitz lesson

Andreas Lubitz intentionally crashed Germanwings flight 9525 into the French Alps, killing all 150 people on board.  Lubitz had a history of mental illness, and should not have been flying a plane that day.  Many news reports were quick to blame German privacy protection laws.  (For example the Times headline was “German obsession with privacy let killer pilot fly;” Time magazine said “German privacy laws let pilot ‘hide’ his illness from employers;” and the Washington Post report was “Crash challenges German identity, notions of privacy.”).

But before we jump to privacy protection laws, let us begin with the real question: what systems could we have had in place to stop him from flying, even if we did not care about privacy at all?

Of course, this question applies not just to pilots, but also to members of so many professions we trust not to hurt us.  The police, of course, come to mind immediately as an example.   How about a barber who holds a razor blade at your neck?  How about an industrial worker with access to chemicals with which she could cause an explosion or a catastrophic fire? 

The question also applies to matters other than mental illness.  Anyone with an infectious disease, even a mild cold, is a potential threat in a hospital ward with many immune-suppressed patients. 

If the airline had known the extent of Lubitz’s illness, we can suppose that they would have grounded him[1].  German privacy laws did play at least some role in keeping details of his illness from his employer.  But let us see how this would have played out if there were no such laws.  Germanwings would likely have fired Lubitz: however the monetary aspects may have been handled, there is no question that Lubitz would not have remained a pilot (and possibly not even become one in the first place).  But Lubitz apparently loved to fly: becoming a pilot was his dream.  If he knew in advance that his medical record would be shared with his employer, he may have not have sought treatment in the first place.  Thereby, he could have ensured he had a pristine medical record, and been qualified to fly.

The crucial point is that decisions based on data are only as good as the data themselves.  In real life, data are often dirty, for a variety of reasons.  Data cleaning is difficult but necessary before any data analysis is attempted.  By creating incentives for falsifying data, we virtually guarantee that the data will be dirty.  Think of how many times you have intentionally provided incorrect or incomplete data, such as your phone number, on a web form.  If you feel the web site has no legitimate need to know your telephone number, you choose to give it some junk that satisfies the site’s requirement that this field not be left blank. The benefits to you of lying about your phone number are so much lower than those to someone like Lubitz lying about their health.  Surely, with his career on the line, the prospective pilot would do his best to hide any disqualifying condition.

So, what is the solution?  Separate testing.   Employees in critical positions of trust should be tested, randomly, periodically, and when there is any doubt.   This is how many other requirements are managed.  Consider alcohol or drug use.  Pilots are expected to be sober when they show up to fly.  About a dozen pilots are caught every year with blood alcohol level above the limit.  If someone showed up to work visibly drunk, presumably they would be tested and disqualified.  This testing has nothing to do with medical records.  Pilots can tell their doctors about their alcohol use.  Pilots with alcohol abuse problems can seek treatment and receive help without jeopardizing their careers, as long as they manage to show up to work sober.  We could test similarly for other conditions.  If someone is going through an acute mental crisis, a trained mental health practitioner will likely be able to see that right away.  In fact, co-workers may be able to guess something is wrong, just as they would if the employee showed up drunk.  Are such tests expensive?  Yes.  Are they fallible? Yes.  But so are judgments made even from perfect and complete medical records.   Certainly, judgments made on incorrect and intentionally misleading records will be worse, not to mention the health costs to employees of seeking care only outside the system and the needless loss of privacy for professionals in positions of trust.

[1] There is a complicating wrinkle in the case of Andreas Lubitz.  He had just seen a psychiatrist and had been given a medical note that he could have used to excuse himself from work that day.  Lubitz chose to tear up the note and fly.  Should the psychiatrist have alerted the airline directly?  Possibly.  Extreme actions are justified when there is danger of significant imminent harm.  To me, this is the difference between the doctor advising you to stay home when you are sick, so as not to infect others, and the doctor contacting the relevant health authorities to recommend involuntary quarantine.  The latter path of action is sometimes justified, but the threshold has to be kept high.