Researchers using digital data in the field of humanities and social sciences mainly work with social media data originating from commercial platforms such as Facebook, Twitter or Instagram. These data are often collected in real time. In the following paragraph, we will discuss some French examples of such projects in order to highlight the different possibilities of scraping or gathering online data and to reflect on both the opportunities and challenges that come with the different approaches.

Twitter-data are certainly one of the favourite digital-born sources to be studied by digital humanists. French examples are the study carried out by Badouard (2016) on the #jenesuispasCharlie hashtag after the attacks against the editorial staff of Charlie Hebdo in January 2015, and Frédéric Clavert’s project on the commemorations of the First World War [Clavert 2018]. As noticed by Dagiral and Pailler, “one of the main constraints in building corpus from Twitter is the inability to obtain the entire stream, even for an extremely short period of time; there is not technical reason for this restriction, which actually results from the choices made when designing the site” [Dagiral and Pailler 2018, 117] (our translation). In order to have the entire stream, Twitter does offer the researchers to purchase data. However, this option strongly depends of available funding and raises some concerns about research ethics as well.

Next to the rather time-consuming tasks of building ones own dataset, researcher mostly rely on available data, which are shared online, often through platforms such as Github[5] or through special collections made available by archives, libraries, and other cultural heritage institutions. In 2010, the Library of Congress (LoC) reached an agreement with Twitter to archive all the tweets issued since 2006 [Raymond 2010]. It took the library until 2013 to complete the data collection for the period 2006-2010 – a fact demonstrating the enormous time and work required for indexing such a vast dataset. The volume of data (170 billion tweets and more than 300 terabytes of information from 2006 to 2010) led the LoC to announce in December 2017 that it would restrict collection to data concerning specific themes and events. For technical reasons but also ethical (consent) and legal (privacy rights) issues, data collected by the LoC are still not available to researchers [Zimmer 2015].

However, other institutions with smaller perimeters are offering access to Twitter collections. Examples include the archives of the INA (French National Institute for Audiovisual) and the BnF (French National Library), which are both related to the French Internet legal deposit. These institutions do not collect all the Twitter data, but have Twitter archives linked to their archiving perimeters. Although these archives are really valuable, they are biased on choices that do not necessarily coincide with the expectations of researchers. One respondent to our survey revealed the potential use of the French BnF archives, whilst underlining certain limitations: “I generally find it very difficult to work on traces, and particularly old ones. I use the BnF web archives, but the search tool isn’t very good. On top of that, anything that isn’t from ‘public’ pages isn’t archived” (ICS interview, May 2018). This may lead to research questions that are based on the availability of relevant archives (thus merely updating questions that researchers have already been aware of for some time). “The idea of the patrimonial value of web content” influences the criteria for safeguarding data and “threatens the existence of content that does not fit the criteria set by web archiving organisations for the preservation of data [...]. Certain contents and formats can also be excluded – this is particularly true of images in 1990s web archives, and also banner advertising and pop-ups” [Schafer 2018, 53].

This small overview of the ways researchers can access social networks sources (researcher's collect, use of sample made by others, purchase of data, institution’s collections) already underline several challenges in terms of sustainability and highlight several temporalities at stake:

• Researchers (and librarians) who collect data in real time in Twitter run the risk of lacking distance and potentially overlooking important trends. As explained by Thomas Drugeon, who works at INA in France and launched an emergency Twitter collection during the Charlie Hebdo attack in 2015:

  I added hashtags to the collection at lunchtime on the day after the events in the editorial offices at Charlie Hebdo. We acted more quickly for the events of 13 November. The Twitter flows weren’t organised in quite the same way either. The hashtags were more varied in November, whereas in January nearly everything was concentrated on the #jesuischarlie hashtag. In November, at least five hashtags emerged and we could observe movements, cycles too, for example day/night reflecting the different time zones at international level. For #jesuischarlie we missed the peak at the beginning. [...] Valérie Schafer: And did you also collect #jesuisahmed, #jenesuispascharlie? Thomas Drugeon: No, we didn’t capture them ourselves, but we could access them indirectly, as a side effect. We didn’t want to start adding hashtags along the way as we wanted to make sure that we would have a homogeneous set.[6]

• Researchers who react after an event will not be able to fully access data unless they buy them from Twitter or use datasets which were created by others (where available), may they be individuals or librarians. Researchers then face other challenges: although the datasets may be documented, stressing the importance of metadata [Pomerantz 2015], they are reliant on the choices of others – maybe even robots.
• As pointed out by Thomas Bottini and Virginie Julliard, who used INA archives to analyze the semiotic dimension of the “gender theory controversy” that played out on Twitter,[7] researchers need to “consider the often non-explicit archiving logic of these databases, as well as changes in their architecture and editorial policies” [Bottini and Julliard 2017, 47]. The massive availability of data in archiving collections is not synonymous with “turnkey” corpora [Barats 2016] [Brunet 2012] as collecting involves specific criteria and curation, which claim for a knowledge of archiving choices.

Coping with the sheer volume of data is a further challenge. The Info-RSN project, funded by the French National Research Agency (ANR), which focused on the circulation and sharing of information on Twitter and changes in journalism [Mercier and Pignard-Cheynel 2018], may help to illustrate this problem. The researchers used a private company to collect the data, i.e. all the tweets that contained a URL link to one of the 31 French news entities included in the survey. The raw data collected corresponded to roughly 40 million tweets, and it took one and a half years to develop tools and ultimately obtain a final corpus of 17 million tweets (after the elimination of duplicates and tweets that were outside the research criteria).

From a methodological point of view, building large datasets renews the traditional criteria mobilised for the constitution of research corpora (representativeness, exhaustiveness, homogeneity, etc.). The heterogeneous nature of data (alphanumeric, audio, visual or timestamped data), their ephemerality (including for the short term), the design-specific limitations of platforms (websites, socio-digital networks, internet portals), and finally the hardware-specific nature of devices and means of reception (computers, tablets, smartphones) make corpus-building a highly complex and critical hermeneutic practice. For researchers working with open access datasets, it can be difficult to detect the blind spots of a search query or to reconstruct the selection criteria that have informed the creation of the corpus. The archiving of Twitter by the BnF and INA can serve as a speaking example here again. INA compiles its archive through the public Twitter API, whereas the BnF uses Heritrix - the crawler also used by Internet Archive and many libraries around the world. The results are very different in terms of data as well as metadata: the BnF’s Twitter archives look more like screenshots, whereas the INA collections are more flexible, allowing comparisons of metadata while giving access to results that no longer look like the original tweet. This may be a disadvantage or not, depending on the type of research being conducted: studying a Facebook or Twitter news feed from a semiotic and discursive point of view requires being able to retain the format of the broadcast.

Despite their apparent accessibility digitised texts also raise challenges. Some of them are common with born-digital heritage. Some databases are unstable over time. This is the case of private press databases like Factiva,[8] which depend on agreements concluded between the publisher of the database and the owners of the data that is collected. Digitisation is not neutral either, and affects the digitised text [Föhr 2017] from a semiotic point of view:

The source text [...] is then turned into various types of semiotic text, including a first, the “digital” text per se, in which it is encoded using 0s and 1s. This text is generally illegible for humans. A second, known as an “image text,” can be displayed on a screen or printed on paper and read as such, but any analysis of this version is typically “manual.” Then come the dynamic text, the annotated text and the edited text, and finally the readable, analysable text. So [...] digitisation does not produce a single “digital” copy of the text, but an entire galaxy of digitised texts which, interlinked and organised hierarchically, open up new avenues for interpretation and analysis. [Meunier 2018]

Factiva, for example, collects and digitises the texts without preserving the layout and format of distribution. These choices should be accessible to researchers whether they are stable or not. Etienne Brunet compared the Frantext[9] textual database with the French works on Google books. Frantext provides the researcher with a defined and explicit database for collection criteria [Brunet 2012].

Dealing with datasets as “sources” for humanities and social sciences research not only requires a critical understanding of the production and indexical logic of databases and its structured information, but also a hermeneutic reflection on the dual nature of digital data as both content and form. “When working with digital objects it’s essential to remember that what they look like on the screen is a performance,” says Trevor Owens [Owens 2020, 6], head of the Digital Content Management department at the Library of Congress in Washington D.C. Next to the challenge of determining the authenticity of data in terms of their physical integrity, one has to consider the performative volatility of data representations in different formats and on different devices. New skills such as interface criticism which make us think about the framing of “in-form-ation” [Drucker 2014] are therefore necessary in order to critically evaluate the “integrity of representation” of digital data [Föhr 2017].

Finally, there is also the question of internal metadata and of metadata [Pomerantz 2015] added by the researcher during her/his annotations. Regarding the first ones, let’s mention the case of retweets. Although retweets are useful to study the popularity or even the virality of a tweet, the number of retweets that researchers may find in Web archives is linked to the exact time the tweet was archived, meaning that they may miss the real popularity of a given tweet if it was archived at the beginning of its dissemination and before the climax of its popularity. Regarding the second ones [Salvador 2017], the durability of taxonomies and their readability remain a central question: the choice of metadata and taxonomy is not always stable even within an organisation and this can make longitudinal or diachronic approaches more difficult, as the researcher has to consider not only changes in the Web, but also in its archiving, interfaces, supports, algorithms, descriptions and taxonomies. A rigorous documentation of metadata is therefore needed to facilitate searchability, analysis and quotations.

Another question we addressed in our surveys is: what happens to the data at the end of the project? In the field of biology, Timothy H. Vines and his team conducted a survey on data sustainability through the analysis of 516 articles published from 1991 to 2011. They noted that scientific data are lost two years after the publication of an article, after which the chances of accessing scientific data would fall by 17% per year. This was mainly due to the failure of storage tools or obsolete email addresses that prevented any contact with the authors [Vines et al. 2014, 95]. Is the situation different in social sciences and humanities? Probably not! The answers to our survey show a great difficulty when it comes to implementing a sustainable storage and sharing approach in a medium and the long-term perspective (i.e. periods of up to ten years or beyond ten years, respectively).

A first finding of ours is, that testimonials show a focus on the “here and now” rather than on future uses of research data. The mastering of data and tools is often driven by a learning-by-doing approach, shifting the focus to pragmatic solutions rather than investing time into more fundamental speculations about open science in a sustainable perspective. Many researchers highlight the large amount of short-term difficulties: “I don’t have much time to think about the long term, because I am too busy with my current collection and analysis problems! And above all, I don’t have the time or the technical means to go into it in detail” (ICS interview, May 2018). However, many of the scholars interviewed keep screenshots on their computers and external hard drives. They do their own archiving, but without sharing it or without any guarantee that they will be able to transfer their archives from one machine to another in order to preserve the data. This problem is sometimes “solved” by the use of online storage and sharing tools (OwnCloud, Dropbox, Google Drive, Huma-Num), “for practical reasons but without any real thought about the long term, especially after the project has ended” (ICS interview, July 2018).

But doing research is of course an iterative process that comes with a lot of surprises and unforeseen challenges which force the researcher to adapt methods and tools as the project progresses. It would be an error to underestimate the researcher’s adaptability and flexibility in this regard. As one historian confessed: “My current project (traces of the First World War on Twitter) was not originally designed as a research project. So initially, data archiving, storage and so on were not part of my plan. Today, they are increasingly central to my reasoning, as the project will end in 2019. For archiving, I can use the DH Now project. I also think that adapting to the GDPR is something to contemplate for the long term” (History interview, May 2018).

In terms of maintenance let’s return to the trio of data, corpus and results we underlined: they all can have similar stakes, but they involve different actors, research problems and management methods. Institutions such as the BnF preserve the data, without necessarily keeping the corpus created by the researchers. This complementary approach is currently being explored in the Internet Lab platform (made available to the researchers of the Web90 and ASAP projects in 2016), or in the prospective study conducted by Moiraghi (2018). Although data maintenance can be carried out by legal repositories and institutions, it can also be a private sector initiative, for example, for the Newsgroups of the Usenet community: these exchanges between early adopters of computer-mediated communications are a rich source of information for digital historians, and the maintenance of this data is entirely dependent upon Google’s interest in preserving these exchanges. Yet users were rapidly disappointed by the organisation of these archives, which were experienced as unclear and difficult to use and search [Poulsen 2009]. Camille Paloque-Bergès notes that “the non-standardised archives of the Web are the result of the succession of transitional and informal states of digitalised document cycles; this succession is due to the development of ‘wild’ documentation on the Internet before professional archivists started their work” [Paloque-Bergès 2017].

The problem of citability of research data and final research outputs is of course not a new one. As Emmanuelle Bermès (BnF) noted, this concern existed way before the era of digitalisation: from 2006 onwards, the BnF was concerned about the sustainability of identifiers, notably for websites. A choice had to be made between “opaque identifiers” based on existing standards such as UUID (Universal Unique Identifier), or more “significant” identifiers that facilitated the citability and accessibility of resources. “This duality illustrates the problem libraries have always had: the dilemma of choosing between communication and conservation” [Bermès 2006] (our translation). There are further issues to address: “Indeed, sustainability and unicity both have a problem of scale. It is now commonly accepted that identifiers can be reused for different websites or even destroyed on the web. Sustainability cannot therefore be considered as ensuring that data is ‘eternal’, but rather that it ‘lasts long enough’ for the needs of the institution managing the resource” (idem). Sustainability therefore involves other challenges such as the independence of the authority maintaining the data or the ability to integrate pre-existing models and envision post-existing models.

Maintenance requires the consideration of both the temporality of the research questions, methods, and tools and of the stakeholders and users involved. It requires time, effort and resources that are rarely available to researchers if not planned from the outset. It is rare for the researcher’s institution to carry out this maintenance. Researchers may therefore decide to rely on ad hoc solutions. However, the question of the responsibility for data preservation remains: within an institutional framework, research is not the sole property of researchers. They could rightly argue that there is a mutual obligation of maintenance within the institution they have developed their research. Yet such solutions have been more often provided for article repositories and platforms than for research data. In this respect a system like HAL[10] in France offers an effective response that is increasingly used by researchers for making articles available through open access (preprints, author drafts, final publications, etc.). Similarly, sound archive repositories, like the “Phonothèque” at the MMSH (the Maison méditerranéenne des sciences de l’homme, or Mediterranean Centre for Humanities),[11] offer a means of heritagising and preserving collections. But there are hardly any integrated solutions for projects using heterogeneous data. It is worth noting that various new solutions are under consideration at national level in France, for example the “feasibility study for a [shared] warehousing service for simple data.”[12] The way this study will deal with “Big Data, Little Data, No Data” [Borgman 2015] and with the precise scope of “simple data” remains to be seen, as in practice it is generally for complex data that researchers need help and support. But there are also question marks as to the need to create new solutions rather than building on existing ones proposed by research infrastructures in this area and to the utility of a nationwide scheme for an issue that is of international or global scope (not least because of cross-border research projects).

The interviews we conducted indicate that researchers use a wide range of data analysis tools (Nvivo, Iramuteq, Gephi, etc.), and also reveal tools developed by researchers. Thomas Bottini and Virginie Julliard developed their own tool to observe the semiotic dimension of the controversy around gender studies on Twitter. This raises the question of the long-term registration of these “custom-made” tools, their diffusion and their readability in the medium term [Bottini and Julliard 2017]. This is true for the case of the Calico project (communities of e-learning, instrumentation and collaboration), which ran from 2006 to 2009 and examined distance learning platforms and online discussion lists [Blondel et al. 2011]. A set of analytical tools were developed during this project which are no longer available – be it because of professional mobility of researchers or the lack of funding to maintain the tools. This lack of sustainability directly affects the readability and the intelligibility of the results of this research project.

As noticed by J. Smithies and al. “digital projects benefit from being planned and executed with their longevity in mind from the start” [Smithies et al. 2019, 8]. The consideration of collection environments and conditions as well as the documentation work may enrich reflections on the choices of the observable elements [Fuchs and Angel 2018] and the choice of methodologies [Paveau 2014] [Venturini et al. 2014]. A good illustration of this point is an ANR-funded research project on online petitioning. The research group behind this project (Pluridisciplinary Analysis of On Line Petitioning) had access to data from one of the main French-speaking online petitioning websites (lapetition.be): the dataset included 12,000 petitions which had received a total of 3.25 million signatures over the period between 31 October 2006 and 12 February 2015. The richness of the data (data entered by Internet users and generated by the system), as well as their heterogeneity (alphanumeric data, number of signatories, site sections and forms, etc.), led to the development of different approaches, both qualitative (interviews with the designer, “manual” analyses) and quantitative (textual statistical tools), in order to look at the characteristics of online petitioning as well as the limitations and incentives induced by the system that affect the petitioning dynamic. The interviews proved to be essential in order to assess the stability of the system over time and avoid bias in interpreting the collected data. Before the end of the project, i.e. 2018, the website was no longer accessible, confirming the importance of taking screenshots in order to be able to contextualise the data [Barats 2016]. This example underlines the crucial importance of documentation – and the emergence of data stewards as a new and promising profession in the field of digital humanities. As query tools evolve over time and can thus change data collection conditions, a professional documentation of the creation of research data should not only provide technical information (standards, protocols, formats), but must include search modalities used to constitute the corpora. The lack of “thick description” of both the metadata and the process of gathering, curating, and publishing or recontextualising research data is certainly one of the biggest risks when it comes to the long-term and sustainable contribution of digital humanities research to the long tradition of “analogue” knowledge production in humanities and social sciences [Tóth-Czifra 2020].

Until now, historians had been probably better protected from the obsolescence of their sources than their counterparts working with contemporary data, but they are now increasingly concerned by this risk. The ANR Web90[13] project carried out from 2014 to 2018 is a prime example: when it was launched in 2014, searches conducted in the Web archives (may it be the Wayback Machine of Internet Archive or collections of national institutions) only allowed URL requests. Subsequently, the BnF implemented a full-text search in its archives, thus widening the search capacities, affecting the results and influencing data analysis. In addition to changes in data access and the consequent disruption caused by full text and the possibility of classifying the results according to different facets (time, format, domain name), other additional tools provided an answer to a blatant problem: the temporally disparate composition of archived pages, which included anachronistic elements that had not been updated (for example, calendars that are not routinely re-archived for reasons of duplication, etc) [Brügger 2018]. Internet Archive recently introduced a system to date the archives of each element making up a webpage,[14] thus removing the opacity of these temporal patchworks. It is also important to document collection tools - they may sometimes exclude advertisings, gifs or emoticons in the case of automatic collection tools - or consultation terminals that modify the display of data, a central aspect in the case of works that emphasise a semiotic perspective. In the Web archives it should be noted that the fonts and characters may also differ from the original pages, as enlightened by Musiani et al.: “If at the time of archiving, a Web page font was not written explicitly in its original source code, but rather used by default, the default settings set by the browser in its current version will appear on the archived page” [Musiani et al. 2019]. Archiving data can also remove the original formatting of the platform. This is the case, for example, for Twitter data stored by INA as explained earlier: the form and the interface are not preserved, because the choices that have prevailed favour manageability and interoperability between the data. It is also important to note that the archiving or displaying of data can evolve over time: initially, INA archived emojis but did not display in the results to researchers, thus removing the heterogeneity and complexity of data that is characteristic of web writings [Halté 2016].

One respondent noted that the main difficulty “is to reconstruct the temporal dynamics of the discussions studied (often I only have access to the final data)” (ICS interview, May 2018). Elements that have not been collected have to be reconstructed to account for the complexity of the dispositif [15] and its functionalities. This was undertaken by the INA, who collected data collection limit messages via Twitter’s public API and developed documentation on missing data. Research on the history of the Web has also shown that it is difficult (but not impossible) to identify the users of 1990s Newsgroups and establish their sociological profile, or even to consider the representativeness of their posts. The same question could arise for Twitter’s archives and its users. Preservation is also about the readability of the article in the future. Therefore, one should also consider future users. A publication may describe facts that may no longer be readable/understandable fifty years later, as the digital environment studied has evolved or even disappeared. If we imagine, for example, the disappearance of Twitter, how could we guarantee the documentation of the data in such a way as to reproduce their context and their techno-semiotic environment? This is already the case when reading an article about the Minitel - a videotext system broadly used in the 80s and 90s in France, with students, especially strangers to the French context, who have absolutely no knowledge of the technical ecosystem and the small beige box. Some current analyses can be difficult to understand even now: hyperlink analysis and maps can lack readability, for example in the field of scientific controversies when the data are not visible. This difficulty could be further accentuated if Gephi, Iramuteq or Lexico, are no longer usable, or even documentable. If the tools as well as the data disappear, is the analysis still readable? The design of the tools and dispositifs, studied for example by Monnoyer-Smith (2016) or Mabi (2016), introduce restrictions and develop a specific type of research and results (graphs, clouds, link maps [Plantin 2013]), which also claim for digital hermeneutics.

In addition to the different temporalities of research data and the distributed materiality of the research devices and infrastructures, there are also temporal questions that are specific to the individual experiences of researchers. Amongst the researchers who were especially sensitive to data sustainability in our survey, one colleague – who has the ability to carry out “ad hoc programming and programming for the exploitation of XML databases” (DH and Language interview, May 2018) – has been particularly attentive to sustainability issues:

In general, we deal with the volatility of the data on a day-to-day basis: we have transferred our servers three times. Part of the corpus data is hosted on personal servers, outside the research institutions. However, we are fortunate to have redundant servers in the laboratory that we manage internally: all our data are saved in XML format (txt) that is always accessible and guarantees their use in the long-term, and in a compressed form that goes from one server to another several times a day. In each of our projects, a substantial part of out funding is earmarked for 1 / the legal protection of the data, ensuring the sharing and the diffusion of the data entered (without confiscation) and 2 / the saving of data without a subscription (even if the data is hosted by a provider to avoid giving priority to one university project over another. No “cloud” storage is used, but hard drives are purchased and CRON backup programming is used on the websites of both project partners, etc.) Each time, the solution is to reinvent, according to the type of financing and the project itself: an editorial project? A “pure data” project? etc. (DH and Language Interview, May 2018)

The capacity to think about data maintenance and the constant reworking of ideas seems to develop as soon as the researcher is no longer in “discovery mode” and fully immersed into the difficult daily control of tools and data (as underlined by one respondent to our survey: “[...] It has been more difficult since I started trying to integrate source code into the analysis for critical code studies. However, I am interested in discourse, the practices and the uses - accessing the source code and becoming familiar with it can help to understand the active role code plays in human-machine interactions” (ICS interview, July 2018)). For those starting a DH-project with rather limited technical or data management skills, the need for support by computer scientists and engineers might be a necessity right from the start of a project. As underlined by a researcher in ICS: “Accessing (and processing) data seems increasingly dependent either on researchers learning IT methods that are not part of their initial background and that may be costly to acquire, or on working in conjunction with computer scientists (technicians, engineers or researchers) whose methods and research interests may be significantly different. We may therefore witness a narrowing of research aims or questions because of the need to master certain tools” (ICS Interview, May 2018).

Open access is central to the concerns of researchers and of publishers, but is also of importance to European funders, especially in the framework of H2020 projects, thus continuing a movement that began with early initiatives such as ArXiv[20] in the field of physics or biology. Although this study does not intend to detail the policies of open access in Social and Human Sciences, it is important to note that these policies are a quintessentially economic and political challenges [Cotte 2017]. This is a regular subject of debate [Girard 2017] [Valluy 2017], which leaves researchers faced with a paradoxical demand: “first, to publish, then publish more, in highly rated journals, in journals that are increasingly evaluated, graded, ranked and second, to make results public, disseminate them to as many people as possible, get them online as quickly as possible” [Farchy et al. 2010] (our translation).

Open access has found its way through platforms like the French multidisciplinary open archive HAL[21] or the research notebooks platform Hypotheses.[22] These collaborative efforts have definitely fostered best practices and paved the way for a greater acceptance of open access publication and data sharing. Sure, one could argue that the maintenance and the searchability of these platforms aren’t free from biases and institutional dependences too. However, these biases and dependences appears less worrying than the ones specific for private companies with their opaque and questionable system of ranking, visibility and referencing. Beyond the independence of these academic databases, attention needs to be paid to questions of interoperability of data formats. But like the open access, interoperability comes at a price regarding policies of standardisation, negotiations and debates they may raise, technical developments and implementation, updates, gateways, etc., and the human effort it implies. The question of the economic model is clearly posed when funded researchers can publish articles in open access by paying publishers, while others are prevented from doing so because they are unable to pay the high fees.[23]

The question of open access is crucial at all stages of research: it entails strong constraints (as well as opportunities) for the researcher before, during and after the research process, and requires not only specific means (research engineers, data collection tools and short- and long-term storage) but also a detailed knowledge of the rules applying to data sharing (which were for example renewed within European legislation by the recent GDPR - General Data Protection Regulation).[24] These rules also evolve over time, when some works enter the public domain. Controversy may sometimes appear, as seen in the debates between [Ertzscheid 2015] and the Anne Frank Foundation. The researcher who currently uses web archives from archival institutions may for her/his part be frustrated by the legal impossibility to copy archives due to intellectual property rights. But he/she can find comfort in the knowledge that they will be traceable through permalinks.

As Zimmer (2010) pointed out in the case of Facebook, this sharing and quotability raise issues of anonymity. Anonymity can be waived by search engines and betrayed by the cited data – despite the best intentions of researchers [Latzko-Toth and Pastinelli 2013]. It is therefore necessary to anticipate any obstacles to anonymisation, which vary according to the type of data. This difficulty was encountered by the group of researchers working on the ANR-funded project on online petitioning [Barats 2016].[25] As mentioned earlier, thanks an agreement with the manager of an e-petitioning platform, this project had access to the entire database and related data (including texts, petitioners, signatories, comments and timestamps) over a period from 2006 to 2016. The data had been previously anonymised (12,000 petitions and more than 3 million signatories). However, the researchers quickly realised that entering the quotation of comments by signatories via the Google search engine provided the same information online. The research team had to carry out multiple online checks to be certain before publication that no quotes or excerpts compromised the anonymity of the data, thus restricting the dissemination of results. In addition, the agreement did not allow them to respond to requests for data sharing. This example is typical of the limits of open access to digital data and of the ethical, technical and legal constraints that influence the dissemination of results.

Unfortunately, many great digital humanities projects of the past now suffer from their status as “digital wasteland.” Many research project websites are just abandoned at the end of the project; others may be frozen. However, freezing a project is not sufficient to maintain it or guarantee its sustainability. A good example is the Amsterdam-project DDS (De Digital Stad): an online socializing platform created in Amsterdam in the mid-1990s. A team of media archeologists around Gerard Alberts revived the DDS platform using the data that had been preserved, but identified a number of technical issues despite the precautions previously taken:

On 15–16 January 1996, the DDS servers were down for most of the night to allow for a full backup of De Digitale Stad. A full 1-on-1 disk copy of all the servers running DDS was created on 3 Digital Linear Tapes (DLT). DDS congratulated itself with a city frozen in time, preserved “to be studied by archaeologists in a distant future: the FREEZE.” [...] In restoring old data, it soon came to light that the package would not simply unwrap, or defrost. The DLT tapes holding the FREEZE did not easily render their content.  [Alberts et al. 2017]

The planning of the DDS founders here was remarkable. However, being prepared is not always sufficient to avoid risk. One survey respondent described how the earmarking of funds to store data beyond the end of an international research project was still insufficient to keep the data available on the medium term: “The data was hosted by a university institution which could no longer cover the costs of data storage, leading to the loss of all the data. We had planned the financing of data maintenance, but the funding was limited to a ten-year period and could only be temporarily extended, leading to the loss of data that could have been used for future research” (ICS interview, June 2018).

As we know, many projects stop abruptly or are no longer updated. Although their short-term fate could be considered better than that of the many websites that simply disappear, the high risk of the data disappearing has resulted in multiple preservation efforts. For example, the European Publications Office keeps information about European projects within the CORDIS database,[26] and the websites of these projects can provide extremely useful data for funding bodies, stakeholders or historians interested in the history of science or of the European Union. This determination is also shown by the arquivo.pt website of the Portuguese web archives, which has made an effort to preserve European R & D websites.[27]

Some past projects give the feeling that research productions are out of date, but they reflect the state of the digital art of the period: a pioneering 3D production in 1980 rapidly appears to be well below current levels of achievements. Something that was a technical feat in its time becomes one of the first steps of a technique (which may be extremely relevant for analysis by a historian of science). This brings us to a wider examination of obsolescence over time for certain projects and the evaluation of the maintenance efforts necessary to avoid it: this raised the issue of digital databases or research outcomes based on paid images and royalties that must be negotiated on a regular basis.

It also highlights the broader notion of updatism [Pereira and Lopes de Araujo 2016], which seems inherent to the conflicting temporalities of digital studies. Yet the updating of websites and data are different things. The second case can be illustrated by the Twitter archives concerning the terrorist attacks that occurred in France over the last few years. These archives continue to grow, particularly at INA, as the hashtags are reused by users for tweets about different terrorist attacks or for commemorations, thus continuously producing new results. While living archives are more and more developing [Rhodes 2013], updatism is a new challenge that may affect the perimeters, the temporalities of projects and even reviews (possibly claiming for post publication reviews[28] as a way to reflect the evolving nature of data and knowledge).