Understanding the Complexity of Detecting Political Ads

Online political advertising has grown significantly over the last few years. To monitor online sponsored political discourse, companies such as Facebook, Google, and Twitter have created public Ad Libraries collecting the political ads that run on their platforms. Currently, both policymakers and platforms are debating further restrictions on political advertising to deter misuses. This paper investigates whether we can reliably distinguish political ads from non-political ads. We take an empirical approach to analyze what kind of ads are deemed political by ordinary people and what kind of ads lead to disagreement. Our results show a significant disagreement between what ad platforms, ordinary people, and advertisers consider political and suggest that this disagreement mainly comes from diverging opinions on which ads address social issues. Overall our results imply that it is important to consider social issue ads as political, but they also complicate political advertising regulations.


INTRODUCTION
Social media and the public sphere's digitalization have changed the political campaigning landscape in both good and dangerous ways. While social media are creating new opportunities for engaging citizens in political conversations, they have also raised several risks for the integrity of elections and the political debate. For example, online ads can be tailored to specific groups of people, hence polarizing the voter base and distorting the political debate. Advertisers can buy large amounts of ads to flood people's social media feeds and steer public debates on issues that are of interest to them. Anyone, from political parties to interest groups, and specialized advertising companies such as Cambridge Analytica, can steer the political debate through online advertising.
Ad platforms have put forward several measures to mitigate risks and allow for public scrutiny of ads. Twitter and TikTok decided to ban political ads altogether. Google and Facebook allow political ads, but advertisers are subject to a higher degree of scrutiny and limitations. On Google, advertisers can only use geographic location, age, gender, and contextual targeting to target political ads. Facebook does not restrict the micro-targeting of political ads. Advertisers, Measures from both ad platforms and governments are positive developments. However, all of them implicitly rely on the assumption that one can reliably distinguish political ads from non-political ads.
In this paper, we take an empirical approach to test this assumption by analyzing the characteristics of ads deemed political by ordinary people, the characteristics of ads that lead to disagreement, and whether there are differences between what advertisers consider political and what ordinary people consider political. Our analysis is based on a dataset from ProPublica that contains 55k Facebook ads received by U.S. residents, labeled by at least one volunteer as political, and that received three or more votes (Section 2). The dataset was collected by a browser extension that collects the ads users see when they browse their Facebook timeline and allows users to label whether the ads they see are political.
First, we investigate whether ad platforms, volunteers, and advertisers agree on which ads should be considered political (Section 3). All ad platforms agree that ads from or about political actors and ads about elections and voting should be considered political. However, only Facebook and TikTok consider ads about social issues (such as climate change or immigration) as political. Our results show that volunteers disagree on whether an ad is political for more than 50% of the ads in the dataset, and only 83% of the ads labeled as political by advertisers are also labeled as political by a majority of volunteers. Hence, the fundamental assumption that we can clearly distinguish political from non-political ads does not hold, since there is no consensus even on what constitutes a political ad, and volunteers and advertisers label different sets of ads as political.
Next, we analyze the characteristics of ads that are labeled as political by volunteers and advertisers in the ProPublica dataset, which can be useful to inform the debate on definitions of political ads (Section 4). To that end, we gathered data about the advertisers arXiv:2103.00822v1 [cs.CY] 1 Mar 2021 sending political ads and the content of their ads. We hired Prolific users to annotate 2300 ads with the political or social issues the ad is referring to. Our analysis shows that a wide range of advertisers (from political actors to NGOs and businesses) are posting political ads on Facebook and that ads about social issues account for a large fraction of the ads labeled as political; hence emphasizing the importance of including such ads in political ads definitions. Our analysis also shows that the ads labeled as political by volunteers and advertisers are very diverse. We see ads with a clear political message from advocacy groups (e.g., ads addressing abortion issues in the U.S.); but also ads from NGOs that address humanitarian issues and do not seem to directly or indirectly impact U.S. elections or legislation (e.g., ads asking for donations for ending world hunger). As political ads may be subject to higher restrictions, this questions whether it is desirable that the same restrictions apply to both types of ads. More generally, this emphasizes the need to account for the diversity of political ads in devising regulations.
We finally analyze the ads that lead to disagreement among volunteers and between volunteers and advertisers (Section 5). We first observe that advertisers mislabel some ads as either political or non-political (according to the Facebook ToS). Then we find that advertisers seem to underreport ads (that are considered political by volunteers) about social issues, especially the economy and civil and social rights. Volunteers seem to underreport ads (considered political by advertisers) from advertisers such as NGOs and charities, and about social issues, especially civil and social rights and health. Part of the problem may be that the definition of ads about social issues may be too broad and vague, which leads to being interpreted in different ways by people. This also raises the question of whether all ads related to social issues should be considered political, and if not, how to filter social issue ads that are not political.
Because of the high volume of ads, enforcement mechanisms need to rely on automated machine learning (ML) algorithms to detect political ads. However, it is not clear how one should train and evaluate such models since there is disagreement on which ads are political (i.e., the positive examples). To investigate that, we train four classifiers with different groups of positive examples (coming from advertisers and volunteers). We test how they perform over various groups of political ads with varying degrees of disagreement (Section 6). While all classifiers achieve high accuracy in detecting ads everyone agrees are political; their accuracy drops on ads that only a few find political.
Another important question is whether (and to which extent) models trained with labels from advertisers would declare as political the same ads as models trained with labels from volunteers (i.e., reliable detection of political ads). Theoretically, if ads labeled as political by advertisers and volunteers are representative of political ads in general, the resulting models should declare the same ads as political. Our results show that the overlap between different models is relatively high (ranging from 83% to 97%), but that discrepancies in the input data transfer to discrepancies in the output data. This suggests that existing labeled datasets are not providing a representative set of political ads needed to build reliable detection schemes.
Overall, our work suggests that, given the complexity of deciding which ads are political, it would be beneficial to have ad libraries that contain all ads running on the platform, not only ads deemed political by the ad platform. Following this work, we issued a statement together with civil societies asking for "Universal advertising transparency by default" that we submitted to the DSA consultation [17]. However, this crucial first step is not enough because political ads are also subject to higher restrictions; hence, we still need to detect political ads reliably. We hope this study can help policymakers to define political speech and decide on appropriate restrictions and ad platforms to set infrastructures for detecting political ads.

DATASETS
For our analysis we use the following two datasets of ads that users have received on their Facebook timeline: ProPublica dataset. ProPublica, an investigative journalism organization, has developed a browser extension that collects the ads users are receiving on Facebook and allows users to label whether the ads they are seeing are political or not [27]. The extension is currently maintained by the NYU Online Political Transparency Project [25]. While ProPublica was not able to make available all the ads it has collected, it shared with us all the ads for which at least one user has labeled it as being political, as well as all the ads that have the "Paid for by" disclaimer (i.e., the official political ads that have been declared as such by advertisers). This dataset is valuable because it provides us with a unique view of which ads are considered political by "ordinary" people/volunteers. To our knowledge, there are no studies of such data.
For this study, we only kept ads with at least three votes (either political or non-political) and that were received between June 2018 and May 2020; resulting in a dataset of 54.6k ads coming from 7530 advertisers. The median number of votes per ad after filtering is 5. We call the ads that have the "Paid for by" disclaimer the official political ads and the ads that do not have the disclaimer the official non-political ads. Table 1 shows the number of ads in the ProPublica dataset as well as the fraction of official political ads and official non-political ads. Note that this dataset does not contain a representative sample of political ads as they are ads received by people who answered ProPublica's call for action to install the tool.
AdAnalyst dataset. Similar to the extension provided by ProPublica, AdAnalyst collects the ads users see on their Facebook timeline [2]. The AdAnalyst dataset contains over 500k ads from users in various countries. For this study, we keep only ads in English (detected using text-blob python library [35]) and that targeted users in the US between October 2018 and May 2020. For this, we use information about ad targeting available in the "Why am I seeing this ad?" button and select only ads targeted at people who live in the USA or visited places in the USA recently. The resulting dataset contains 9k unique ads (198 ads with "Paid for by" disclaimer and 8802 without). This dataset does not have votes from volunteers.
Ethical review board and reproducibility. Both data collection by ProPublica and AdAnalyst were approved by the respective ethical review boards. The ProPublica data is available to the public through a request form [28]. The 9k ads from AdAnalyst, the data collected from the Prolific studies, and other supplemental material is available at http://lig-membres.imag.fr/gogao/www21.html.

DISAGREEMENT ON POLITICAL ADS
The base to detect political ads reliably is to agree on which ads should be considered political and which ads should not. In this section we look at whether ad platforms, volunteers, and advertisers agree on which ads are political.

Disagreement across ad platforms
The Terms of Services of different ad platforms provide information on which ads they consider as political. We review the definitions and restrictions for political advertising across four ad platforms. Facebook defines political ads as: "Made by, on behalf of, or about a candidate for public office, a political figure, a political party, or advocates for the outcome of an election to public office; About any election, referendum, or ballot initiative, including "go out and vote" or election campaigns; About social issues in any place where the ad is being placed; Regulated as political advertising." The social issues are: civil and social rights, crime, economy, education, environmental politics, guns, health, immigration, political values and government, security, and foreign policy [14].
Everyone with a Facebook account can be an advertiser if they provide a payment method. However, to be able to send political ads, advertisers need to verify their accounts by providing proof of their identity [15]. Advertisers can only send political ads in the country they reside and need to provide proof of residence. Advertisers need to self-label their ads as political and need to provide a disclaimer about who paid for the ad. This "Paid for by" disclaimer appears on the top of the ad frame, after the advertiser's name. Finally, Facebook adds the political ads to their Ad Library [10]. Google defines political ads as: "ads about political organizations, political parties, political issue advocacy or fundraising, and individual candidates and politicians" [18]. The platform imposes no restrictions on political ads, but the platform expects all political ads to comply with local legal requirements. Google considers election ads as a separate category. The definition of election ads depends on the country, but overall it refers only to ads from or about candidates and political parties during an electoral period. Only verified advertisers can run election ads. Election ads can only be targeted by geographic regions (but not by radius around a precise location), age, gender, and contextual targeting options such as ad placements, topics, keywords against sites, apps, pages, and videos. Twitter defines political ads as "ads with political content: that references a candidate, political party, elected or appointed government official, election, referendum, ballot measure, legislation, regulation, directive, or judicial outcome; as well as ads of any type by candidates, political parties, or elected or appointed government officials" [36]. Twitter bans all political ads.
TikTok defines political ads as ads that promote or oppose a candidate, current leader, political party or group, or issue at the federal, state, or local level -including election-related ads, advocacy ads, or issue ads [6]. TikTok bans all political ads. Overall there are three categories of political ads: ads from or about a political figure or political party, ads about elections, and ads about social issues. While the precise definition of political ads varies across ad platforms, the most significant difference is that Twitter and Google do not consider ads about social issues as political while Facebook and TikTok do. While it is certainly a debatable question whether or not social issue ads should be regarded as political, the EU Code of Practice on Disinformation mentions both issue ads and political ads as sensitive content. Our results will show the importance of considering social issue ads as political and why they complicate political advertising regulations.

Disagreement among volunteers
At least three volunteers have labeled each ad in the ProPublica dataset as being political or non-political. The volunteers were given no instructions for what ads they should consider as political, and users were left to decide based on their instinct.
To observe to which extent volunteers agree on what ads are political, Figure 1 plots the distribution of the number of political votes divided by the number of all votes for each ad in the ProPublica dataset. We denote this fraction as fr. A fraction fr = 1 means that everyone agrees that the ad is political, while a fraction fr = 0 means that everyone agrees that the ad is not political. The plot shows that for more than 50% of the ads, at least one volunteer disagrees with the others (fr is neither 0 nor 1), which shows that deciding whether or not an ad is political is debatable for more than half of the cases.
To distinguish ads on which users agree they are political from the rest, we split the ads into four disjoint ad groups based on the volunteer votes. We will analyze them separately in the paper. The groups are defined as follows: • strong political ads: ads with fr = 1, i.e., where everyone agrees that they are political; • political ads: ads with0.5 ≤ fr < 1, i.e., where there is some disagreement, but the majority labels them as political; • marginally political ads: ads with 0 < fr < 0.5, i.e., where there is some disagreement, but the majority labels them as non-political; • non-political ads: ads with fr = 0, i.e., where everyone agrees that are non-political.
There are 26k strong political ads, 19.7k political ads, 7.6k marginally political ads, and 1.3k non-political ads.

Disagreement between volunteers and advertisers
The ProPublica dataset provides data on whether an ad was labeled as political by the advertiser itself (see Section 2). Table 2 presents the overlap between ads labeled as political by volunteers and ads labeled as political by advertisers (the official political ads). The table shows that 96% of strong political ads, and 93% of political ads were also declared as political by advertisers. Hence, most ads considered political by the majority of volunteers are also considered political by advertisers. There are, however, 4% of strong political ads and 7% of political ads that advertisers did not label as political.
The more surprising finding is that advertisers label as political a large majority (74%) of marginally political ads. Looking the other way around, 83% of official political ads are labeled as political by most volunteers. In contrast, 15% of official political ads are only labeled as political by a minority of volunteers, and 2% of official political ads are not labeled as political by any volunteer. Hence, many ads considered political by advertisers are not regarded as political by volunteers. While the reasons can be diverse (this is the subject of Section 5), we conclude that there is currently a significant discrepancy between the ads labeled as political by advertisers and by volunteers.

Takeaway:
The assumption that we can clearly distinguish political from non-political ads does currently not hold as there are significant disagreements between ad platforms, volunteers, and advertisers on which ads are political. Therefore, it is problematic to apply restrictions on political ads if the decision of whether an ad is political depends on the person labeling it.

WHAT GETS LABELED AS POLITICAL
This section provides a general view of ads labeled as political by volunteers and advertisers and analyzes who sends them and what are they talking about. This analysis is relevant for informing the debate on definitions of political ads and understanding the impact of potential regulations. The next section will focus on which ads lead to disagreement.
To interpret the results, we need to know the precise conditions in which the labeling happened. The ProPublica volunteers were given no instructions for what ads they should consider as political, and they were left to decide based on their subjective beliefs and background knowledge. However, volunteers could see if an ad was labeled as political by the advertiser itself (these ads have a "Paid for by" disclaimer on Facebook). We present results separately for ads that run with a disclaimer and ads that run without a disclaimer to isolate the potential effect of the "Paid for by" disclaimer.
Advertisers have to self-declare if they send political ads (as defined by Facebook's ToS). However, there is no public information on how Facebook enforces this policy [33]. Hence, ads labeled as political by advertisers are either a product of their own belief that their ad is political; or the result that Facebook constrained them to label the ad as political to run on the platform (maybe due to false positives in their enforcement algorithms).

Analysis of advertiser categories
To characterize advertisers we analyze their category. Advertisers need to create a Facebook Page and select from a pre-defined list a category for their page such as "Software Company" or "Political Party" [16]. We use the advertiser's ids available in the dataset to collect their category using the Facebook Graph API. Some pages no longer exist, we were able to extract categories for 6476 ProPublica advertisers (82%). Figure 2 plots the breakdown of the corresponding advertisers categories for strong political ads, political ads, marginally political ads, official political ads and nonpolitical ads. We group similar advertiser categories (grouping details can be found in our supplementary material at http://ligmembres.imag.fr/gogao/www21.html). Figure 2 shows that most strong political ads come from political actors (58% w. and 48% w/o. disc.), but a significant fraction of ads also come from NGOs (14% w. and 21% w/o. disc.), communities (4% w. and 4% w/o. disc.), and advocacy groups (3%w. and 4% w/o. disc.). In the political ads group, a smaller fraction of ads come from political actors (24% w. and 25% w/o. disc.), much more from NGOs (36%w. and 37% w/o. disc.), and we also see more ads from advocacy groups (6% w. and 6% w/o. disc.), news media (4% w. and 4% w/o. disc.), and communities (6% w. and 5% w/o. disc.). In the marginally political ads group, only (1% w. and 1% w/o. disc.) of ads come from political actors, the majority (52% w. and 58% w/o. disc.) from NGOs and charity organizations (11% w. and 5% w/o. disc.), some ads come from news media (5% w. and 3% w/o. disc.) and businesses (5% w. and 3% w/o. disc.). In the official political ads group, we see a similar diversity in the advertisers labeling their ads as political. Many countries' specific electoral legislations only regulate (and impose restrictions on) ads from political actors. However, we see that there is a wide range of advertisers pushing political ads online and that volunteers do label ads from these advertisers as political; hence, prompting for updating legislation.
Facebook is explicitly exempting news organizations from labeling their ads as political even if they are about political issues [14]; however, yet do seem to consider these ads as political. This raises the question of whether ads from news media should be treated as political ads. On one side, political journalism is different from political propaganda; on the other side, news media has been used as a tool to manipulate users, and many unauthentic news aggregators are emerging with the purpose of promoting a political agenda [4]. Table 3 presents examples of political ads from different categories of advertisers such as community, NGO, or business. For each ad, the table shows the fraction of political votes divided by all votes from volunteers and whether the ad was labeled as political by the advertiser itself. The table shows that there is a wide diversity of ads getting labeled as political. For instance, we can see an ad from the ice-cream company "Ben and Jerry" (a business) that is inciting  Figure 2: Breakdown of advertisers categories for different groups of ads for ads with and without "Paid for by" disclaimer. citizens to vote, and an ad from the "Democratic Attorneys General Association" (an NGO) that is asking people who should be the V.P. of Joe Biden. Such ads have a clear association with elections. In the table, we also see many ads, such as the ones from the "World Food Programme" and the "USA for UNHCR" (Charities), that address social issues but do not seem to have any evident association to elections or legislation. The critical point to recognize is that ads labeled as political can have a very different level of "politicalness", going from straight advocacy messages addressing abortion issues to ads merely asking for a donation to end world hunger.

Analysis of ad messages
To gather grounded information about the topics of ads labeled as political, we took a random sample of 300 ads with a "Paid for by" disclaimer and 300 ads without "Paid for by" disclaimer from each strong political ads, political ads, marginally political ads,  Figure 3: Breakdown of the political and social issues discussed in ads for the different groups of ads with and without disclaimer.
300 ads non-political ads, and 200 ads without disclaimer from AdAnalyst. While we picked both ads with and without a disclaimer, we did not show the disclaimer in our surveys. We set up a survey on Qualtrics [29] where for each ad, we ask respondents questions about the ad's message. We hired workers through Prolific [26], and we redirected them to fill out the survey on Qualtrics. Each worker had to label 20 random ads from the pool of 2300 ads, and each ad was labeled by three workers. We selected workers that are residing in the USA since all the ads used in the experiments targeted people who lived in or visited the USA. The median amount of time that workers spent on the survey was 12 minutes. Each survey had an instructions page, followed by 20 pages each containing one ad to label. For each ad, we asked the following questions: (1) "Is this ad made by, on behalf of, or about a political actor? (such as a candidate for public office, a political figure, a political party or advocates for the outcome of an election to public office)"; (2) "Is this ad about elections? (such as referendum or ballot initiative, including "go out and vote" or election campaigns)"; and (3) "Does this ad refer to a social issue? (such as civil and social rights, ...)". Workers were allowed to answer yes to all the questions. If workers selected that the ad is about a social issue, we asked them which social issue: "Which social issue is this ad talking about?" Workers had to choose from the following list: civil and social rights, crime, economy, education, environmental politics, guns, health, immigration, political values and governance, security and foreign policy. We considered these social issues because they appear in the Facebook definition of political ads [13]. Workers were allowed to select multiple social issues if needed.
If workers answered no for all three initial questions (the ad is not about a political figure, election, or social issue), they were asked to choose from a list "What topic describes best the ad?". We took the list of 23 topics from the Interactive Advertising Bureau (IAB) categories [11]. Note that we did not ask workers whether the ad is political or not; we just asked them questions about its message. Figure 3 shows the breakdown of the political or social issue discussed in an ad according to Prolific workers for different ad groups for ads with and without disclaimer. For each ad, we pick the ad topic chosen by the majority of workers or mark it as disagreement if no two workers chose the same ad topic or if two topics had an equal number of votes. We attributed all ads about both a political figure and a social issue or a political figure and election to the political figure group, and all ads about both an election and a social issue to the election group. For clarity, all ads for which the majority of workers chose a (non-political) IAB topic are marked as "None of the above" in Figure 3. Figure 3 shows that all groups of ads contain most of the ad topics we consider. We see higher fractions of about a political figure or political party and ads about an election in the strong political ads (78%+8% w. and 65%+7% w/o. disc.) and higher fractions of social issues ads in the political ads (38% w. and 61% w/o. disc.) and marginally political ads (75% w. and 62% w/o. disc.). In the official political ads group, there is also a high fraction (48%) of social issue ads. The non-political AdAnalyst ads are shown as control. Indeed less than 2% of these ads are labeled as being about a political figure, election or social issue. Social issue ads are only considered political by Facebook and TikTok, not by Google and Twitter. However, these results tell us that a large proportion of the ads volunteers and advertisers label as political are about social issues. Hence, it is crucial to consider social issue ads as political as well. Figure 3 shows that some ads (2% w. and 2% w/o. disc. of strong political ads and 1% w. and 11% w/o. disc. of political ads) were not labeled by workers as being about a social issue, a political figure or election. Since there is no expert ground truth, we cannot say whether labels from volunteers or labels from workers are better. Nevertheless, the (non-political) IAB topics that were mentioned the most by workers were society, health & fitness, education and science. This raises questions on where to drawl the line between ads about civil and social right and ads about society; or ads about health as a social issue and ads about health & fitness as a lifestyle.
One might decide that marginally political ads should not be treated as political because only a minority of volunteers labeled them as political. Figure 3 shows that 3%+2% w. disc. and 2%+3% w/o. disc. of marginally political ads do contain ads from a political figure or political party or elections. In addition, 21% w. disc and 20% w/o. disc. ads are about civil and social rights, and 24% w. disc. and 9% w/o. disc. are about environmental politics. The numbers look similar for non-political ads. Marginally political ads do contain a significant number of political ads as defined by the Facebook ToS. These results show that marginally political ads should not be ignored because they might contain ads about social issue and ads where only a few people have the right background knowledge to detect them as political.
Takeaway: Our results show that a large fraction of ads labeled as political are about social issues and do not mention a political actor or elections. Hence, it is crucial to consider ads about social issues as political. Our results also show that a wide range of ads are getting labeled as ads about social issues. Hence, since many legislative projects are considering to severely restrict micro-targeting [23] or ban such ads altogether; we need to decide whether we want ads (with no apparent link to elections and legislation) coming from charities or communities to be subject to the same restrictions as ads that advocate polarizing issues. Such restrictions could hurt a wide range of humanitarian civil organizations.

LEARNING FROM DISAGREEMENT
The previous section showed that a very diverse set of ads get labeled as political. This section analyzes the ads that lead to disagreement among volunteers and between volunteers and advertisers. This analysis is relevant for refining political ads' definition and improving the processes and instructions for labeling ads.

Volunteers vs. advertisers
To understand why advertisers and volunteers disagree on ads being political, we examine separately ads that seem to be underreported by advertisers and ads that seem to be underreported by volunteers.
Ads underreported by advertisers. These are the strong political ads and political ads without disclaimer. Table 2 shows that 4% of the strong political ads and 7% of the political ads are not labeled as political by advertisers. There are several possible (nonexhaustive) explanations: (1) advertisers do not comply with the ToS (e.g., they willingly do not label their ads as political to avoid scrutiny), i.e., volunteers are right; (2) advertisers underreport certain categories of political ads, i.e., advertisers and volunteers have different interpretations of which ads are political; and (3) volunteers misinterpret the ads' message, i.e., advertisers are right. Figure 2 presents the breakdown of advertiser categories and Figure 3 the breakdown of ad types corresponding to strong political ads, and political ads without disclaimer. A significant fraction of advertisers are political figures (48% in strong political ads and 25% in political ads), and a significant proportion of ads refer to a political figure or political party and elections (65%+7% for strong political ads and 17%+7% for political ads). Hence, more than half of strong political ads and political ads without disclaimers are not compliant with Facebook's ToS. These results confirm previous findings that advertisers sometime do not label their ads as political and the need for better enforcement mechanisms [33].
A large fraction of ads without a disclaimer (23% of strong political ads and 61% of political ads) are about social issues. Recall that we excluded from this category ads labeled as social issues but mentioning a political figure or elections. Tables 7 and 8 (in Appendix) show some examples of ads about civil and social rights and environmental politics in strong political ads and political ads without disclaimer. These ads are on topics such as climate change, healthcare, and equal pay, which are very politicized issues in the US, and give valid reasons to volunteers to label them as political.
To understand whether ads about some social issues are less disclosed by advertisers than others, for each ad topic, we compute the fraction of ads that do not have a disclaimer in the strong political ads and political ads groups. Ads about economy (0.15), civil and social rights (0.28), and security and foreign policy (0.27) have the lowest fraction of ads with a disclaimer. In contrast, ads about political figures (0.6), election (0.57), and environmental politics (0.49) have the highest fractions of ads with a disclaimer. This shows that advertisers are underreporting ads about social issues, especially if they are about economy or civil and social rights.
For 2% of strong political ads, and 11% of political ads w/o. disc. workers did not label them as being about a political figure, election, or social issue; which means that no one besides volunteers labeled them as political. Table 4 shows a few examples of such ads. These ads seem to address some issues but are not clearly related to the social issues provided to workers. This raises an interesting dilemma: if someone labels an ad as political (without being forced or by mistake), can they be wrong?
Ads underreported by volunteers. These are non-political ads and marginally political ads with disclaimer. There are 1.3k nonpolitical ads, and 5.6k marginally political ads (74%) labeled as political by advertisers. There are various reasons why advertisers would label their ads as political while all/most volunteers labeled them as non-political: (1) advertisers might be forced to label ads as political (even if they are not) because of false positives in the enforcement mechanisms implemented by the ad platform; (2) advertisers might think that disclaimers would bring more attention to their page; (3) advertisers understand better why their ads should be political, and volunteers underreport such ads; etc. Figure 3 shows that a significant fraction (14%) of non-political ads are labeled as not being related to a political figure, election or social issue by workers; meaning that no one besides advertisers are considering these ads as political. Table 4 shows a few examples of such ads. Indeed, the majority of these ads do not seem to be political. Since substantial restrictions are envisioned for political ads, it is essential to know what enforcement mechanisms are put in place by ad platforms to understand what is the impact of false positives in their algorithms. Non-political ads mislabeled as political is also problematic when building automated detection methods that use political ads labeled by advertisers to train models. Hence, it is important to look for poisoning attacks when building such models. Figure 3 shows that the majority of non-political ads and marginally political ads without disclaimer are related to civil and social rights (21% and 20%), health (20% and 16%) and environmental politics (22% and 9%), while only a few refer to political actors (2% and 1%) or elections (2% and 2%). Figure 2 shows that these ads come mostly from NGOs (53% and 58%), news media (4% and 3%), businesses (4% and 3%), and charities (16% and 5%), while only a few (1% and 1%) come from political actors. Hence, it seems that volunteers underreport many ads about a social issue, especially about civil and social rights and health, and ads from advertisers such NGOs, and charities. Table 7 and 8 (in Appendix) present examples of ads about civil and social rights and environmental politics in non-political ads and marginally political ads. We see that most of these ads are related to social issues, but volunteers might not consider them as political because there is no apparent association with elections or legislation.
Takeaway: Two main factors contribute to disagreement between advertisers and volunteers: (1) advertisers mislabel ads as political or non-political (maybe to avoid scrutiny; maybe because they are forced to label their ads as political by enforcement mechanisms put in place by ad platforms); and (2) both advertisers and volunteers underreport ads about social issues. Part of the problem may be that the definition of ads about social issues is too broad which leads to different interpretations among people. This raises the question of whether all ads related to social issues should be considered political, and if not, how should we filter social issue ads that are not political. For example, one possibility would be to consider as political only ads about social issues that could directly or indirectly impact elections or legislation or that address polarizing issues.

Volunteers vs. volunteers
To investigate which ads lead to disagreement among volunteers, we check if there is more disagreement on ads coming from specific advertisers and ads about particular political or social issues.
To see if ads from certain categories of advertisers lead to more disagreement, for each advertiser category, we group all corresponding ads (from strong political ads, political ads, and marginally political ads). Figure 4 shows the ECDF of fr for each group. Ads with a fr close to 0.5 have the highest level of disagreement (half of the volunteers label them as political and half as non-political). We split the analysis on ads with disclaimer and ads without a disclaimer since the disclaimer might have impacted how volunteers voted. We see in Figure 4 that the distributions shift to the right (more political votes) when ads have a "Paid for by" disclaimer. However, we cannot attribute this shift solely to the presence of disclaimers because ads with disclaimers might also have messages that are "more political".
The plot shows that at least 10% of ads in each advertiser category has = 1, which means that at least some volunteers are not bothered by the fact that the ad is coming from non-traditional political actors. Figure 4 shows that ads coming from political actors and public figures achieve the highest agreement, 85% have = 1. Besides, ads from communities and advocacy groups tend to be seen as more political, while ads from charities as less political. Ads from other advertisers such as NGOs, causes, news media, education, and businesses are somewhere in between, leading to the highest level of disagreement. To get definite proof if the advertiser category influences the decision (and not the message of the ad), we would need a conjoint analysis that tests the same ad message with different advertisers but our data does not permit such analysis. In any case, platforms and policymakers should clarify how much consideration should be given to the advertiser when labeling ads as political.
To see if ads from certain ad topics lead to more disagreement, for each ad type, we group all corresponding ads (from the 1800 ads labeled by Prolific workers in strong political ads, political ads, and marginally political ads). Figure 5 shows the ECDF of fr for each group (we only show groups for which we have more than 20 ads labeled). We can see that the highest agreement is among ads that mention political figures and elections, while, as expected, the highest disagreement is on various social issue ads. We performed a pairwise Kolmogorov-Smirnov statistical tests between the distributions. Ads about elections and political figures are statistically different than the rest; but most of the social issue ads are not statistically different between them.
To see why for a particular social issues, some ads have higher fr than others, Tables 7-8 (in Appendix) show examples of civil and social rights ads and environmental politics ads for different ad groups. We see that the ads address a wide range of topics (e.g., abortion, wildlife, violence, hunger), they call for various actions (e.g., sign petitions, surveys, donations, call an elected representative) and try to provoke various sentiments (e.g., pride, anger, fear). Ads that address climate change and pollution are seen as more political, while ads about wildlife protection are seen as less political. Besides, ads that refer to problems in the U.S. (ad from NRDC) are seen as more political than ads that refer to problems in other countries (ad from Care2). While these are only anecdotal examples, they emphasize the complexity of deciding which ads are political. Limitation: There are other reasons for disagreement that we could not analyze with this dataset. For example, the background knowledge of volunteers might impact how they vote (the political nuance of an ad is only recognized by some) or the political ideology of volunteers impacts how they vote. These questions are essential for recruiting moderators, and we leave them for future work.
Takeaway: Ads from NGOs, causes, news media, education, and businesses and ads on social issues lead to the highest disagreement among volunteers. To distinguish better political from non-political ads, we would need policy recommendations that clarify the perimeter of social issue ads. This raise a multitude of questions such as: Should we treat ads about more politicized issues differently than ads about less politicized issues? Should we treat social issues depending on the country? Should we treat ads that call for precise actions differently than ads that just inform citizens? Should we define social issues at a smaller granularity (in both topics and locality) than currently? How should the system adapt to emerging social issues? How much weight should be given to the advertiser's identity (as opposed to just the ad content)?

CLASSIFICATION AND DISAGREEMENT
Traditional supervised classification algorithms create models from positive and negative examples that we feed in the training phase. The previous sections showed significant discrepancies between ads labeled as political by advertisers and ads labeled as political by volunteers. Hence, this raises the question of whether classifiers trained on one or the other would result in significantly different models. Intuitively, if the training examples are biased, the models will be different, while if the training examples are representative of political ads in general, the resulting models will be similar. This section investigates how discrepancies in positive labels from advertisers and volunteers impact the resulting classification models.
For the evaluation we split the ProPublica dataset in two equalsize slices of 28k ads: 1 and 2 . We use 1 as the training and validation dataset and 2 as the holdout/test dataset. We build four models using four different sets of positive labels but the same negative labels. As negative examples, we took 7.5k ads in English from AdAnalyst without the "Paid for by" disclaimer (see Section 2). To build the different models, we used Naive Bayes. While Naive Bayes is neither new nor sophisticated, it was shown by Silva et al. [33] that it achieves very high accuracy for detecting political ads and outperforms other methods. The classifiers only take as input the ad's text, and as pre-processing, we deleted all Html tags, stop words, and punctuation. We used Count Vectorizer for text embedding [32].
We performed 10-fold cross-validation for each classifier over its specific training-validation dataset that contains 8000 positive and 7.5k negative examples. Table 5   ads), it is essential to limit the rate of false positives (non-political ads labeled as political); hence, we are interested in true positive rates for a 1% false positive rate. The table shows that all classifiers achieve high accuracy of over 95%, but only , , and achieve true positive rates of more than 90%. The lower true positive rate of 1 (86%) is expected as it has a more challenging task because it is trained and tested with more debatable political ads.
The main challenge in evaluating the classifiers is that we do not have a gold standard collection of political and non political ads. Table 5 only tells us how good the models are at identifying the same kind of political ads with the ones they were trained on, but not how good they are at identifying political ads in general. Hence, we look next at how these models perform on detecting other kinds of political ads then those they were trained on.
We use the four models to make predictions for all ads in 2 . To predict that an ad is political, we took the threshold corresponding to a 1% FRP for each classifier. Table 6 shows how well the four models are at identifying official political ads, strong political ads, political ads, marginally political ads, and non-political ads in 2 .
As negative examples, we used 1000 ads from AdAnalyst in English that do not have a disclaimer and were not used for training. Table 6 shows that has the lowest number of false positives, while has the largest number. For detecting strong political ads, all models detect more than 95% of ads. For detecting political ads the and perform the best (detecting over 94% of ads). For marginally political ads, 1 and , perform equally well (over 85% detection), while has a 82% detection. For non-political ads, labels 85.1% as political while labels 86.8% as political.
The detection rates of and are similar across different datasets, with performing better especially on marginally political ads and non-political ads. Hypothetically if the resulting classifiers would label as political the precise same ads, the input data is representative of political ads, and who is labeling the training data (be it advertisers or volunteers) does not matter. To understand whether and label the same ads as political, we computed the fraction of ads labeled by both models as political over all ads for different ad groups in 2 . The data shows that the two models have an overlap of 97% in strong political ads, 94% in political ads, 83% in marginally political ads, and 84% in non-political ads. These results show that the overlap is relatively high, but discrepancies in the input data do transfer to discrepancies in the output data. Hence, we need to consider how biases in labeling are impacting classification results and whether this may lead to unfairness against certain advertisers.  [31] considered the question of the autonomy of politics. The author concluded that the current situation of politics is reflected in three different ways: outright extinction, autonomy or weakening, which leads to different ways of perceiving, identifying, and defining politics. Warren [38] proposed that the concept of politics should help to clarify normative interests in politics, that the definition of politics should embrace everyday understandings of politics, and serve explanation. He suggested that politics can be define by two attributes: power and conflict.
Analysis of political conversations online: Hersh [20] described how political campaigns changed across time and concluded that social media has a large impact on peoples' decisions. Maruyama et al. [24] showed that Twitter activity could affect people's vote decision. The authors experimented during the 2012 U.S. Senate election in Hawaii. The results showed that people who actively participated in Twitter discussions changed their opinion about their candidate more often than people who did not use Twitter. Kou et al. [22] analyzed the development of public discourses on social media. The authors showed that during the "Umbrella movement", conversations on Facebook (mostly used by Hong Kong citizens) emphatized with protesters, while conversations on Weibo (mostly used by mainland China) emphatized with the government. Political advertising: Silva et al. [33] created a tool for collecting ads from Facebook and implemented several supervised classifiers to detect political ads during the 2018 election in Brazil. They detected a significant number of political ads that did not have the official "Paid for by" disclaimer. Ribeiro et al. [30] analysed ads send by the Russian interference in the 2016 US elections and found that ads were send to people less likely to report them. Edelson et al. [9] presented a clustering-based method to discover advertisers engaging in a potentially undeclared coordinated activity and proposed recommendations for improving the security of political advertising transparency. Furthermore, Bolden et al. [5] summarized problems with the Facebook Ad Library, such as the lack of clear policies and data systematicity. Finally, Ali et al. [1] proved that Facebook's ad delivery algorithms effectively differentiate the price of reaching a user based on their inferred political alignment with the advertised content, inhibiting political campaigns' ability to reach voters with diverse political views. Our paper focuses on a more foundational question of what should be considered political advertising. Several other studies have pinpointed problems with the Facebook ad ecosystem without focusing on political advertising such as discrimination [34], lack of transparency [2,3], and security and privacy problems [21,37] .

CONCLUSION
Many agree that online advertising especially political adverting needs to urgently be regulated, but one missing key is how to reliably detect political ads. This papers attempts to dissect some of the complexity of labeling political ads. To our knowledge, this is the first study to show how ordinary people label ads as political, why they disagree and what are the implications for policymaking and enforcement algorithms.
Our paper shows that volunteers seem to underreport ads from NGOs, and charities (that are considered political by advertisers) and advertisers seem to underreport social issue ads (that are considered as political by volunteers). While disagreement can be alleviated through better guidelines to a certain degree, many ads addressing societal and humanitarian issues are intrinsically hard to label. We believe that the community needs a gold standard collection for political ads and to better define the perimeter of social issue ads. We hope our analysis can help policymakers and ad platforms to refine the definitions of political ads and their regulation.

CREDO Mobile
"We can transcend the darkness of this moment by joining the struggles of past and future freedom fighters. That is how, when we reach the end of our lives and look back on these heady moments, we will find peace in the knowledge that we did our best." -Ady Barkan.

Granite State Progress Education Fund
Stop anti-abortion shame, stigma and hate from New Hampshire politicians. Sign the petition to support abortion access for all Granite Staters! 0.83 w.
Physicians for Human Rights Doctors and nurses are standing up against human rights abuses across the world. Join our community and learn more about our work 0.75 w/o. International Rescue Committee Women and girls in crisis zones face discrimination, violence, and a lack of equal opportunities. Learn how we're working to change that. 0.66 w/o.  Advertiser Text fr disc.

Strong Political National Parks Conservation Association
No organization has won more victories for the national parks over the past century than NPCA -but we can't do it without you. Please donate to protect our nation's magnificent public lands. Colorado's first-in-the-nation oil and gas regulations work to protect our communities. 1 w/o.

Care2
Botswana is considering lifting the ban on hunting elephants. We must act NOW and convince Botswana to maintain their stance on protecting these endangered elephants from poachers! 0.55 w.

American Bird Conservancy
The Endangered Species Act is under attack. Despite the fact that 99% of species shielded by the Act -including Bald Eagles and California Condorshave avoided extinction, opponents in Congress are threatening to undermine this bedrock environmental law. Add your name to ABC's petition and tell the government to help protect endangered birds now 0.83 w.

NowThis
Women are equally impacted by climate change, and it's critical that we have them equally involved in the solution 0.66 w/o. NRDC Plastics never break down. And that's becoming a real problem for those of us that depend on the Gulf of Mexico and Mississippi River 0.8 w/o.

Defenders of Wildlife
The support from our donors has helped us win many battles for wildlife, but there is always more to be done. Our love of animals is endless, so we are ready to fight tirelessly for imperiled wildlife that can't speak for themselves. Support Defenders today and help us continue the fight for wildlife! 0.33 w.
National Audubon Society Action Fund Climate change threatens the birds we love. Sign up and we'll alert you to actions you can take to protect birds and the places we all need. 0.17 w.

Non-political National Park Foundation
Working together, you can help us have a powerful impact on our spectacular national parks. Your support right now will go to work immediately to protect the places that matter most for future generations. 0 w.
The Nature Conservancy The challenges facing our natural world are growing every day. Please, make a tax-deductible gift to give nature and wildlife a future. 0 w.

Oceana
Sea lions are drowning in mile-long "walls of death" off the California coast. Let them die... or help us save them 0 w. IFAW IFAW protects animals and the places they call home. With your help, we can continue to make a difference. Let's get to work. 0 w.