Sep
10
Tulsi Gabbard is right – Google is biased


On the night after the first Democratic debates,
Tulsi Gabbard was the single most Google-searched candidate by far [1]. A perfect opportunity for the new candidate
to raise crucial donations and climb up in polls in order to qualify for the next debates. But that would be too easy. In the heat of Tulsi’s massive surge in
popularity, Google suspended her advertising account for 6 hours and so drastically limited her ability to direct newcomers to her campaign website[2]. Tulsi is asking for $50 million to compensate
for damages. Google’s response?”We have automated systems
that flag unusual activity on all advertiser accounts… …and we do so without bias toward
any party or political ideology.” They have an algorithm and it is unbiased. The classic argument goes that machine-learning
algorithms are mathematical, and by their very nature, neutral and unbiased. But this unchecked theoretical view of engineers
in Silicon Valley crumbles in reality. In our reality, algorithms reinforce biases
they learn about from their training data. The new invisible hand of the modern discourse
are the machine-learning algorithms that are used by tech companies to recommend us shopping items, organize social media feeds, or personalize search results. These algorithms start off with a small set
of very simple instructions and programmers then feed them pools of data to learn from
on their own. Machine-learning algorithms are good at navigating
complexity – much more efficiently than humans. They can quickly skim through large databases
and prioritize certain values over the others. Today, algorithms are increasingly more often
entrusted witdh critical decision-making, including court sentencing, granting loans
and benefits, and even hiring for jobs and academic placements. [7] But there is a catch. Much of the development and implementation
of algorithms happens in secret. Their formulas are proprietary and users rarely get to even know the variables that make up their equations. Often times machine-learning algorithms make
decisions that not even their developers can understand why and how they arrived to them
and yet they just seem to work. [7] But mathematics cannot solve everything. The result of machine-learning algorithms
is solipsistic homogeneity – a process of finding associations, grouping data into categories
and creating a structure of sameness. The training data is always paramount to any
algorithm. If social or political biases exist within
that data, the algorithm is most likely going to incorporate them. Often times, it’s the historical data that
carry negative social footprint into the automation. In 2018, Amazon was looking for a way to automate
its hiring system. To recruit new engineers more quickly, an
artificial intelligence was developed. The system would scan through past resumes
and search for the best candidates on the web. But because the historical data showed predominantly
male resumes, the AI “learned” men are preferred to women. The algorithm automatically downgraded all CVs with the term “women’s” or attending women-only schools. When Amazon learned about this they tried
to repair the algorithm but soon they found out no matter what they did, it would always
find new forms of bias. So they decided to kill the algorithm and
return to the traditional hiring methods. [11] Similar to Amazon’s hiring AI, Google’s advertising algorithm also mirrored cultural
biases of historical data. A study found that the system shows ads for
high-income jobs to men disproportionately more often than it does to women. [8]
In other cases, users can attempt to feed the algorithm with biased information and
manipulate its outcome. Not so long ago Google Search autosuggest
feature used to rely heavily on user-input data. Until users learned how to easily game the
system to manipulate its rankings or just to troll the search engine with a cesspool
of bigotry. So Google made a decision to drastically interfere
with its search algorithm removing entire dictionaries of non-advertiser friendly terms. [10] Artificial intelligence is also used to predict criminal behavior that judges rely on to determine
their sentencing. But not even this realm is immune to algorithmic
biases. One such widely used algorithm flagged African
Americans as higher risk although they didn’t re-offend twice as mush as white Americans. Similarly, white Americans were labeled lower risk but did re-offend twice as much as African Americans. [9] Machine-learning algorithms are still very weak at understanding nuances of human language. Under the pressure from advertisers, YouTube
cracked down on extremist content by automatically flagging and demonetizing videos containing
a whole vocabulary of key words. But the algorithm is not capable of differentiating
between content that is truly extremist and one that is educational or merely reporting
on it. YouTube’s workaround was to give mainstream
media an exclusive pass, automatically alienating independent creators and journalists in the
process. [12] [13] The success of machine-learning algorithm stands and falls on the availability of good
data. The catch is there will always be less information
about minorities which will always lead to higher likelihood of invalid statistical patterns
about minorities. [14] A perfect manifestation of this reality,
is Amazon’s facial recognition tool that misidentified women for men 19% of the time and brown and black women for men up to third of the time. [15] [16] Not always is it the algorithm that should be blamed for all the bias. Sometimes corporate or organizational interest
of its creators can hugely interfere with its delivery. As Google grew to become a dominant search
engine worldwide, it slowly began offering more and more services that directly competed
with the market of providers that relied on Google search to reach their customers. [5 a,b] When the company launched Google Finance,
it began prioritizing it over the organic search results for relevant key words, even
though Yahoo Finance claimed the title of being most popular among users. This practice then expanded to Google Health, Google Reviews, Maps,video, travel and bookings and email. Prioritizing its own products allowed Google
to steal up to 34% of the search traffic. Now that percentage is even higher, as Google
Search offers instant answers and a wider range of Google products that make users stay
on Google longer and thus generate more ad revenue for the company. [17] [18] [19]
This is not a critique of whether Google should be allowed to push its own products as a private
company. Rather, it’s to show yet another vector
for bias to sneak into the algorithm and show that its search engine is not as neutral as
Google would have you believe. Corporate bias is a powerful factor. And corporate bias is especially important
to political insiders. Long time Google Executive Eric Schmidt has
been working hand-in-hand with the Democratic party, both with Obama and Hillary Clinton
campaigns. There was a lot of effort from Google insiders
trying to get Hillary Clinton elected. This included implementing features that would
manipulate Latino vote in key states or investing in startups and groups that would support Clinton campaign with technology, data and advertising. [20] [21] Tulsi Gabbard probably doesn’t enjoy the same level of insider connection with one
of the most influential tech companies in the world. So whether temporary suspension of her account
in a critical moment was just an error of the algorithm or was intentional, is a speculation
at this point. Had Tulsi had people on her side at Google
headquarters, this suspension might have never taken place or would have been much shorter. Google is refusing to give answers to crucial
questions: What variables triggered the automated system
to suspend her account? Was it flagged by the algorithm and then suspended
manually? Or was the decision made by the algorithm
alone? What unusual activity led to the algorithm
flagging Tulsi’s Ads account? Spending significantly more on ads on Google
after she became the most searched candidate could only be expected as the most rational
move a presidential candidate could make. Definitely not an unusual activity. Everybody’s strategy would be to capitalize
on the search traffic. It’s very difficult to understand the reasoning behind suspending her account under these circumstances. This practice of unaccountable moderation
is an industry standard across all major social media platforms [3 a,b]. Routine censorship raids on social media gave the right the argument to accuse Silicon Valley of liberal bias. [4] Whatever the case is, the presence of bias is undeniable. Algorithms are mathematical, but they can
only learn from people. A good step forward would to be admit the
bias exists and open up the source code of the machine-learning algorithms, so that we
can study these biases in real time as they arise. Secret development of artificial intelligence
by unaccountable tech corporations is a recipe for dystopian control of the information flow
and monopolization of Internet markets. Tulsi Gabbard learned this the hard way.