Everybody Lies: Big Data, New Data, and What the Internet Can Tell Us About Who We Really Are

In the movie Minority Report, psychics collaborate with police departments to stop crimes before they happen. Should Big Data be made available to police departments to stop crimes before they happen? Should Donato have at least been warned about her ex-boyfriend’s foreboding searches? Should the police have interrogated Stoneham?

First, it must be acknowledged that there is growing evidence that Google searches related to criminal activity do correlate with criminal activity. Christine Ma-Kellams, Flora Or, Ji Hyun Baek, and Ichiro Kawachi have shown that Google searches related to suicide correlate strongly with state-level suicide rates. In addition, Evan Soltas and I have shown that weekly Islamophobic searches—such as “I hate Muslims” or “kill Muslims”—correlate with anti-Muslim hate crimes that week. If more people are making searches saying they want to do something, more people are going to do that thing.

So what should we do with this information? One simple, fairly uncontroversial idea: we can utilize the area-level data to allocate resources. If a city has a huge rise in suicide-related searches, we can up the suicide awareness in this city. The city government or nonprofits might run commercials explaining where people can get help, for example. Similarly, if a city has a huge rise in searches for “kill Muslims,” police departments might be wise to change how they patrol the streets. They might dispatch more officers to protect the local mosque, for example.

But one step we should be very reluctant to take: going after individuals before any crime has been committed. This seems, to begin with, an invasion of privacy. There is a large ethical leap from the government having the search data of thousands or hundreds of thousands of people to the police department having the search data of an individual. There is a large ethical leap from protecting a local mosque to ransacking someone’s house. There is a large ethical leap from advertising suicide prevention to locking someone up in a mental hospital against his will.

The reason to be extremely cautious using individual-level data, however, goes beyond even ethics. There is a data reason as well. It is a large leap for data science to go from trying to predict the actions of a city to trying to predict the actions of an individual.

Let’s return to suicide for a moment. Every month, there are about 3.5 million Google searches in the United States related to suicide, with the majority of them suggesting suicidal ideation—searches such as “suicidal,” “commit suicide,” and “how to suicide.” In other words, every month, there is more than one search related to suicide for every one hundred Americans. This brings to mind a quote from the philosopher Friedrich Nietzsche: “The thought of suicide is a great consolation: by means of it one gets through many a dark night.” Google search data shows how true that is, how common the thought of suicide is. However, every month, there are fewer than four thousand suicides in the United States. Suicidal ideation is incredibly common. Suicide is not. So it wouldn’t make a lot of sense for cops to be showing up at the door of everyone who has ever made some online noise about wanting to blow their brains out—if for no other reason than that the police wouldn’t have time for anything else.

Or consider those incredibly vicious Islamophobic searches. In 2015, there were roughly 12,000 searches in the United States for “kill Muslims.” There were 12 murders of Muslims reported as hate crimes. Clearly, the vast majority of people who make this terrifying search do not go through with the corresponding act.

There is some math that explains the difference between predicting the behavior of an individual and predicting the behavior in a city. Here’s a simple thought experiment. Suppose there are one million people in a city and one mosque. Suppose, if someone does not search for “kill Muslims,” there is only a 1-in-100,000,000 chance that he will attack a mosque. Suppose if someone does search for “kill Muslims,” this chance rises sharply, to 1 in 10,000. Suppose Islamophobia has skyrocketed and searches for “kill Muslims” have risen from 100 to 1,000.

In this situation, math shows that the chances of a mosque being attacked has risen about fivefold, from about 2 percent to 10 percent. But the chances of an individual who searched for “kill Muslims” actually attacking a mosque remains only 1 in 10,000.

The proper response in this situation is not to jail all the people who searched for “kill Muslims.” Nor is it to visit their houses. There is a tiny chance that any one of these people in particular will commit a crime. The proper response, however, would be to protect that mosque, which now has a 10 percent chance of being attacked.

Clearly, many horrific searches never lead to horrible actions.

That said, it is at least theoretically possible that there are some classes of searches that suggest a reasonably high probability of a horrible follow-through. It is at least theoretically possible, for example, that data scientists could in the future build a model that could have found that Stoneham’s searches related to Donato were significant cause for concern.

In 2014, there were about 6,000 searches for the exact phrase “how to kill your girlfriend” and 400 murders of girlfriends. If all of these murderers had made this exact search beforehand, that would mean 1 in 15 people who searched “how to kill your girlfriend” went through with it. Of course, many, probably most, people who murdered their girlfriends did not make this exact search. This would mean the true probability that this particular search led to murder is lower, probably a lot lower.

But if data scientists could build a model that showed that the threat against a particular individual was, say, 1 in 100, we might want to do something with that information. At the least, the person under threat might have the right to be informed that there is a 1-in-100 chance she will be murdered by a particular person.

Overall, however, we have to be very cautious using search data to predict crimes at an individual level. The data clearly tells us that there are many, many horrifying searches that rarely lead to horrible actions. And there has been, as of yet, no proof that the government can predict a particular horrible action, with high probability, just from examining these searches. So we have to be really cautious about allowing the government to intervene at the individual level based on search data. This is not just for ethical or legal reasons. It’s also, at least for now, for data science reasons.





CONCLUSION



HOW MANY PEOPLE

FINISH BOOKS?

Seth Stephens-Davidowitz's books