Facebook isn't free: zero-price companies overcharge consumers with data

Abstract

The emergence of dominant companies that don't charge money to their primary consumers poses serious challenges to current antitrust law around the world. This paper suggests an approach to regulating these ‘zero-price’ companies that considers the data consumers give up to use them as the ‘price’ they pay. The ‘data as price’ model acts as a starting point to assess whether consumers are being ‘overcharged’ by Facebook in the status quo compared to how much data they would give up in a more competitive social media landscape. By surveying thousands of participants and assessing a litany of relevant behavioural considerations, this paper finds that customers are overpaying for Facebook, and that this may come at a serious welfare cost to millions of consumers. While further analysis is warranted, there is substantial cause for concern, and for critical re-evaluation of the standards generally used by antitrust regulators around the world to regulate companies such as Facebook.

Introduction

This paper assesses whether consumers give up too much data to ‘free’ technology platforms by asking the following questions:

• Phase I: What are the welfare effects of current data extraction?

• Phase II: What data extraction would persist in a competitive, de-shrouded social media market?

• Phase III: How much better off would consumers be amidst a de-shrouded, competitive market?

This paper surveys Americans about Facebook, but can apply globally to any ‘zero-price-to-consumers’ company. The data show substantial consumer harm from data extraction and suggest that de-shrouding and increased competition (as would be provided by sound privacy and antitrust regulatory measures) can help.

Srinivasan (2019) chronicles Facebook's increasing data extraction as competitors were beaten and acquired. Because of Facebook's ‘zero-price’ business model, ‘it does not make much sense to focus the competitive assessment on prices’ (Valletti, 2019). Situating ‘zero-price’ companies into a traditional antitrust framework requires two things: first, a careful assessment of what consumers give up, which behaviourally informed surveys are best positioned to contribute; and second, a strong finding that, under the right circumstances, competition could remedy the harm done. This paper offers both.

Definitions and concepts

Data extraction: Collecting data for third-party monetization. Data extraction is distinct from data collection, which occurs when companies collect individualized data for purposes besides third-party monetization (e.g., Netflix personalizing recommendations) (Rodriguez, 2020).

Supra-competitive extraction: More data extraction than would exist in a competitive market.

De-shrouded extraction: Transparent data extraction where consumers fully understand what data are collected and what happens post-collection.

Data as price: Measuring the amount consumers pay for goods using the data extracted from them. While prior work has discussed using data as price (Eben, 2018), none has quantified it with behaviourally informed consumer-preference surveys, nor operationalized it within antitrust law.

While data, unlike money, can be duplicated, it remains a useful proxy in an antitrust context. Other possible substitutes for money (e.g., a barter market where legal analysis is a pseudo-currency) are equally duplicable. An hour of legal analysis can be copied or widely distributed simultaneously via webinar. Data could also differ in that they are necessary for interacting with social networks (making it an input). While this is true for data collection (collecting your name/university), data extraction (third-party monetization) is severable from social media. Another possible difference is lack of scarcity. However, marginally increasing data extraction is still costly to individuals.

Others use data extraction as a quality metric (OECD, 2018). However, global antitrust law rarely employs quality theories (Sage & Hammer, 1999). Additionally, quality is reduced by features of goods themselves, rather than what consumers give up to access them. Lastly, people generally tolerate free low-quality products more than overpriced ones.

Other models use attention (not data) as a form of price (Wu, 2017). These models, however, suffer because consumers cannot use the product without providing attention.

Survey methodology

Qualtrics/Mechanical Turk

This paper follows Campbell and Kay (2014) and Sunstein (2018) in using Qualtrics/Mechanical Turk. To avoid ordering effects, individuals were each only asked one substantive question.

Participants

A total of 2326 US participants (slightly more males than females; median age = 30–49 years) were paid $0.17. Participants of previous studies were excluded. Questions received 100+ responses, enabling cross-question comparison.

Representativeness

Participants were recruited from the pool of US Mechanical Turk users. The sample appears fairly nationally representative of US Facebook habits. In total, 79% of respondents use Facebook daily (compared to 74% nationally), 14% weekly (compared to 17% nationally) and 7% monthly or less (compared to 10% nationally) (Pew Research Center, 2018; US Census Data).

Randomization

Randomized question order was unnecessary because each respondent received one substantive question. Answer choices were rotated.

Testing

Surveys were tested a dozen times, establishing baselines of attentive respondents.

Duration

Average duration was 70 seconds, consistent with pre-launch tests (median = 69 seconds) and with respondents only answering one substantive question and demographic questions.

Limitations

In-person experiments were infeasible for the single-substantive-question surveys, though they would have ensured participant focus. Similarly, introducing willingness-to-pay (WTP) and test questions would improve consistency and understanding. However, median values minimize individual outlier disruption. In-person replication can test robustness.

Phase I: current welfare effects of data extraction

Overview

This section measures welfare loss in the status quo compared to a hypothetical non-extractive ideal using three similar questions (to assess framing effects) (Kahneman et al., 1991). While WTP and willingness-to-accept (WTA) questions can be challenging, respondents were largely social media users, avoiding challenges of familiarizing respondents with the underlying material.

Willingness to pay

The WTP question summarizes Facebook's current privacy policies (using Mahmoodi et al., 2018) (Appendix A1). Then, individuals answered how much they would pay for perfectly protective data policies.

Worth to you

The next design avoided an unwillingness to pay for previously free products. Individuals imagined receiving more protective policies and stated what that change was worth to them (Appendix A2).

Willingness to accept

Respondents were told that their data were currently perfectly protected and asked their WTA more extractive policies (Appendix A3). This triggers an endowment effect by endowing strong protections (Knieser et al., 2014).

Results

Demographics and results are appended (Appendices A4 & A5). The medians are fairly low ($10–25) considering the lifelong benefits. This may be anchored by ‘software as a service’ offerings (Spotify, Netflix, etc.) charging ~$10 monthly. Replication can frame payments as monthly, soliciting higher lifetime values.

Endowment effects

These findings reveal a 1.75 endowment effect for privacy (WTA/WTP). This is small compared to other studies: ‘[A] meta-analysis … found an average [endowment effect] … of 7.2 [over a range of goods]’ (Knieser et al., 2014). ‘Superendowment effects’ have also been found for Facebook use generally (Sunstein, 2018). Replication can use discrete choice experiments in order to ‘avoid some of the distortions of [WTA and WTP]’ (Sunstein, 2018).

An endowment effect of 1.75 is also smaller than in other privacy experiments (Acquisti et al., 2014, finding a 5.7 endowment effect). However, privacy endowment ‘depends critically on the context … and how the problem is framed’ (Acquisti et al., 2014).

Endowment effects may be depressed because it is ineffective to endow privacy protections that people know they don't have. Distrusting the WTA question's premise (that Facebook is currently very privacy protective) reduces anger at losing privacy protections. While the analysis could be limited to people who believe Facebook is currently privacy protective, they may be preference outliers. An approach identifying people who accept the question's premise by analysing the 75th percentiles yields a (very large) 17.5 endowment effect.

Endowment effects may also be depressed because of difficulties in clawing back privacy that we have lost. Technically, disclosed data are (largely) permanent and protections are less meaningful if other providers already have similar information. Additionally, individuals might be unconcerned about Internet privacy if others are in the same boat. Historically, as well, privacy incursions are rarely reversed and ‘most people … view [privacy erosion] as inevitable’ (Lafrance, 2016).

Possible harm

Notably, the WTA's 75th percentile is $875. This staggeringly suggests that 25% of Facebook users may think that they are being injured nearly $1000 by Facebook's current policies. Importantly, these individuals are not minimal Facebook users. Four-fifths of individuals in the 75th percentile use Facebook daily. When casual ‘I Agree’ clicks generate over $1.5 billion in aggregate harm (Appendix A6), regulators should take notice.

This shows that individuals may be flagrantly ignorant of Facebook's data policies – a failure of law and policy. Alternatively, these individuals may feel that they gain more than $875 in value (Sunstein, 2018).

International effects likely mirror these US results. Facebook is widely used globally. This analysis should be replicated broadly to assess the harm done to consumers around the world across diverse regulatory and privacy expectation environments. Similarly, while respondents were over 18, children suffer equal harm from privacy infringements. Using data as price provides a unique mechanism to recover for data-related antitrust injuries against children and adults.

Considerations

Survey results can differ from real-world behaviour (Sunstein, 2018). Replications should observe market behaviour by, for example, examining discounts necessary to attract customers to extractive rewards programmes. Existing work shows consumers will pay ‘about $2.28 to conceal their browser history, $4.05 to conceal their contacts, $1.19 to conceal their location, $1.75 to conceal their phone's identification number, and $3.58 to conceal their texts’ (Savage & Waldman, 2013).

Worth to you and the ‘untethered effect’

Worth-to-you values, strikingly, exceeded WTA values (Appendix A5). Why?

This paper proposes an ‘untethered effect’, whereby valuations increase when they are untethered from real-life financial situations. WTP relates to actual ability to pay (Sunstein, 2007) and income (Breffle et al., 2015), and both WTP and WTA implicitly encourage people to consider personally paying/receiving money. By contrast, worth to you solicits abstract worth without regard to ability to pay or financial need. Notably, the untethered effect here is even larger than the endowment effect.

The ‘untethered effect’ is important to understand and quantify. For one, disentangling valuations from wealth may reduce anti-poor biases in regulation. However, untethered estimates may produce inaccurate estimates. For example, US juries may calculate inefficient awards because they do not personally imagine paying or receiving money. Experts and attorneys should be mindful of untethered estimates (though facing real litigants may adequately tether). Additionally, further analysis can explore the impact of combining the untethered and endowment effects (‘how much is it worth to you to maintain current privacy?’).

Phase II: data extraction in a competitive, de-shrouded social media market

Overview

Demand and supply simulations estimated consumer drop-off and increased revenue from increased extraction. These results approximated the likely level of data extraction in a competitive, de-shrouded market.

De-shrouding

Even with more competitors, markets with misleading or mystifying pricing do not produce efficient outcomes. Current data extraction is highly shrouded: ‘[It takes] 244 hours per year to read [most consumers’] privacy policies’ (OECD, 2018). Therefore, 97% of young adults ‘consent to legal terms and services conditions without reading them’ (Cakebread, 2017). Extraction ‘is ubiquitous and often invisible … [precluding] informed choices’ (FTC, 2012).

Here, data extraction was de-shrouded with brief, clear explanations of data policies based on Mahmoodi et al. (2018).

Demand simulation

A conjoint analysis simulated a competitive social media market (Appendices A7 & A8). Respondents chose between social media platforms with different data policies and a ~5% quality difference. Five percent fewer friends on the platform was a proxy for a small but significant quality difference that might persist in a competitive market, loosely analogized from US antitrust market definition practice (Salop et al., 2017).

Response rates were compared to a baseline question with a 5% quality difference but no data policy differences (in both, advertisers could target based on public posts). Analysing differences from the baseline isolates the effect of each data policy on consumer behaviour amidst a 5% quality difference.

Demand simulation results

Demographics and results (Appendices A9 & A10) show tremendous drop-off from extraction across the board. This suggests that people would flee extractive platforms if comparable options existed. However, several concerns exist. It is possible that the 5% quality difference was too small to represent actual expected quality differences under competition. Additionally, more analysis can explore the lack of (and occasionally peculiar) variation between different types of extraction.

What explains large drop-offs compared to real-life dominance of extractive platforms? For one, consumers may face quality reductions exceeding 5% to guarantee data privacy. Additionally, a focusing illusion may have inflated drop-off by adding artificial salience to parts of the online experience that are generally unimportant. People may have also rejected the premise of the questions (that the companies were otherwise identical).

Additionally, individuals may profess a commitment to privacy that they wouldn't act on: ‘91% of consumers think companies shouldn't be accessing their data without their consent’ (Fast Company, 2019). But ‘[many] consumers will not consciously spend time to select higher privacy settings for their services’ (OECD, 2018). De-shrouding, however, may obviate this ‘privacy paradox’ (Kokolakis, 2017). Kummer and Schulte (2016) studied an app market with simplified expressions of privacy preferences, and they ‘observe[d] fewer downloads for apps which ask for more privacy-sensitive permissions’.

At base, the results here should alarm regulators. The status quo dominance of extractive services may illustrate massive market failure considering that ~60% of consumers rejected even minimally extractive options amidst de-shrouding and viable alternatives.

Analysing demographic effects

Demographic questions are shown in Appendix A11. To analyse demographic effects on the demand simulation, fixed effects on the ‘question #’ variable helped control for the expected variation of answers between questions.

The regression was framed as follows:

• • Dependent variable: Answer to the substantive question (coded as binary)

• • Independent variables: Gender, race, politics, school and age

• • Fixed effects

The results are shown in Appendix A12.

Preliminary results indicate that only age was statistically significant at a 5% level. The positive coefficient shows older respondents choosing less extractive policies over more friends, an important but unsurprising finding. It is also noteworthy that none of the other coefficients were statistically significant (even after correcting for collinearity). These findings suggest that extraction is equally unfavourable across gender, racial, educational and political lines.

Supply simulation

To estimate extraction revenue, respondents answered questions as hypothetical business owners advertising on social media (Appendices A13 & A14). Because of the difficulty of answering these questions, the supply simulation is a first approximation that should be verified with (currently unavailable) data from actual ad buyers. Categories addressing more complex extraction necessitated modified framing (Appendix A15).

Demographics and results are appended, but they are primarily useful for integration with the demand simulation (Appendices A16 & A17).

Integrating results from the demand and supply simulations

To conceptualize integrating the simulations, imagine 100 Facebook users, each of whom generates $27.61 of advertising revenue (Flynn, 2018). Total quarterly revenue equals $2761. If increasing data extraction reduces demand by 20%, 80 users remain (in a simplified model). If that same policy increases revenue 50%, Facebook users each generate $41.42 in ad revenue. After considering user drop-off and revenue increases, $3313 of total revenue results. The ~$500 revenue increase (from $2,761) suggests that Facebook would rationally adopt the policy.

For this analysis, the baseline result was that 83% of consumers chose the platform with more friends when data policies were identical. Additional work is needed to understand this outcome, but for now, assume 17% of people preferred a platform with fewer friends (perhaps for anonymity or exclusivity).

The results shown in Appendix A18 indicate that nearly every increase in extraction reduces total revenue. The only revenue-positive extraction is ‘age, gender and city’ data (contingent on using the mean value). If the median is more accurate (which it likely is, in order to minimize outlier influence), even rudimentary data extraction may be unprofitable in a competitive, de-shrouded market.

This suggests that widely used data-extraction business models may be based on deception, oligopoly or both. It would also imply that the values obtained from Phase I of this analysis regarding consumer harm (to the tune of billions of dollars) may accurately represent consumer harm.

Replication and limitations

Replications should experiment with quality gaps exceeding 5% and use real ad-revenue data (if available). Replications should also explore the additive effects of different extractions. A company using certain extractive techniques (that caused some users to leave) may not lose additional users by implementing new extractive techniques (if sensitive consumers have already left). Further analysis should also ask whether extraction would cause consumers to switch from platforms they are already using (rather than assessing new platform choices). Myriad other challenges exist when modelling highly complex industries replete with network effects, stickiness, minimal data portability, etc.

One key objection is that companies may become more extractive when facing steeper competition (to preserve revenue). This makes sense in the heavily shrouded status quo. However, in a de-shrouded data extraction environment, as modelled in the demand simulation, consumers who receive understandable data policies and have comparable-quality alternatives steer clear of data extraction. The integration of the supply and demand simulations suggests that a de-shrouded, competitive environment could yield a dramatically less extractive equilibrium than the status quo.

Phase III: assessing consumer harm under de-shrouded competition

The Phase I survey was modified to study the likely harm under the extraction policies that Phase II suggests came closest to being viable in a competitive, de-shrouded market. Individuals were faced with one WTP (Appendix A19) or WTA (Appendix A20) question. Demographics and results are shown in Appendices A21 and A22.

To compare consumer harm under the status quo to consumer harm under de-shrouded competition, Phase III values were subtracted from Phase I values (Appendix A23). The results are striking and discordant with expectations. It appears that for half of the relevant measures, individuals found extracting ‘the age, city and gender you listed on your profile’ to be more harmful than extracting a wider range of personalized information (Appendix A24).

These peculiar outcomes likely occurred because age, gender and city feel especially concrete to consumers. It is easier for consumers to imagine their age, gender and city being extracted than picturing every post and message they ever wrote. Future research should explore the impact of the focusing effect and the ways in which policy wordiness and concreteness affect outcomes.

Conclusion

This paper finds that there would be dramatically less extraction (and likely less harm) in a world with de-shrouded data policies and more robust social media competition. Regulators should think concretely about analysing data as price in order to leverage antitrust law and behavioural insights towards regulating zero-price markets. Shrouding and insufficient competition in the market for data are concerning, and behaviourally informed antitrust interventions have the potential to improve consumer welfare dramatically.

These findings stand apart from the existing literature. First, this paper is the first to calculate the value that consumers put on Facebook privacy. Second, no prior work has modelled consumer choice amidst a competitive, de-shrouded social media market.

This paper constitutes a jumping-off point for regulators around the world regulating zero-price companies that do not fit neatly into existing antitrust frameworks. While data do not represent a perfect proxy for price, they provide an analytically useful tool that empowers regulators to leverage traditional antitrust tools in new competition environments.