1

Generally, legal updates regarding copyright law are mostly seen as a technical matter, followed in detail by only a handful of copyright experts and industry figures. As an exception, the new Directive on Copyright in the Digital Single Market (DSM)^,  [1]  [2]has become one of the most controversial legislative pieces in the European Union’s (EU) history. The opposing petition was signed over 5 million times, [3] which gives a sense of the interest sparked by the DSM. In addition, dozens of demonstrations were held in numerous European cities on March 25, 2019 to protest against the DSM. Even the European Wikipedia site was shut down for an entire day in a protest against “dangerous copyright laws”. [4] Despite ferocious campaigning - led primarily by US tech giants Google/YouTube and Facebook, opposing artists, as well as internet freedom activists - the controversial DSM was approved by the European Parliament on March 26 2019 and endorsed by the Council of the EU on April 15, 2019. It was published in the Official Journal of the EU on May 17, 2019 and entered into force on June 7, 2019 (article 31 DSM). Since it is a directive (not a regulation), EU Member states have two years to implement the DSM into their national legislations, which is until June 7, 2021 (article 29[1] DSM). The DSM is the first major update to EU copyright rules in nearly two decades (2001).

2

The DSM’s most substantial and controversial articles are article 15 (formerly article 11) and article 17 (formerly article 13). While article 15 DSM focuses on online news content and grants a new right for press publishers over the use of their press publications by information society service providers (“link tax”), this paper focusses on article 17 DSM. This article will make online platforms and aggregator sites liable for copyright infringements if they do not take pro-active steps to enter into licensing agreements with rightholders and remove the non-licensed material on their platform. Although it is mainly aimed at Google/YouTube and Facebook, article 17 DSM will also impact other online content-sharing service providers (CSSPs), in particular smaller CSSPs, and has the potential to negatively affect them to a much greater extent than it would the big tech companies. The main reason for this is that article 17(4) DSM places great emphasis on the adoption of automatic filtering systems and introduces a financial risk for CSSPs if they do not successfully implement such filtering systems. In addition, article 17(9) DSM requires CSSPs to operate an effective complaint and redress mechanism for users in the event of disputes over the disabling of access to, or the removal of, works uploaded by CSSPs. Article 17 DSM therefore not only requires CSSPs to guarantee the unavailability of non-licensed content, it also requires CSSPs to guarantee user’s limitations and exceptions, such as parody.

3

This paper is structured as follows: Section 2 will explain the rationale of the DSM’s new liability regime, while Section 3 will briefly outline the relation between article 17 DSM and the E-Commerce Directive (ECD), [5] and allude to some inconsistencies between these directives. Section 4 will describe why having access to a strong upload-filter is essential for CSSPs to be able to remain competitive under the new era of the DSM. In Section 5, this paper will examine why small CSSPs will not have access to filtering technology - unless a provider is able to develop a sufficiently sophisticated and competitive filtering technology and willing to license it to small CSSPs against a reasonable fee - thus creating the risk that the DSM’s filtering requirement may substantially harm competition amongst CSSPs in the EU. In Section 6, this paper will examine why today’s content filtering technologies are subject to significant inherent limitations with regard to their accuracy, efficiency and affordability. Section 7 will describe what a CSSP will have to do to meet the requirement of having an effective and expeditious complaint and redress mechanism. Last but not least, in Section 8, this paper will describe why under the new DSM era larger CSSPs may gain a new competitive advantage over small CSSPs, which may lead to more market concentration, and whether small CSSPs can do anything to remain competitive.

4

The music industry has suffered great losses in the past decades from declining CD sales and, more recently, electronic downloads. The new digital business models such as subscription based models were not able to make up for this loss - whether through paid subscriptions to Spotify or Apple Music, Internet radio from Pandora, or videos on YouTube. [6] Big media companies and collective management organizations (CMOs) are at odds with digital music providers - especially free, ad-supported music services such as YouTube, for allegedly not returning significant revenue to the music industry. According to a Google report, from October 2017 to September 2018, YouTube paid more than USD 1.8 billion in ad revenue to the music industry. [7] This deal is, however, not lucrative enough for them. The International Federation of the Phonographic Industry (IFPI) claims that for every USD 20 Spotify returns to the music industry, YouTube returns only one dollar. [8] Thus, the worldwide music industry is fighting for more money and returns, in particular from tech giants like YouTube or Facebook.

5

In addition, the public perception and reputation of big tech companies suffered greatly from their insufficient actions to fight the spread of hate speech, violent videos and copyright infringements on their platforms. It suffered more still from recent scandals involving data breaches, the Cambridge Analytica scandal, as well as the role of tech companies and social media platforms in recent political elections and generally from their alarmingly increasing dominance in certain markets. Today, more than ever, tech companies are subject to the highest scrutiny of regulators and policymakers alike.

6

Moreover, safe harbor regimes are in turmoil, not only, but especially, in Europe. There seems to be a much wider global trend against safe harbor jurisdictions. Such a trend aims to impose proactive monitoring and filtering obligations on Internet service providers (ISPs). Some authors argue that the introduction of article 17 DSM is rooted in the discourse about the “Internet threat”, which reflects a gradual shift in the perception of ISPs from being “mere-conduits” to “active gate-keepers” of content uploaded and shared by users. [9]

7

On a political level, the question of “the fair remuneration of authors and performers and of the difference in bargaining power when they license or transfer their rights” was raised in a communication from the Commission to the European Parliament, the Council, the European Economic and Social Committee and the Committee of the Regions. [10] According to an additional communication from the Commission, online intermediaries have a duty to provide a safe online environment to users by ensuring that illegal content is removed promptly and proactively, and they should adopt effective proactive measures to detect and remove illegal content online and not only limit themselves to reacting to takedown notices which they receive.  [11]

8

This eventually led to the introduction of article 17 DSM, which is, in such a context, not surprising and may even seem overdue. The rationale of article 17 DSM is based on the assumption that online intermediaries that operate on an ad-funded business model (e.g. YouTube, Dailymotion, Vimeo) - as opposed to companies that operate on a subscription-based business model (e.g. Spotify, Apple Music) - do not obtain licenses from rightholders for the works which they store on their platforms. With the goal of closing the value gap, article 17(4) DSM forces CSSPs to apply best efforts to enter into licensing agreements with rightholders, CMOs and big media companies. This requirement should strengthen the negotiation power of rightholders, which should eventually lead to more favorable licensing terms for the rightholders and media companies.

9

It is important to note that it is disputed whether the value gap actually exists because there is no robust empirical evidence for its existence. The idea of a value gap was developed by the music and entertainment industry, which used this term in music industry global reports without having any empirical evidence. [12] Different trade associations representing the music industry created the term “value gap” sometime around 2015 as a slogan. Since then, they have concertedly and constantly used this term in numerous public and government relations campaigns. [13] The Draft Directive’s Impact Assessment confirms this assumption and states that “the limited availability of data in this area […] did not allow to elaborate a quantitative analysis of the impacts of the different policy options”. [14] The European Copyright Society also pointed out that the DSM’s proposal is not founded on any solid scientific - in particular economic - evidence. [15] Different reports come to the same conclusion and indicate that there is no clear evidence on the effects of copyright infringement in the digital environment, the scale of it, or the effectiveness of more aggressive enforcement strategies. [16] A report commissioned by the European Commission, which was released only upon the filing of an access request by the Pirate Party’s MEP Julia Reda, [17] states that there is no “robust statistical evidence of a displacement of sales by online copyright infringements”. [18]

10

There is no doubt that the policy goal to redistribute resources from big platforms to creators for the use of their works in the platform economy is well-intended. After all, the policy goal of our copyright system should be for creators to allow the European public to enjoy creative content in all ways made possible by digital technology against a fair compensation system. [19] However, it remains to be seen whether the implementation of the DSM in the EU Member States’ national legislation will have the desired effect, since it is uncertain whether the value gap actually exists. Even then, assuming that it exists, it is still not clear if article 17 DSM will have the desired effects. Hence, it is questionable whether the EU has chosen an appropriate means to achieve this policy goal by introducing article 17 DSM given its potential negative side effects. These effects include harm to smaller companies and European competition amongst CSSPs in general, as well as harm to freedom of expression in particular, due to the over-blocking of content, as will be highlighted in this paper.

11

The DSM’s new liability regime shall apply to CSSPs, which are, according to article 2(6) DSM, defined as providers “of an information society service of which the main or one of the main purposes is to store and give the public access to a large amount of copyright-protected works or other protected subject matter uploaded by its users, which it organizes and promotes for profit-making purposes”. This broad definition and, in particular, its term “large amount” create considerable uncertainty leaving it to the courts to define what a “large amount” means.  [20]

12

The second part of article 2(6) DSM states that “providers of services, such as not-for-profit online encyclopedias, not-for-profit educational and scientific repositories, open source software-developing and-sharing platforms, electronic communication service providers as defined in Directive (EU) 2018/1972, online marketplaces, business-to-business cloud services and cloud services that allow users to upload content for their own use” shall not fall under the DSM’s definition of CSSP. This excludes Wikipedia, open access repositories, and open source sharing platforms because they do not operate for profit. Internet access providers and telecom service providers are not covered by the definition because it is not their main purpose to give the public access to a large amount of copyright protected works. Also, online marketplaces such as eBay, whose main activity is online retail, are not covered for the same reason. [21]

13

In theory, the new regime for CSSPs should be coherent and complement the ECD, which introduced on a horizontal level a framework of conditional liability, enabling the development and functioning of online services in various forms. However, the emphasis on the adoption of upload-filters under article 17 DSM and the encouragement to invest in such technologies may well result in a clash with articles 14 and 15 ECD. [22] These articles provide safe harbor protection and prohibit Member States from imposing a general monitoring obligation on ISPs that fall under one of three categories, i.e. mere conduit, caching and hosting. [23]

14

Under the ECD, ISPs are qualified as a hosting service according to article 14(1) ECD as long as they meet the so-called neutrality test developed by the European Court of Justice (CJEU) and are granted safe harbor protection. [24] However, according to article 17(3) DSM, when a CSSP performs an act of communication to the public or an act of making available to the public – which will most likely always be the case because article 2(6) DSM’s definition of CSSP already entails these acts – it shall no longer be protected by article 14(1) ECD. Accordingly, the ECD’s safe harbor protection will no longer apply to CSSPs covered by article 2(6) DSM.

15

The DSM’s monitoring obligation through the use of filtering clashes with article 15 ECD and will blur the line between active and passive hosting providers. This distinction was determined in detail in a number of cases by the CJEU, who has slowly associated the requirement of knowledge and awareness of copyright infringing content with the active role of an intermediary, which goes beyond the mere provision of services neutrally and the technical automatic processing of data provided by the intermediary’s users. [25] The CJEU also specifically recognized that monitoring obligations to prevent copyright infringement would be in violation of the ECD. [26]

16

Therefore, article 15 ECD’s prohibition for Member States to impose general monitoring obligations on ISPs will be seriously compromised with the new obligations under article 17 DSM – and this, notwithstanding the fact that article 17(8) DSM states that “the application of the provisions in this article shall not lead to any general monitoring obligation”. Indeed, this statement only makes sense if you argue that the monitoring obligation for CSSPs does not apply to “any content”, but only to content for which the rightholders have provided the CSSP with the necessary information. However, as Frosio quite logically observes, the introduction of any filtering technology “de facto imposes a general monitoring obligation as in order to filter unwanted content, all content must be monitored”. [27] Thus, the statement set forth at article 17(8) DSM is only true if CSSPs can obtain a license for all copyright protected works from all rightholders. However, this is an unlikely scenario since CSSPs will always host some amount of unlicensed content and will therefore need to ensure that such non-licensed content is not available on their platform by monitoring all content. In addition, as Bridy describes, the most prevalent filtering technologies “work by screening every piece of user-uploaded content in real time against that universe of works. No file escapes the system’s surveillance. If such functionality does not amount to general monitoring, it is hard to imagine what would”. [28] Thus, article 17 DSM will bring a systematic inconsistency within EU law and effectively repeals article 15 DSM for storage providers covered by article 14 ECD.

23

As we have seen, article 17 DSM provides several strong incentives for companies to invest in the development of copyright filtering technologies. However, developing and maintaining an upload-filter requires significant resources. As mentioned, YouTube has already invested over USD 100 million in its proprietary “Content ID” filter using a filtering technology called fingerprinting. [38] According to a 2016 Google report, “Content ID” is responsible for 98% of the content management. With regard to the music industry, even 99.5% of reported sound recording copyright claims were automated through Content ID. [39] To give an idea of the scale of content: 500 hours of video material are uploaded on YouTube every minute (or 82.2 years every day). [40] YouTube describes its filter as a technology that blocks videos which match items identified by a small, trusted group of rightholders. [41] Since “Content ID” is a proprietary filter, we cannot analyze the tool in detail and must rely on (rare) publicly available information about it. At any rate, in view of the significant number of YouTube users, the massive amount of content uploaded to the platform, and YouTube’s efforts and investments in its filter, there are many reasons to believe that to date, “Content ID” may well be the most sophisticated upload-filter.

24

However, YouTube is not the only platform that has developed its own filtering technology. Audible Magic is a US based private company and currently the only third-party provider that offers content recognition solutions, which it licenses to universities and social media platforms. It has a growing list of social media clients, which includes, for instance, Facebook, Vimeo, Viacom, DailyMotion, and Tumblr. [42] YouTube initially licensed Audible Magic’s digital fingerprinting technology, but ultimately decided to build its own proprietary system. Confusingly, both companies now refer to their systems as “Content ID”,  [43] which led to an (on-going) trademark dispute over this name. [44]

25

Also, smaller platforms in the EU have made important investments in the development of such filters. For instance, Soundcloud spent over EUR 5 million to build its own filtering technology and dedicates seven full-time employees – out of approx. 300 employees [45] - to maintain the technology. [46]

26

The above shows that, contrary to claims made by proponents of article 17 DSM, copyright filtering tools are expensive, both to develop and maintain, and such cost may be too much too bear for startups and small companies. The European Commission’s policy rationale for proposing a mandatory filtering obligation (outlined in a leaked impact assessment) reflects a misunderstanding of the technological and economic realities around content filtering. In this impact assessment, the European Commission claims that the cost of filtering tools would be negligible for startups: “it is estimated that a small-scale online service provider can obtain such services for less than 900 euros a month” . [47] This estimate is almost solely based on comments submitted by Audible Magic to the US Copyright Office in a study about the effectiveness of the U.S. Digital Millennium Copyright Act’s safe harbor provisions. [48] However, as Engstrom and Feamster explain, this estimate is only accurate for an incredibly small number of CSSPs. [49]

27

According to Audible Magic’s website and its current pricing, a CSSP hosting less than 10,000 song files per month has to pay a monthly fee of USD 1,000 in order to license Audible Magic’s filtering technology for audio files only (video files have to be licensed separately at the same rate). [50] However, 10,000 song files is an incredibly low volume for a CSSP. To put this in perspective, when Soundcloud was only five years old, users were already uploading twelve hours of audio content every minute (or two years of material per day). [51] This amounts to 172,000 files per month, with an average song length of 3 minutes.

28

As such, developing or licensing a content filter comes at a relatively high cost. A survey of CSSPs reported that medium-sized platforms engaged in file-hosting services paid between USD 10,000 and USD 50,000 a month of licensing fees just to use Audible Magic’s filtering tool. [52] Another source states that, for mid-sized streaming companies, Audible Magic is quoting on average USD 30,000 to 60,000 per month of licensing fees. [53] However, it is worth noting that the licensing fees paid to the third-party provider amount to only a portion of the total costs associated with fingerprinting software. As noted by Urban et. al., licensed content filtering systems are not a turnkey service. In addition, filtering systems require integration with existing systems and additional operational work, such as tracking and managing user appeals. [54] The platform of any CSSP must be altered or augmented to perform the fingerprint lookups and comparisons against a fingerprint database. This is also a substantial software integration task. [55] Under article 17(4) DSM, this is what a CSSP will potentially need to do in order to meet the requirement to make best efforts to ensure the unavailability of specific works for which the rightholders have provided the CSSPs with the necessary information (i.e. fingerprinting and other information, which then needs to be compared against the CSSP’s data base). Thus, the cost for CSSPs to source filtering technologies from third parties and implement them in their internal systems is likely to be much greater in absolute terms than the European Commission’s initial projection.

29

The fact that article 17 DSM applies to “copyright-protected works or other protected subject matter uploaded by its users” is another challenge, because it is not restricted to certain specific copyrighted works such as audio or audio-visual works, for which most of the current filtering technologies are designed for, but for any kind of work. [56] This makes it even more difficult for a CSSP to efficiently filter the entire uploaded content and will create challenges for any CSSP applying article 17 DSM. Indeed, because rightholders can provide CSSPs with the relevant information about any kind of copyrighted work, including images or text files, a DSM-compliant filter will have to be capable of recognizing any kind of copyrighted content. However, fingerprinting tools are narrowly tailored to particular media types (for instance, an audio fingerprinting tool cannot be used to match copyrighted text files), and such tools only exist for a small subset of the many types of copyrighted content available online. [57] This will imply additional costs, since text, image, audio and video content must be separated and will likely each require a separate tool or technology (in case a CSSP’s platform allows one to upload these different kinds of works).

30

An additional aspect of the requirement set by article 17 DSM is that the cost of filtering systems also makes it harder for young companies to attract investors and compete with incumbents. This is confirmed by a survey in the US and the EU, which indicates that a majority of investors would be “uncomfortable investing in businesses that would be required by law to run a technological filter on user uploaded content”. [58]

31

The effects of economies of scale also weigh in favor of large tech companies when it comes to the filtering requirement. Once a technological product is built, the costs of offering an additional unit decrease with increasing scale. Economies of scale are a key advantage for large businesses, as such businesses can afford to invest in expensive and specialized capital machinery, whereas it may not be viable or cost-efficient for smaller businesses to either buy/license or invest in specific technologies, thus creating an entry barrier for such companies, where having the technology in question is a legal requirement. Large companies are also more likely to have a large workforce that can be assigned to separate tasks in order to boost productivity. [59] These factors all contribute to the fact that it will likely be big companies with sufficient financial and human resources and a large user-base that will be in a position to develop a sophisticated content filter, rather than smaller players with no such resources and a much smaller user-base. For instance, it seems doubtful that any other company than YouTube would reasonably be in a position to invest USD 100 million in a proprietary content filtering technology. And even if another company did choose to invest in the development of its own filter, the end product would presumably not be as powerful as YouTube’s “Content ID” and might not even be compliant with the requirements of article 17 (4) DSM (which requires CSSPs to make best efforts in accordance with the high industry standards of professional diligence and to take all the steps that would be taken by a diligent operator to achieve the result of preventing the availability of unauthorized works taking into account best industry practices and the effectiveness of the steps taken).

32

The German Federal Commissioner for Data Protection and Freedom of Information, Ulrich Kelber, also comes to the conclusion that smaller CSSPs will not be able to develop their own filtering technology and that they will have to license it from a third party. He raises an interesting point in connection with data collection. According to Kelber, the filtering obligation will ultimately create an oligopoly of a few providers of filtering technologies, through which more or less all the Internet traffic of the relevant platforms and services will run. Thus, these providers would receive and collect extensive data. [60] This would allow these companies to collect data about the users of their clients’ (i.e. other CSSPs) platforms in addition to the data already collected in connection with their own platform. At this point, however, it seems more likely that large tech companies will rather take advantage of their technological, structural and financial edge over smaller CSSPs and keep their more advanced filtering technology for themselves. This allows large CSSPs to gain in market power instead of supporting their competitors by licensing out their filtering technology to smaller CSSPs on a voluntary basis. However, it remains to be seen how powerful third-party providers like Audible Magic will become and how they will use the collected data from their clients.

33

In sum, it seems unlikely that smaller companies will be able to develop sufficiently strong content filters on their own, primarily due to their financial and structural disadvantage compared to larger companies. Hence, in the new DSM world, smaller CSSPs will have no choice but to license their filters from a third party – provided they can even afford it given the probable high licensing fees that may be imposed upon smaller CSSPs. Adding to the challenge is that there is no guarantee that a third-party provider such as Audible Magic will be able to develop a content filter that meets the requirements of the DSM and is offered against an affordable fee. Whether or not this is the case will mostly depend on whether the third-party provider can attract investors that are willing to inject funds into the improvement/development of its filtering technology. This, in turn, will depend on whether there is a sufficient pool of potential clients, which are CSSPs that are willing to invest a substantial amount of money and internal resources in order to license a third-party filtering technology.

34

As mentioned, upload-filters already play an important role for CSSPs today, and their role will only gain in importance under article 17 DSM. The role of upload-filters under article 17 DSM might be somewhat overstated and may stem from the false impression that filtering technologies are more developed than what they actually are. For instance, in a video for the European Commission, Audible Magic advertised the benefits and affordability of filtering technologies,  [61]giving the (false) impression that such filters were efficient, accurate and affordable. As part of its lobbying efforts, Audible Magic stated that its technology is accurate to about 99%. Even if this statement is true, an algorithm that misidentifies about one in every 100 pieces of audio content does raise a number of issues. To put the range of acceptable false positive rates into perspective, e-mail service providers, for instance, consider that any false positive rate higher than about 0.1 percent is too high to be used for spam filters, due to the potential limitations on speech that could arise as a result of legitimate e-mail messages being misidentified as spam. [62] Thus, an accuracy rate of 99% might sound high at first glance, but taking a closer look may well prove insufficient for a tech company filtering millions of files on a daily basis.

35

In their report that examines the current state of content filtering technology, Engstrom and Feamster find that all tools currently available to identify potentially infringing material – from hash-based filtering to fingerprinting – are limited in their capacity to accurately identify infringements.  [63]The authors observe that “critically, all content filtering technologies are at best capable of simply identifying the contents of a file, not making the often complex determination as to whether the use of a particular file constitutes an infringement”. [64] Filtering technologies have a role to play in the online ecosystem to identify and remove infringing material. However, due to the mentioned limitations, the range of infringing activity that filtering tools can effectively address is rather narrow. Even for media types for which filtering tools already exist, the tools are only capable of matching content; however, they are not capable of determining whether or not the use of a particular work constitutes an infringement. This determination generally requires the intervention of humans - courts and legal practitioners, since such identification does, depending on the applicable law, also entail being able to correctly identify copyright concepts such as fair use and fair dealing, as well as specific copyright exceptions such as parody or criticism. In the EU, the list of exceptions is already quite rigid and comprises twenty-one exceptions, nearly all optional, which describe exhaustively when a copyright protected work may be lawfully used without the rightholder’s approval. [65] The lack of EU-level harmonization in relation to copyright exception and limitations makes it even more challenging for filtering systems to be effective. This would require filtering systems to ascertain on a case-by-case basis the infringing nature of content in a geographically sound manner, namely taking into account the diverse existing national exception regimes. [66] For these reasons, it seems rather unlikely that in the near future an algorithm will be able to accurately identify copyright infringements, in the EU or elsewhere. This would, however, be necessary to avoid false-positives.

36

Even works belonging to the public domain present challenges for upload-filters as evidenced for instance by YouTube’s Content ID, which struggles to recognize the difference between copyrighted material and works belonging to the public domain. A German music professor tested Content ID and uploaded public domain recordings of copyright-free music pieces by Bartok, Schubert, Puccini, and Wagner, recorded before 1963 (and therefore in the public domain under German law). All these music pieces were blocked by Content ID and the professor had to appeal to numerous takedown requests.  [67]Another professor even uploaded a ten-hour video of white noise, only to have it flagged five times for copyright infringement. [68]

37

The above shows that filtering technologies are still in a rudimentary state and are far from being able to identify copyright infringements and are not even capable of accurately matching content. Therefore, there will be (too many) false positives, or, in other words, non-infringing content that will be (over-)blocked. How article 17 DSM addresses this issue will be analyzed in the following Section 7.