Good ol’ Boys Club (top definition)
Male discrimination against females. Men who refuse to speak to, deal with, or work with women… — Urban Dictionary 2011

The 2019 Canadian Federal Election was held on October 21, 2019, to elect members of the House of Commons to the 43rd Canadian Parliament. During the election season, Facebook ads had been adapted by all candidate parties as a campaigning platform. All the ads are accessible through Facebok Ads Library.

In March 2019, Facebook launched the Ad Library — a direct response to the Cambridge Analytica scandal that hit the platform in early 2018 (Silva et al. 2020). Through its API, we are able to see who pays for an ad, the gender and age group distributions reached by the ad, the dates the ad is published and other information.

Through the analysis in gender distribution of all parties ads, we found that although both male and female (and all genders) citizens have the right to vote in Canada, the equal inclusion of the female audience in election campaigns is not reflected in all parties’ targeting strategies. The findings indicate that three out of the six parties (Bloc Québécois, Conservative Party, and PPC) have their ads reached substantially more male audience than female audience, despite that there were more female users than male users on Facebook in Canada.

Overall Spending

The table below displays the overall spending and impressions of each party. As the two biggest political parties in Canada, both Liberal Party and Conservative Party had spent more than 1 million for Facebook ads during the election season between June 1st, 2019 to October 21, 2019.

The Liberal Party has put 12199 ads, which is the largest amount of ads among all parties, and six times more than the Conservative Party. It is estimated that Liberal Party has spent at least more than 1.7 million on Facebook ads during the election season, and the ads had made at least more than 185 million impressions. The regional and demographic distribution of people reached by the ads is examined in the following sections.

Gender Distribution

According to Statista, at the time of the election season (October 2019), women accounted for 53.2 percent of Facebook users in Canada. The gender distribution of the ads does not reflect the gender ratio of Facebook users in Canada, as shown in Figure 2. PPC’s ads spending demonstrated the largest proportion of male audience; 74.73% of its ad spending has been distributed to reach male audiences. The ads of Conservative Party and Bloc Québécois also had reached more males than females. The Green Party, on the other hand, has 64.45% of ad spending for reaching female audiences, which is the highest among all parties.

Each ad has its own gender distribution. If an ad has reached more female audiences than male audiences, it is labelled as a “more-female-reaching” ad and vice versa. The number of “more-female-reaching” ads and “more-male-reaching” are counted for each party, which is presented in the table below.

The proportions are then calculated to have an idea if a party has extremely skewed gender distribution in their ads; the result is presented in the table bebow. The results show that PPC, Bloc Québécois and Conservative Party have a substantially large portion of ads that reached more male audiences than female audiences. Out of PPC’s 66 ads, only 1 ad reached more female audience than male audience. For Bloc Bloc Québécois, more-female-reaching ads only accounted for 10% of all their ads; only one-third of Conservative Party’s ads reached more female audience than male audience.

Gendered Ads Narratives

To further investigate in the differences messaging by gender, TF-IDF is conducted to extract keywords from more-female-reaching ads and more-male-reaching-ads respectively. The results are shown in the figure below.

For some parties, the segmentation of gender is obvious. For example, although there are overlapped keywords such as “tax” and “ahead”, Conservative Party stressed “maternity” benefits and helping “parents” in their more-female-reaching ads. In Green Party’s more-male-reaching ads, “Elizabeth May”, which is Green Party’s female leader was not on the keyword list; and instead of environmental issues such as “fracking”, they focus more on the economical keywords such as “economy” and “workers”. Liberal Party’s more-female-reaching ads stressed on strengthened gun controls (“ban” “military-style” “assault” “rifles” “restrict” “handguns”); their more-male-reaching ads seemed to focus on economic topic as well (“middle” “class”). NDP’s ads also have different messaging towards different genders. Their more-female-reaching ads have distinctive keywords “education”, “healthcare” and “housing”, whereas their more-male-reaching ads have “corporations”, “telecom” and “governments”.

The keywords of PPC’s more-female-reaching ad were generated from only one ad; the original creative body for this ad is “Watch our Veterans Policy announcement in Gatineau with 42 PPC veterans on stage. #PPC2019”. Since there is only one more-female-reaching ad, it is not clear if PPC had crafted the ads to deliver different messaging for female audiences.

So Is Canada’s politics still an old boy club?

For some parties in Canada, it might still is. Conservative, PPC and Bloc Québécois, had targeting strategies that resulted in reaching substantially more male audiences than female audiences. Because how interests or behaviours of the audience are targeted is not disclosed, we can not know exactly what had led to the larger portion of male audiences reached by the ads of certain parties. However, it raises the question if their targeting strategy indirectly reflect their hidden mentality towards women’s participation in politics.

Through TF-IDF keyword extraction, we also found that the ads have the tendency to speak about economy-related issues to men and environmental and societal issues to women. This gender bias in keywords and narratives might occur naturally also due to targeting specific interests that influenced the gender distribution of the people reached. The question raised by this phenomenon is if women are expected to “care” about politics for only certain concerns?

In short, the gender bias in distribution and messaging of the 2019 Canadian Federal Election on Facebook Ads reflects the implicit limitation of women’s participation in politics in Canada. Through the distribution landscape, we see that half of the parties are indirectly targeting a much larger amount of male audiences than female audiences. Through the different messaging, women (and men) more likely to be exposed to only certain topics and therefore polarize their social roles.

References

Facebook Ads Library: https://www.facebook.com/ads/library/

Silva, Marcio, Lucas Santos de Oliveira, Athanasios Andreou, Pedro Olmo Vaz de Melo, Oana Goga, and Fabricio Benevenuto. 2020. “Facebook Ads Monitor: An Independent Auditing System for Political Ads on Facebook.” In Proceedings of the Web Conference 2020, 224–34.

“Statista.” 2021. Statista. https://www.statista.com/.

The Guardian, The Guardian. 2019. Canada Election 2019: Full Results, October.

Woolf, Max. n.d. Facebook-Ad-Library-Scraper. Github. https://github.com/minimaxir/facebook-ad- library-scraper.

When a brand name is used as a verb, you know that brand has become ubiquitous in our lives. For many years now, “Googling” has become the primary way to retrieve information for many of us who lives in this highly digitalized society; and these search results we consume are shaping our world views. This blog post attempts to tell you how Google shapes our world views in problematic ways.

Problem 1. Perpetuating Racial and Gender Bias

The author of Algorithms of Oppression, Safiya Noble (2018), spent six years studying Google’s search algorithm and demonstrates that Google search results can reinforce and perpetuate gender and racial bias. She presented disturbing case studies where the keyword “black girl” prompted results encompassed websites and images of porn. Moreover, the search results constantly represent white people as the norm and standard while people of colour are depicted as the “others”. For example, when searched “white teenagers”, the image results showed positive wholesome depiction of white teenagers. On the other hand, when searched “black teenagers”, they were often depicted as gangsters.

While Google has improved a lot on those issues examined in the book, some deep stereotypes are still found from their search results today. When putting the keywords “boss at work” for Google Image Search, most of the resulting images represents “boss” as men. On the other hand, the keyword “secretary” not only prompted female-only images, some of them are explicitly sexual. These search results normalize two harmful narratives: men are dominant and authoritative; women are submissive and objectified.

Problem 2. Google’s Filter Bubble

In the names of relevance and personalized experience, Google manipulates the search results according to your personal data it collected through your search history, browsing history, shopping history, location, and even your email.

In other words, the information you are exposed to is what Google thinks you will prefer. As a result, point of views that are different than yours would be excluded in the search results, and we only see and hear what we like. Internet activist Eli Pariser calls this a filter bubble, which is a similar idea to “echo chambers”. The filter bubble’s effects are the most apparent when it comes to political ideologies, as when people are searching political topics, they are more likely to be presented with search results that are conforming with their existing political views. As a result, these existing political views are reinforced and even more polarized by the information presented.

Problem 3. Anticompetitive Practices

Google has dominated the search engine market for the last decade. As of January this year, Google holds an 91.86 percent Search Engine market share world wide (Statcounter, 2021). Although the majority of Google’s revenue comes from advertising, it has expand its markets to products and services that we also heavily depend on such us Gmail, Google Drive, Google Maps and many more.

October 20, 2020, the U.S. Justice Department filed an antitrust lawsuit against Google. The U.S. Justice Department lawyers accused Google of abusing its dominance in search engine market to enrich its own business empire and maintain its monopoly status. The Markup also showcases that top Google search results are often Google products. Inspired by The Markup’s work, I conducted an experiment to search for “flight Toronto” in four different search engines — Google, Bing, Yahoo and Duckduckgo.

Case Study: Searching “flight Toronto”

When I searched “flight Toronto” with Google Search, the first two results are ads from Air Canada and Flighthub. Right after the ads is the integrated interface of Google Flights, then is the real search results. What’s wrong with offering another Google service within the Google Search platform?

To see the wrongdoing of this practice, we have to consider a couple things —results presented first have higher chance to be clicked; a service that is presented in your face without a “click” has a higher chance to be used. Google Flights’s integrated interface is placed before the first organic search result. And although it’s placed after the ads, it doesn’t require user to click on a link. This is a way that Google can stifle competitors subtly but effectively. As The Markup stated, the effects of placing Google’s own products on the search page are stark:

In the nine years since Google Flights and Google Hotels launched, those sites have become market leaders. They garnered almost twice as many U.S. site visits last year as each of their largest competitors, Expedia.com and Booking.com, even though we found Google Flights doesn’t always show users all the options.

Results On Other Search Engines

I input the same keywords for three other search engines — Bing, Yahoo and Duckduckgo — and scraped the search results (method can be found in the last section). All three engines rank Expedia as the first result in the searches and none of the engines rank Google Flights in the top 5. However, by placing the interface directly in Google Search, Google Flight establishes a status of a “default” service to its massive amount of users and impacts the online travel industry abruptly.

The “Default” Bubble

It’s important to note that racial and gender bias exist not only in Google Search, but other search engines as well. The filter bubble is also not unique Google but also to Facebook and other platforms that operate on personalized algorithms.

What is unique to Google is its “default” status of not only one, but several essential services and products such as Google Search, Gmail and Google Maps. As U.S. Justice Department lawyers describe, Google has a role of the gatekeeper of the Internet, and such power is being abused to enrich their business empire by expanding the sense of default for other Google products in unfair ways. It’s really important to know that such conduct not only hurts Google’s competitors, but also hurts us as consumers as our cognitive mechanism is exploited to favour choices from a single source.

Method: Search Results Scraping by Scrapy-Splash

Discliamer: Since I am a beginner to web scraping, I might not be aware of more efficient ways to scrape search results. If you know a better way, please share!

Data Collection: Create Scrapy Spiders and Get Dynamic Content Through SplashRequest

To see the search results of “flight Toronto” in other non-Google search engines, three Scrapy Spiders were created to scrape first-page search results of same keywords on Bing, Yahoo, and Duckduckgo. Because the web pages are dynamically generated by JavaScript scripts which are executed by browsers. Because Scrapy is not a browser, it can not handle JavaScript scripts and the content would be empty. Therefore, Splash is used to handle JavaScript rendering.

Three separated Spiders for the three search engines are created respectively. Only the code for the Duckduckgo Spider is shown here, since all Spiders share a similar logic.

The Spider gathers the title, domain, and the index of each non-ad search result and saves it as a csv file searches_ddg.csv. The csv file is then loaded into a Python note book by Pandas.

import pandas as pd
sr_ddg = pd.read_csv("searches_ddg.csv")

The data frame looks like this:

Data Preprocessing

The last column result_index is zero-based, which means the initial result of the sequence is 0 rather than 1. For more a intuitive reading, I change it to one-based number by add each index with 1 and create a new column called result_rank.

sr_ddg["result_rank"] = sr_ddg["result_index"] + 1
sr_ddg.drop(['result_index'], axis = 1)

The final data frame looks like this:

Other keywords will be explored in future works. I have tried “hotel Toronto” and observed that Google places an integrated interface of Google Hotels in before the organic search results as well. If you have a suggestion for keywords to be explored, don’t hesitate!

The Google Search Bubble was originally published in The Startup on Medium, where people are continuing the conversation by highlighting and responding to this story.

Amazon’s Super Bowl commercial #BeforeAlexa was streamed to millions of viewers. In this 90-second ad, Alexa was portrayed as several obsolete roles in different historical scenarios — a maid that lowered room temperature by breaking a window of her master’s house; a joker that told a joke by deprecating himself to amuse the queen; a secretary received the command to “delete the tapes” for a President Nixon character. The commercial paints a picture that Alexa is now all in one our new maid, new joker and new secretary (the implied power relation is another topic). From the eyes of consumers, Alexa seems to save a good amount of labour for us today. However, if you have read Anatomy of an AI System, you will know that behind every execution of a simple command you give to Alexa, there is a massive chain of human labour that are out of our sights — the mineral extraction for making lithium-ion batteries, the manufacturing of electronic parts, the tagging and labeling of data sets, and more.

Lithium and Cobalt Mining

Lithium is a key material for making lithium-ion batteries, a type of rechargeable battery commonly used for smart devices and electric cars. The rise of smart electronics has led to the rise of lithium mining projects and complex supply chains. Lithium mining has sparked controversy for its environmental cost and the damage to the indigenous communities in some of the mining sites such as Bolivia. Moreover, according to Washington Post, a consumer company typically deals with multiple suppliers who also deal with multiple suppliers. The complexity of the supply chain makes the sourcing of lithium difficult to track. Cobalt is another key element for powering lithium-ion batteries. The demand for cobalt also led a massive number of mining projects in Democratic Republic of the Congo (DRC). Human right issues in cobalt mining operations are not rare. Amnesty reported that in some sites, the workers, including children, are working in harsh and dangerous conditions without protection for $1 USD per day.

Chinese Electronic Factory Workers

To lower the cost for electronic manufacturing, multinational companies find their supplies from China for cheaper labor cost. In 2011, China Labour Watch (CLW) revealed the horrible working conditions in ten Chinese electronics factories who assembles products for well known companies such as Dell, Salcomp, IBM, Ericsson, Philips, Microsoft, Apple, HP, and Nokia. According to the report, some workers were required to work overtime up to 160 hours per month. Majority of these factories did not offer wage that covers basic living costs. The intensity of working pace also lead to long-term occupational illness and injury.

Crowdsourced Online Workers

AI learn from data, and data need to be organized into training data sets. To compile training data sets, manually labeling, tagging and flagging thousands of data features are common. To save the cost of hiring someone to do this type of repetitive tasks, companies outsourced the tasks to people who are willing to take on these tasks for compensations that are far below minimum wages. Amazon’s Mechanical Turk is the largest crowdsourcing platform to the date. According to this paper, the median hourly wage earned by Turk workers is only about $2 USD per hour; only 4% of them earned more than $7.25 USD per hour.

And Us

Again, AI learn from data, including ours. As we are interacting with Alexa or other AI systems, our interactions are turned into data and transmitted to the companies behind the systems. In other words, as we are sending out commands to the AI and provide feedback about the accuracy of their responses, we are also providing labor.

AI Are Essential. Or Are They?

The commercial humorously exaggerated the inconveniences without an AI Virtual Assistant, and they are simply untrue. Without Alexa, we walk a few steps to turn up the knob of the thermometer; we ask a real human right beside to tell us a joke, and we set a reminder ourselves without the need of having a secretary. Behind each simple command, there is a massive labour that sustain an AI system that’s hidden from us. As a consumer, we should know that the stack required to sustain an AI system is not just data modeling, servers or networks; it goes further into capital, labour and nature. And as Kate Crowd pointed out, we should all start doing thought experiments to compare the the cost of the planetary labour network Alexa requires to play our favourite song versus the cost of us hitting the play ourselves.

AI and Their Human Slaves was originally published in The Startup on Medium, where people are continuing the conversation by highlighting and responding to this story.

What are the most common ways data breaches occur?

Not surprisingly, hacking is by far the most frequent used way to breach data. Poor security practices are the second most common way that data got breached. In this case, data are exposed not necessary by malicious attacks, but by, for example, being in the process of unproper transmission. Lost or stolen media (portable disk drive, laptop…etc) are the third in the list. The forth is accidental publication of data, usually by internal employees. Note that it has happened more than deliberate attempts of stealing data internally, which is the fifth on the list.

Which industries are the most affected?

The web industry is the most affected industry. “Web” as an industry is broad and vague. Having a look at the list, the companies categorized in web can actually be further categorize to more specific industries. What are they have in common is that they all offer their services or products through the Internet. Healthcare is the second most affected industry, followed by finance, government and retail.

Are there more and more data breaches each year?

The year of 2011 is the year that had the most data breaches happened. The trend went down to the lowest in 2017, and bounced back high in 2018. The trend went up until 2019 then the number of incidents dropped in 2020. Although I have received two data breach notifications this year, there is no incident recorded on Wikipedia, yet.

Method: Web Scraping by Scrapy

Disclaimer:

1. Wikipedia allows web scraping as long as the bot is not going to fast. See here.
2. Wikipedia actually offers their entire database for direct download. See here.

Scrapy and Beautifulsoup are two of the most popular web scraping tools today. Compared to Beautifulsoup, Scrapy is considered a full-ledge tool that allows bigger scale scraping project. Using Beautifulsoup is enough for scraping a single table from a single web page. I am using Scrapy because I want to be more familiar with this tool.

Data collection: Create a Scrapy Spider

Following the Scrapy tutorial, I created a simply spider that scrapes the table of the list of data breaches on Wikipedia. The scraped data is then stored as a JSON line object. The code to create the spider is the following.

After the scraped data is store, it is loaded to a Python notebook as a data frame for further exploration. I used read_json() from pandas to do the job. lines is set to be True because the JSON line file is separated by line.

import pandas as pd
df = pd.read_json("breaches.jl", lines=True)

Data Cleaning: Text Preprocessing

In the data frame, some of the values have extra ‘\n’ symbols to be removed. Also, the first row somehow is just a list of None values, we will remove those as well.

df = df.replace('\n','', regex=True)
df = df[1:]

The data after cleaning will look like this.

Data Exploration: Plot the data

A function is created to plot the bar charts. The arguments are the column, the colour of the bars, and the title of the plot. Then the function was called to showcase the incidents per industry and per data breach method respectively.

from matplotlib import pyplot as plt

def plot_counts_bar(col, colour, title):
  fig = plt.figure()
  fig, axs = plt.subplots(1, 1, figsize=(10,6))
  ax = col.value_counts().sort_values(ascending = False)
       [0:5].plot(kind='barh', color = colour)
  for i in ax.patches:
    ax.text(i.get_width()+.1, i.get_y()+.1,
    str(round((i.get_width()), 0)), fontsize=10, color=colour)
  plt.title(title)
  plt.xlabel('Counts')

# Incidents per method
plot_counts_bar(df['Method'], "#F35B04", 'Top 5 Data Breach Methods')

# Incidents per industry
plot_counts_bar(df['Type'], "#1653f3", 'Top 5 Affected Industries by Data Breach')

Finally, a new data frame is created. The new data frame df_year has two columns — the years and the counts of incidents in each year. The data frame is used to create a line chart that showcases the trend of the incident frequency over the years from 2004 to 2020.

# New data frame
df_year = df["Year"].value_counts().rename_axis('year').reset_index(name='counts')
df_year =  df_year.sort_values(by='year')

# Line chart
fig = plt.figure()
fig, axs = plt.subplots(1, 1, figsize=(10,6))
plt.plot(df_year["year"], df_year["counts"], marker='o', color = "#3DDC97")
plt.title('Data Breach Incident Counts Per Year')
plt.xlabel('Year')
plt.ylabel('Counts')
plt.grid()
plt.show()

Background

In 2018, two bills were set out to end territorial ratings — often referred to as “postal code discrimination” — in the automobile industry. Bill 44, titled Ending Automobile Insurance Discrimination in the Greater Toronto Area Act was struck down in the Ontario legislature on November 1st, 2018. Bill 42, Ending Discrimination in Automobile Insurance Act, remains alive. It passed second reading last March and is currently referred to the Standing Committee. Before bill 42 can become a law, postal code, which is highly correlated with income, race, unemployment rates and educational attainment, is legally used in risk classification algorithms and systems in Ontario.

Risk classification was developed to group people together according to the risks they represent, so people are protected from carrying more expense of someone who is more risk-prone. However, the classification relies on algorithmic systems that perpetuate discrimination by assigning levels of risk based on people’s area of residency, gender, marital status and more. Insurance companies justify such discrimination based on an individual’s area of residency, gender, marital status as “fair” because they are statistically significant indicators of risks. But what is “fair”? Actuarial fairness aims to function as a neutral and technical concept that serves the industry market. However, it can conflict with human rights principles when so-called “fair discrimination” perpetuate discrimination among society. Coming back to the automobile insurance example of using postal code as a risk factor in Ontario, premium costs are weighted more on drivers in higher-risk neighbourhoods. Most of these higher-risk neighbourhoods are also underprivileged neighbourhoods. Therefore, algorithms for calculating premiums based on postal codes are likely to result in “digital redlining” — practices that create and perpetuate inequities between marginalized groups through digital technologies. The seemingly “fair” discrimination of postal code failed to justify its validity in a social and civic context.

Postal code or area of residency is not the only problematic risk factor. According to the Financial Services Regulatory Authority of Ontario, presently, a personal automobile insurance profile is created based on the type of vehicle, driving record, annual mileage, area of residency, age, gender, and marital status. Only three out of the seven factors are directly related to “driving.” Age is a biological factor that people have no control with, and marital status largely depends on different views of family values. None of these factors should be burdened with discriminatory insurance pricing and should be removed from the automobile insurance algorithms.

Digital Redlining

In the 1930s, the US government sent appraisers to grade neighbourhoods in 239 cities, colour-coding them green for “best,” blue for “still desirable,” and yellow for “definitely declining.” The lowest grade was “hazardous,” which was marked with red lines on maps. The “redlined” areas were the ones financial service providers diminished as credit risks. Although the assessments were made by accounting for local amenities and home prices, the residents’ minority racial and ethnic demographics were the central part that contributes to a neighbourhood’s undesirability. Redlining today refers to the systematic denial of various services in selected areas by loan and insurance companies either directly or through the selective raising of prices. Because financial resources were harder to get in these areas, the redlines established a vicious circle for the neighbourhoods to stay underprivileged, even until today.

The effect of redlining in the automobile insurance industry in Ontario is perpetuated by modern automobile insurance algorithms. In the age of Big Data, insurance is evaluated by data analytic models and machine learning algorithms that are not as direct but may be even more damaging than colour-coded maps; this is because they are not limited by geographic boundaries and their inner works are often very complex and uninterpretable:

“Software that makes decisions based on data like a person’s ZIP code can reflect,
or even amplify, the results of historical or institutional discrimination.”

— ProPublica.org

We can see the effect in the Greater Toronto Area’s automobile insurance industry, which is the largest and most multicultural city in Canada. According to Kanetix, the average Toronto insurance rate is $1948. The highest estimated rate is $2590, and the lowest estimated rate is $1533, which results in a large gap of $1057.

The most expensive area consists of four postal codes that cover eight neighbourhoods out of 140 in the City of Toronto. Based on the demographic information of the eight neighbourhoods retrieved from the City of Toronto’s Neighbourhood Profiles, six out of the eight neighbourhoods are designated as Neighbourhood Improvement Areas (NIAs) — vulnerable communities with socio-economic struggles recognized by the City of Toronto. 7 out of the 8 neighbourhoods have a higher than average percentage of residents identified as visible minorities. Half of the neighbourhoods have higher than average percentages of households in poverty. 7 out of the 8 neighbourhoods have higher than average unemployment rates. Finally, all neighbourhoods listed have low higher education attainment compared to the City of Toronto as a whole.

It is evident that postal codes are not just codes for mailing routes but codes for income, ethnicity, and educational attainment. It interplays with multiple social dimensions and thus carries the baggage of historical and structural discrimination. Allowing postal code to be a risk factor allows discrimination in race, income, education level, and employment status.

Actuarial Algorithmic Discrimination

The postal code is not the only risk indicator that perpetuates discrimination. Under section 22 of the Ontario Human Rights Code, the law explicitly allows four exceptions in insurance contracts — age, gender, marital and family status or handicap — as long as they are bona fide and reasonable. Except for “handicap”, all exceptions are applied to automobile insurance in Ontario. According to the Financial Services Regulatory Authority of Ontario, presently in Ontario, a personal automobile insurance profile is created based on not only factors that are directly related to driving (type of vehicle, driving record and annual mileage), but also four discriminatory risk indicators including area of residency, age, gender, and marital status.

Age and Gender

Among these four factors, age and gender are innate features that people do not have control with. Although age and gender correlate with the number of automobile accidents,there is no evidence that these two factors are the direct cause of these risks. However, male drivers under 25-year-old are charged considerably higher insurance premiums because they are seem as higher risk drivers. It is evident that if starting to drive at the same age, older drivers would have more driving experiences contributing to lower risks; age is just a proxy for experiences. There is also no evidence that age is the cause of higher-risk and driving record alone should be the risk factor. Both age and gender are non-casual and immutable factors that should be not be included in the risk classification algorithms.

Marital Status

Unlike age or gender, marital status is a product of choice in general. However, it largely depends on an individual’s personal values and attitudes towards cohabitation and the individual’s life stage, which should not be burdened by discriminatory insurance prices. Traditionally, marriage is the favoured status because it is associated with traits of stability and responsibility. Statistically, married individuals do have fewer accidents than single people. However, marital status is not a direct cause of the risks. Algorithms that use marital status as a variable in determining the premium would contribute to a single cohabitation model’s normative status and the stereotypes that single individuals are unstable and irresponsible.

Area of residency, age, gender, and marital status have long been featured in the statistics used to determine automobile insurance premiums in Ontario. Although insurance companies claim that there are correlations between these characteristics and risk level, the causal connections are not proven. Therefore, instead of relying on an individual’s actual circumstances and behaviour, the algorithms’ assessment of individual risk is based on generalized assumptions and stereotypes of a group. As a result, discrimination on marital status, family status, age and gender are embedded and perpetuated in the actuarial risk classification algorithms.

Alternatives

Area of residency, age, gender, and marital status have long been the features used in actuarial algorithms for determining premium rate in Ontario’s automobile insurance industry; it is time to end such discriminatory practices. To change the status quote, four alternatives are proposed:

1. The inclusion of area of residency, age, gender and marital status as risk factors in actuarial algorithms should be made illegal by the law.
2. A shift from feature-based risk assessment to a behaviour-based ground is necessary.
3. All variables and their weights in determining insurance premium should be made understandable and auditable.
4. A legal definition of “fairness” in the automobile insurance industry should be stated to include the welfare of the society should be imposed.

Since individuals are assigned to a predetermined number of risk groups, they do not have the autonomy to demonstrate their unique characteristics that distinguish them from the average class members. As an alternative, the industry should adapt to a behaviour-based approach to replace the feature-based approach in conducting risk assessment in the automobile insurance industry. For example, “black box” insurance (or telematics insurance) has risen in recent years. “Black box” here does not refer to the mysterious inner working of an algorithm, but a physical device that is mounted to a car for a period of time to measure and record vehicle speed, location, traveled distance, frequency of driving, and time of day the car is in motion. Other driving performance indicators also measured, including how hard the brakes are applied, how rapid the acceleration level is, and how sharp of a corner is taken. In Ontario, black box insurance is mostly used for determining the “discount” over a determined premium. However, it is apparent that automobile insurance premiums should be largely, if not entirely, based on such personalized individual determination.

Due to the nature of complex actuarial algorithms and pricing models, transparency has been a challenge for consumers. EU’s General Data Protection Regulation (GDPR) mandates that users have the right to demand explanations of algorithmic decisions made for them, including insurance risk systems. There are technical difficulties associated with transparency in modern algorithms in which not even the creator of an algorithm understands the inner working of the algorithm. However, the industry can explicitly inform the consumers what variables are considered in risk classification and how changes of these variables contribute to the final premium determination. By identifying the relationships between inputs and outcomes, we can detect possible biases and get a direction for fixing the transparency problem.

Finally, the definition of “fairness” in the insurance industry should be expanded to the impact of the welfare of society and civilization instead of just the sustainability of the industry. To allow discrimination simply on the basis of statistical averages would only serve to perpetuate traditional stereotypes with all their invidious prejudices, which is absolutely unfair and unethical. Relying on historical data and current definitions of fairness in the insurance industry, the algorithms and the insurance products they support will always be establishing past norms, reflecting the accumulated unfairness, and paralyzing the social progress of our society. Fairness simply cannot be examined in a self-contained statistical ground. Fairness is a dynamic, social and civic issue that should be consistently audited, adapted and debated by all members of society.

References

About FSRA. (n.d.). Retrieved November 18, 2020, from https://www.fsrao.ca/about-fsra

BIG DATA AND THE ROLE OF THE ACTUARY. (2018). The American Academy of Actuaries. Retrieved November 12, 2020, from https://www.actuary.org/sites/default/files/files/publications/BigDataAndTheRoleOfTheActuary.pdf.

City of Toronto. (2018, December 05). Neighbourhood Profiles. Retrieved November 16, 2020, from https://www.toronto.ca/city-government/data-research-maps/neighbourhoods-communities/neighbourhood-profiles/

Discussion paper: Human rights issues in insurance. (n.d.). Retrieved November 12, 2020, from http://www.ohrc.on.ca/en/discussion-paper-human-rights-issues-insurance

Financial Services Commission of Ontario (n.d.). About FSCO. Retrieved November 18, 2020, from http://www.fsco.gov.on.ca/en/About/Pages/default.aspx

Ito, J. (2019). Supposedly ‘Fair’Algorithms Can Perpetuate Discrimination. Wired, April, 2.

Jan, T. (2018). Redlining was banned 50 years ago. It’s still hurting minorities today. Washington Post, 28.

Kagan, J. (2020, August 28). Black Box Insurance. Retrieved November 16, 2020, from https://www.investopedia.com/terms/b/black-box-insurance.asp

Kanetix. (2019, February 26). Kanetix.ca Reveals Ontario’s Most Expensive Cities for Auto Insurance. Retrieved November 12, 2020, from https://www.newswire.ca/news-releases/kanetix-ca-reveals-ontario-s-most-expensive-cities-for-auto-insurance-819924015.html

Kirchner, L. (2015). When big data becomes bad data. ProPublica, September, 2.

Understanding Auto Insurance Rates. (n.d.). Retrieved November 12, 2020, from https://www.fsrao.ca/consumers/auto-insurance/understanding-auto-insurance-rates

Wiegers, W. A. The Use of Age, Sex, and Marital Status as Rating Variables in Automobile Insurance”(1989) 39. UTLJ, 2, 149-at.

Discrimination in Ontario Automobile Insurance was originally published in The Startup on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Rise of 6G

In the 2030s, the world has entered a new paradigm of wireless communication. The sixth generation (6G) of mobile communication is widespread globally and offers data speeds that are 100 times faster than 4G LTE networks and five times faster than 5G networks (Alsharif et al., 2020). The mobile ultra-broadband of 6G has achieved terabits-per-second wireless transmissions and fuels the development of applications whose bandwidth requirements were not met by 5G such as holography and mixed reality (Zhang et al., 2019). Super Internet-of-Things (Super IoT) comes along with 6G and provides hyper-connected experiences to every corner of consumers’ life (ibid); personal data is constantly being collected and transmitted. On the other hand, the mass increase in computing power allow some technology bottlenecks to be unblocked. For example, increasingly realistic computer-generated imagery (CGI) is finally indistinguishable from real-world objects by human eyes. The rendering of CGI human characters reaches the level of high fidelity that is claimed to be crossing the uncanny valley. Additionally, the development of sophistificated NLP has also had a breakthrough; neural language models finally pass the Turing Test and are widely applied in different applications across industry.

Crossing the Uncanny Valley

With the rise in quality of CGI and NLP technologies, realistic and interactive virtual influencers and cyber models shine on the stage of social media platforms. Followers can interact with the virtual influencers by commenting on their posts, chatting with them in messaging channels, and even through newly developed holographic applications and mixed reality technology. Mixed reality, sometimes referred to as hybrid reality, merges real and virtual worlds and produces new environments and visualizations where virtual influencers and physical humans can coexist and interact in real time (Flavián et al., 2019). The interactive experience with virtual characters has never been this eloquent and mesmerizing. High engagement rates lead to a high volume of interactions, and massive amounts of data are collected from their followers. The data not only teaches the virtual influencers how to accurately portray the “ideal” to their followers, but also influences their ideas of “ideal”. The continual development and pioneering tech surrounding virtual realism is promising and continues to have strong corporate support.

Seeing Will No Longer be Believing

The combination of CGI and AI in virtual humans, leading to their increasing accessibility, is both exciting and concerning. As technology advances in CGI are made, a key requirement is to reduce the amount of hand-editing that is needed in the current process. In the future, anybody can produce content for little or no cost to a degree of fidelity that is only possible for Hollywood studios in 2020. This hints at a future in which all of us can produce virtual humans in astonishing fidelity and place them into digital worlds with unparalleled levels of detail (Dormehl, 2019). Moreover, since the virtual characters can be created with any appearance and given any personality, we may see a change in diversity and representation in all industries that involve virtual influencers. On the other hand, negative outcomes will come along. The democratized access to sophisticated technology may lead to abusive use in creating fake personas who act as a compelling tool for spreading disinformation. This is already happening in 2020, where deepfakes or synthetic media are created to mislead viewers with highly realistic digital content (Tolosana et al., 2020). In the 2030s, it will be even harder to distinguish the fake from the real. Seeing will no longer be believing. The ability to generate realistic imagery at scale will have a serious impact on how modern societies think about evidence and trust. Moreover, the freedom to create diverse personas for virtual influencers will play to companies’ advantage by appealing to different audience segments. Rather than hiring one influencer that is appealing to multiple audience segments, companies can create multiple virtual influencers with whom all audience segments can identify themselves — thus resulting in exploitation of audiences in the name of diversity.

Exploitation of Human Emotions

These probable negative outcomes urge discussions on the surrounding ethical issues. The realism of virtual characters is a breeding ground for deceptive practices. For example, virtual influencers may be portrayed as real humans who have their own agency, allowing the company to stay behind the curtain. Moreover, when registering an account, the personal information of a virtual influencer is fabricated, and thus within a legal grey zone, considering they are designed for profit-making. Putting aside the realism of the virtual influencers that can fool the human brain, humans have been known to form emotional attachments to fictional characters (Keen, 2006). Unlike the fictional characters in books or movies that have their destiny set, or characters in video games that are pre-programmed, virtual influencers’ personas are constantly changing with the goal to engage. In the future, the amount of personal data transmitting online through our interactions on social media platforms and the advancement of data analytic techniques will permit companies to know us even better—which restaurants we go to, who our best friends are, which styles we dress in, what news stories catch our eyes, and many more details like our mood, taste, and even values and worldviews. The virtual influencers are then constantly evolving based on the personal data they are fed. Therefore, how they dress, how they speak, which events they attend, who they are friends with, and their biological features—every aspect of their creation is designed to win an emotional response that leads to engagements. The calculated display of characters is clearly an exploitation of human emotions.

In preferable scenarios, policies and regulations are made to impose transparency in the usage of virtual influencers. Social media platforms such as Facebook, Instagram, TikTok, and Twitter have gradually implemented business accounts that are separated from personal accounts. However, it is evident that they should also start separating real human accounts from virtual character accounts, and they should make it mandatory to reveal the companies they are representing.

Contributors

Thank you Paul Hodgetts, Hidaya Ismail, Diego Mamanche Castellanos and Thomas William Rosenthal for sharing your ideas. Thank you Thomas William Rosenthal and Carli for reviewing and editing.

References

Alsharif, M. H., Kelechi, A. H., Albreem, M. A., Chaudhry, S. A., Zia, M. S., & Kim, S. (2020). Sixth Generation (6G) Wireless Networks: Vision, Research Activities, Challenges and Potential Solutions. Symmetry, 12(4), 676.

Dormehl, L. (2019, August 05). Cgi Faces Will Soon Be Indistinguishable From Real Ones. Here’s How. Retrieved December 16, 2020, from https://www.digitaltrends.com/cool-tech/cubic-motion-scanning-technology/

Flavián, C., Ibáñez-Sánchez, S., & Orús, C. (2019). The impact of virtual, augmented and mixed reality technologies on the customer experience. Journal of Business Research, 100, 547–560.

Fowler, D. (2018, April 2). The fascinating world of Instagram’s ‘virtual’ celebrities. Retrieved from https://www.bbc.com/worklife/article/20180402-the-fascinating-world-of-instagrams-virtual-celebrities

Halie, J. (2020, September 18). What are the ethical implications of CGI influencers promoting brands? Retrieved from https://talkinginfluence.com/2020/09/18/ethical-implications-cgi-influencers/

Keen, S. (2006). A theory of narrative empathy. Narrative, 14(3), 207–236.

Tolosana, R., Vera-Rodriguez, R., Fierrez, J., Morales, A., & Ortega-Garcia, J. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. arXiv preprint arXiv:2001.00179.

Zhang, L., Liang, Y. C., & Niyato, D. (2019). 6G Visions: Mobile ultra-broadband, super internet-of-things, and artificial intelligence. China Communications, 16(8), 1–14.

On October 6, 2020, the The House Antitrust Report on Big Tech led by staffers for the House Judiciary Committee’s in the United States is released, with a conclusion that Amazon, Apple, Facebook, and Google’s anti-competitive stance has “hindered innovation, reduced consumer choice and weakened democracy.” (Fung, CNN Buisness, 2020) In other words, the monopoly status of the tech giants has impacted the society negatively. As a consumer, it is not obvious to see the negative impact on a personal or individual level. One might even welcome the penetration of their lives by the Big Tech. Google, for example, provides services that cover almost all aspects of one’s life from email, search engine, navigation, to health — all for free and all so convenient.

In essence, at the core of their products, it is “convenience” that the Big Tech is selling to consumers. They deliver goods to our doors the next day, they tell us which route is the best to get to our destination, and they connect us with family and friends that are thousands of miles away — they are hard to resist. The ability of the Big Tech to provide the utmost convenience is a combination of their state-of-art artificial intelligence systems and their possession of big data collected by such systems. They have achieved what Yang Qiang (2018) called a “closed loop” — services provided by an information system generate data that are fed back into the system for self-refinement. In the case of the Amazon Echo smart home device, users purchase a device that comes with a set of convenient affordances. In using the product, the users then become a resource for the company as their voice commands are collected (Crawford & Joler , 2018), revealing their preferences, desires, and behaviour patterns. To get the access of the free and convenient services, we consumers participate in the loop and welcome the companies to collect our personal data in the name of improving the systems.

However, the compromise is not just on a personal level. As Crawford and Joler (2018) stated, “every small moment of convenience requires a vast planetary network, fueled by the extraction of non-renewable materials, labor, and data.” Coming back to the example of Amazon Echo, the amount of resources needed to execute the commands is much greater than the energy and labour it would take to switch the lights or play the music. The effort of building and maintaining the stack of these digital systems is not just virtual, but is embodied in the chain of exploitation in nature and people. As a consumer who interacts with the products but not the production chain, we are put in the position to focus on the functional aspects of the products and the elevation of life quality, but not so much on the social consequences of the AI products on a larger scale. The closed loop of AI products that we participate in with the Big Tech is not just a loop for self-refinement, but a vicious cycle that makes profits for the companies, provides conveniences to the consumers, but is unhealthy to the society as a whole (Bengio, 2020).

According to Yang Qiang (2018), the Third Industrial Revolution was triggered by the internet and mobile technologies; driven by big data, AI technologies are inciting a Fourth Industrial Revolution. It can be argued that whoever possesses AI technologies has the power to shape our behaviours and values. Bengio said that science in AI has reached a level of maturity that makes it very useful for companies. In fact, they are using AI to form an enormous ecosystem that has a mechanism to strengthen their kins by default. As a consumer, we should know that the stack required to sustain an AI system is not just data modeling, servers or networks; it goes further into capital, labour and nature. The cost of the systems are not just our own personal data on an individual level, but on societal, environmental, economical and political levels (Crawford & Joler , 2018).

References

Chao, W. (2018). Yang Qiang: The Fourth Revolution. UNESCO Courier, 22–24.

Fung, B. (2020, October 10). ‘Near-perfect market intelligence’: Why a House report says Big Tech monopolies are uniquely powerful. Retrieved November 2, 2020, from https://www.cnn.com/2020/10/10/tech/apple-amazon-facebook-amazon-monopoly-data/index.html

Kate Crawford & V. Joler. (2018). Anatomy of an AI System: The Amazon Echo as an Anatomical Map of Human Labor, Data and Planetary Resources. https://anatomyof.ai/.

Šopova, J. (2018). Yoshua Bengio: Countering the monopolization of research. UNESCO Courier, 18–19.