If you’re reading this, there’s a good chance you have a cookie on your computer or your smartphone that says how much you like weed—or at least how you feel about the idea of weed being legalized.
The cookie reflects a score that was calculated for you by CampaignGrid, a digital advertising company that has spent the last five years creating a database with information on over 120 million potential voters. It then uses this information to determine the likelihood of a given voter being pro-legalization. The higher your score, the more likely your opinion on pot is favorable. If you’re young, make over $85,000 and live in Maryland, your score is probably 85 or so. If you are registered as a Republican in a sparsely populated town in Wyoming, your score is likely lower than 50.
These scores weren’t just calculated for fun; they were used to get out the vote for legalization in Oregon, Florida and Alaska, all states where pro-marijuana groups had hired CampaignGrid to help them win over voters. If you lived in one of those states, and your score indicated you were likely in favor of legalization, you would have seen ads as you surfed the Internet that encouraged you to leave the hotbox and get to the ballot box. If your score was in a range that indicated you were “persuadable” on the issue, pro-legalization groups would have targeted you with ads about the benefits of marijuana, doctors’ assurances that it’s safe, and testimonials from people who want to be able to use it for medical reasons. If you had a low score … *crickets.*
Brian Franklin, a political consultant at Impact Politics, which was pushing medical marijuana in Florida for United for Care, says that his organization surveyed voters and found evidence the online ads had an effect on those who were undecided in the lead-up to the election. “People who were predicted to be undecided voters were moved by online ads,” said Franklin. “We hit who we wanted to hit and the ads were effective.”
Florida failed to pass legalization last fall by a hair, but pot was legalized in Oregon and Alaska. And though that’s not concrete proof that its methods worked, CampaignGrid is taking a victory lap anyway. “I think it’s fair to say we contributed to those successes, though there were other players involved as well,” says CampaignGrid president Jordan Lieberman.
The advertising industry has been using algorithmic targeting techniques to sell consumer goods to people on Facebook and Twitter for years, but the application of those techniques to politics could swing elections, change laws, and upend the old models of voter persuasion.
“This kind of behavioral and demographic targeting has been going on [in business] since the early 2000s,” said Jonathan Mayer, a lawyer and computer scientist who specializes in privacy and security issues as a graduate fellow at Stanford University. “The private sector is now targeting people using precise geotargeting and beacons, which political campaigns are just starting to dabble in it. The news—good or bad—is that they are catching up quickly.”
CampaignGrid and similar firms could play a large role in the elections of the future. Their data sets, coupled with predictive analytics, give them the ability to mobilize—and potentially sway—large groups of voters for any cause willing to pay for the technology. (This time around, CampaignGrid was only showing pro-legalization ads, but it’s a mercenary-for-hire, and could work with the other side in the future.)
This was the first campaign in which CampaignGrid firm rolled out what it calls “the marijuana cookie.” Understanding how CampaignGrid created that cookie, and how it assigns scores based on the collected data that results, reveals the planning, preparation and inexact science that goes into voter profiling in the digital age.
Fort Washington, Pa.-based CampaignGrid was founded in 2007, but didn’t get rolling until 2010. It now has a database of over 120 million people that contains up to 170 “attributes” per person. In a PowerPoint presentation obtained by ProPublica in 2012, CampaignGrid said that a file on a given voter might reveal the following: “Lives in Pennsylvania’s 13th Congressional District, 19002 zip code, Registered primary voting Republican, High net worth household, Age 50-54, Teenagers in the home, Technology professional, Interested in politics, Shopping for a car, Planning a vacation in Puerto Rico.”
Lieberman, who worked on a Congressional campaign and ran a magazine about campaigning before joining the firm in 2010, says the files aren’t quite as specific as that. He called the claims in the PowerPoint presentation a “typo” and said that the CampaignGrid employee who created it “used aggressive language about our ability to target.” He added: “It was very ambitious and inaccurate.”
According to Lieberman, much of the voter data in CampaignGrid’s files was purchased from other brokers—such as the political data-mining behemoth L2, which has a massive database of voting behavior, census data, and “exclusive lifestyle and issue data.” L2’s data might include the magazines you subscribe to, or your shopping habits as gleaned from the loyalty cards you use to get discounts. CampaignGrid’s big idea was taking this data, and layering on additional “attitudinal information” about voters; the company’s chief analytics officer Alex Gochtovtt defines this as “what motivates people to action.”
The marijuana cookie is an example of one of these attitudes. CampaignGrid started with a telephone survey in May 2014 of over 16,000 people across the country drawn from its database. Over 12,000 of those contacted agreed to answer questions about their politics, though they had no idea when they did so that they were being contacted on behalf of a company that already knew a lot of information about them.
“That has an unsavory quality to it,” says Mayer. “Calling you to ask you about the details in your file without telling you has a bad odor.”
But the survey wasn’t just about what was in potential voters’ files—it was about what wasn’t there. After peppering the subjects with general prompts, such as “Are you an active voter?” and “Are you likely to vote on referendum issues?”, the surveyors got to the question CampaignGrid couldn’t mine from an existing database: “How likely are you to support marijuana legalization?”
“We didn’t ask about whether their support was for recreational versus medicinal use,” says Lieberman. “We thought that would be privacy-invasive. Compliance with the survey goes up if it’s less intrusive.”
Gochtovtt says the survey found an equal split between people who were for, against, on the fence about, and ambivalent regarding marijuana legalization. They then passed the survey results to their data science team to come up with a way to predict how someone in their database would vote. Because the survey subjects were sourced from the company’s files, the data scientists had up to 170 other discreet pieces of data on each of these 12,000 respondents, which they used to come up with those people’s likely verdict. After their initial 12,000-person survey, CampaignGrid created an algorithmic model to score the millions of other people in the database using the same methods. The scores the model predicted matched the ones the respondents have given, which meant that the algorithm had worked.
Gochtovtt said that the most telling factors in the model were age and party affiliation—young Democrats, for example, tended to support marijuana legalization. (Yes, they really needed an algorithm to figure that out.) “Other factors” come into play in different geographical areas, though, says Lieberman.
Whether CampaignGrid’s techniques amount to a dark art, or a potential privacy nightmare, is a different question from whether they’re effective. While Lieberman says that the whole voter-targeting process is “privacy-respective,” CampaignGrid’s online advertising material trumpets the firm’s ability to target individuals, not just people who fall within a given demographic. “While other voter-targeted platforms typically use modeled voter data – which leads to inefficient and frankly inaccurate targeting – our approach allows you to target specific registered voters online,” says its website. (Emphasis ours.)
Gochtovtt walked that claim back a bit in a phone interview. While CampaignGrid does have individual names in its database, he said, the firm strips those names before they do online targeting. “We cannot target specific named individuals since we require a minimum of 1,000 individuals in a cell before we can target,” he said.
These marijuana-sentiment cookies, which CampaignGrid placed onto hundreds of millions of digital devices in total, weren’t easily detectable. They originated on a cookie-serving platform called Turn, which then used Google’s Doubleclick for Advertisers’ ad-serving platform to channel them into targeted ads. Gochtovtt says that a user’s score goes through so many layers of transformation that it’s essentially unrecognizable. “Scores get bundles, bundles get IDs, and those IDs get put out there,” says Gochtovtt. “I shouldn’t be able to go on my neighbor’s laptop and open their cookie and see what their score is.”
Mayer, the Stanford lawyer and computer scientist, was skeptical of CampaignGrid’s claim that it would be impossible for an outsider to figure out whether a person had been deemed a legal pot supporter by the algorithm. “Having reverse-engineered advertising segments in my research in the past, I haven’t found the process difficult,” he said. Mayer said it would be more difficult to do if the cookie values were encrypted. (They’re not; CampaignGrid says its campaigns “leverage our partners’ platforms and the security they use,” but the cookies themselves are encryption-free.)
For people who don’t want their drug preferences known by an advertising company, the scale of CampaignGrid’s data-gathering could be worrisome. Also worrisome is that users can’t view their own marijuana-sentiment cookies. Google has an ad preferences manager that says what gender, age, and interests it associates with you. Data broker Acxiom created a site in 2013 called Aboutthedata.com where users can see some of what the data giant knows about them. But there’s no way for a CampaignGrid target to do the same thing. Last year, the Federal Trade Commission asked Congress to force all data brokers to provide more visibility into their files and give people the ability to correct inaccuracies.
“We’ve talked about that,” says Gochtovtt. “Because we deal with such a varied data set, we’re not comfortable with that level of transparency, but we think it’s coming.”
Instead of letting people see their scores, CampaignGrid offers the opportunity to get rid of them through an advertising industry opt-out mechanism. So, take heed, privacy-conscious potheads opting out, Internet users can “just say no” to drug-attitude-based targeting.