Archives, Access and Artificial Intelligence

Front matter

LISE JAILLANT (ED.)

ARCHIVES, ACCESS AND

ARTIFICIAL INTELLIGENCE

Chapter 1: Artificial Intelligence and Discovering

the Digitized Photoarchive

X.Y. Han, Cornell University | Vardan Papyan, University of Toronto | Ellen Prokop, National Gallery of Art, Washington, DC | David L. Donoho, Stanford University | C. Richard Johnson,

Chapter 2: Web Archives and the Problem of Access: Prototyping a Researcher Dashboard for the UK Government Web Archive

Mark Bell, The National Archives, London | Tom Storrar, The National Archives, London |

Chapter 3: Design Thinking, UX and Born-digital Archives: Solving the Problem of Dark Archives Closed to Users

Chapter 4: Towards Critically Addressable Data for Digital Library User Studies

Chapter 5: Reviewing the Reviewers: Training Neural Networks to

Read Peer Review Reports

Martin Paul Eve, Birkbeck, University of London | Robert Gadie, University of the Arts,

London | Victoria Odeniyi, University of the Arts, London | Shahina Parvin, Brandon

Chapter 6: Supervised and Unsupervised: Approaches to Machine Learning for Textual Entities

Chapter 7: Inviting Al into the Archives: The Reception of Handwritten Recognition Technology into Historical Manuscript Transcription

AFTERWORD: Towards a new Discipline of Computational Archival Science (CAS)

Lise Jaillant, Loughborough University, UK

Digital archives are transforming the humanities and the sciences. Digitized collections of newspapers and books have pushed scholars to develop new, data-rich methods. Born-digital records (“items created and managed in digital form”) are now better preserved and managed thanks to the development of open-access and commercial software. Digital humanities have moved from the fringe to the center of academia. Yet, the path from the appraisal of records to their analysis is far from smooth.

Cultural heritage organizations face at least three main challenges. First, the volume of digital archives makes it extremely difficult for archivists to assess records. Applying Artificial Intelligence (AI) and machine learning (ML) to archives is still at an experimental stage, but AI/ ML could become an integral part of archival processes.” To manage the sheer bulk and potential sensitivity of records, archivists will also rely on creators to help them make appraisal and selection decisions at the point of deposit.

Second, most born-digital collections are currently closed due to a wide range of reasons (including technical issues, copyright, and data protection). Regardless of whether archives are digital or not, archivists need to balance individual rights and the public interest in the context of the General Data Protection Regulation (GDPR) in Europe. Nobody would reasonably claim that all born-digital data should be unlocked and openly accessible. Yet, it is important to recognize that “dark” archives contain vast amounts of data essential to scholars — including email correspondence, drafts of manuscripts, digital photos and videos. Within current legal frameworks, making born-digital archives more accessible is an urgent priority to fully make sense of our cultural heritage.

1 Ricky Erway, Defining “Born Digital,” OCLC Research, November 2010, URL: https://www.oc Ic.org/content/dam/research/activities/hiddencollections/borndigital.pdf [last accessed: Mar. 29, 2021].

2 Artificial Intelligence (Al) is a large concept designating the creation of intelligent machines that can simulate human thinking capability and behaviour. Machine Learning (ML) is an application or subset of Al that allows machines to learn from data without being programmed directly. In practice, the terms “Al” and “ML’ are often used interchangeably.

Third, data science and AI are becoming essential tools, but very few scholars (particularly in the humanities) have been trained to master these research methods, a skills gap which in turn has an impact on the training we offer to students. This is a central topic of a recent White Paper from the Alan Turing Institute on The Challenges and Prospects of the Intersection of Humanities and Data Science. According to Barbara McGillivray and her co-authors, “the challenge here is to find a way to train and upskill humanities researchers in quantitative and computational methods, while at the same time incorporating the basic principles from these methods throughout undergraduate and graduate degrees, so humanities graduates are well equipped to lead projects but also potentially undertake careers in research software engineering and data science for arts and humanities.” The authors suggest setting up basic courses in data science and software engineering to “offer the foundational skills to support humanists in having structured and informed conversations with computer scientists and data scientists needed in interdisciplinary projects.”3

Automation, Access and AI are becoming keywords to decipher our history. We do not suffer from a lack of records, but from too many records — often locked away in dark archives. Access to dark archives is central but needs to be complemented with data-rich methodologies. How can we shed light on born-digital and digitized archives? How can we give greater access to archives currently closed to the public? What is the role of automation and AI? Archives, Access and AI addresses these central questions and explores crossovers between various disciplines to improve the discoverability, accessibility and use of born-digital archives and other cultural assets.

1. Applying Al to Archives

Archivists have commented on the digital revolution and its impact on archives for the past three decades. But it was not until the mid-2000s that the scholarship on digital preservation started growing. In addition to the preservation of digital materials,* commentators have examined the impact of the digital revolution on

3 Barbara McGillivray et al., The Challenges and Prospects of the Intersection of Humanities and Data Science: A White Paper from The Alan Turing Institute, London 2020, see 21, doi:10.6084/M9.FIG SHARE.12732164.

4 From 2005 to 2007, the UK funder Jisc supported the PARADIGM (Personal Archives Accessible in Digital Media) project, undertaken by the Bodleian Library in Oxford and the John Rylands Library in Manchester. The overall aim was to examine the issues in preserving personal digital materials, and to produce best-practice guidelines. In 2007, the Arts and Humanities Research Council (AHRC) funded the two-year Digital Lives project, led by the British Library in partnership with University College London and the University of Bristol.

appraisal.” Focusing on the preservation of born-digital and digitized records, or on the selection of these records, is not enough. Access and the production of new knowledge are issues that need to move to the center of the scholarly debate. In particular, Artificial Intelligence can be used by archivists to identify sensitive records, but also by researchers to process large amounts of digital archival data. AI has the potential to transform archives, but it also brings new challenges (including ethical challenges).

The closure of libraries, archives and museums due to the COVID-19 pandemic has highlighted the urgent need to make archives and cultural heritage materials accessible in digital form. Yet too many born-digital and digitized collections remain closed to researchers and other users due to privacy concerns, copyright and other issues. Born-digital archives are rarely accessible to users. For example, the archival emails of the writer Will Self at the British Library are not listed on the Finding Aid describing the collection, and they are not available to users either onsite or offsite. At a time when emails have largely replaced letters, this severely limits the amount of content openly accessible in archival collections. Even when digital data is publicly available (as in the case of web archives), users often need to physically travel to repositories to consult web pages. In the case of digitized collections, copyright can also be a major obstacle to access. For instance, copyrightprotected texts are not available for download from HathiTrust, a not-for-profit collaborative of academic and research libraries preserving 17+ million digitized items (including around 61% not in the public domain).

The primary aim of the project was to develop ways to secure the personal archives of individuals in the digital era. In 2008, the Andrew Mellon foundation funded the futureArch project at the Bodleian Library to find solutions to the problem of born-digital but also hybrid archives (composed partly of paper materials). In particular, Bodleian Electronic Archives and Manuscripts (BEAM) worked on digital preservation infrastructure and researcher interfaces for hybrid archives. The 2010s saw the development of guidelines to preserve email archives. See Christopher J. Prom, Preserving Email - DPC Technology Watch Report, Digital Preservation Coalition 2011 (rev. ed. 2019).

5 See Ross Harvey/Dave Thompson, Automating the Appraisal of Digital Materials, in: Library Hi Tech 28 (2/2010), 313-322, doi:10.1108/07378831011047703; Kate Cumming/Anne Picot, Reinventing Appraisal, in: Archives and Manuscripts 42 (2/2014), 133-145, doi:10.1080/01576895.2 014.926824; Anne Gilliland, Archival Appraisal: Practicing on Shifting Sands, in: Caroline Brown (ed.), Archives and Recordkeeping: Theory into Practice, London, 2014; William Vinh- Doyle, Appraising Email (Using Digital Forensics): Techniques and Challenges, in: Archives and Manuscripts 45 (1/2017), 18-30, doi:10.1080/01576895.2016.1270838; Victoria Sloyan, Born- Digital Archives at the Wellcome Library: Appraisal and Sensitivity Review of Two Hard Drives, in: Archives and Records 37 (1/2016), 20-36, doi:10.1080/23257962.2016.1144504; André Vellino et al., Assisting the Appraisal of E-Mail Records with Automatic Classification, in: Records Management Journal 26 (3/2016), 293-313, doi:10.1108/RM)-02-2016-0006.

Archives, Access and AI is particularly timely. “Born-digital archives” are among the new research priorities highlighted in the UKRI (UK Research and Innovation) Infrastructure Roadmap Progress Report (2019): “Ihe complexity of ‘born-digital’ archives [...] and the challenges of archiving for discovery across many different formats raise significant questions about how to preserve, catalogue and make available these materials discoverable and accessible in a coherent fashion, in perpetuity.” The report adds that “this is a fertile area for the arts, humanities and social sciences to explore natural crossovers with other research domains.” A recent UKRI report on UK’s research and innovation infrastructure reiterates this priority on access to cultural collections using new technologies.”

Machine Learning applied to data in libraries and other cultural institutions is also at the center of current debates in the US, in Europe and elsewhere. Ryan Cordell recently wrote a Library of Congress report on ML,® which built on previous work in the same field — including Thomas Padilla’s report on Data Science, ML and AI in libraries for the OCLC (Online Computer Library Center). In particular, Padilla gives examples of applications of AI/ ML to enhance descriptions of records at scale: “semantic metadata can be generated from video materials using computer vision; text material description can be enhanced via genre determination or full-text summarization using machine learning; audio material description can be enhanced using speech-to-text transcription; and previously unseen links can be created between research data assets that hold the potential to support unanticipated research questions.”? Archives, Access and AI builds on this booming interest in new technologies applied to born-digital and digitized collections.

While many digital materials are completely “dark” and inaccessible to users, other records are open in theory, but difficult to find in practice. As Mark Bell, Tom Storrar and Jane Winters show in this edited collection, the UK Government Web Archive (UKGWA) is open to anyone with an internet connection, but discoverability is an issue for several reasons, including the inadequacy of keyword search to find relevant materials. The problem of searching huge amounts of records was also at the center of a 2019 article by Winters and Andrew Prescott. “With the rise of very large born-digital resources such as e-mail archives, Wikileak dumps and web

6 UKRI, Infrastructure Roadmap Progress Report, 2019, see 59.

7 UKRI, The UK’s Research and Innovation Infrastructure: Opportunities to Grow our Capability, 2020, see 3, https://www.ukri.org/wp-content/uploads/2020/10/UKRI-201020-UKinfrastructur e-opportunities-to-grow-our-capacity-FINAL.pdf [last accessed: Mar. 31, 2021].

8 Ryan Cordell, Machine Learning + Libraries, Washington D.C., 2020, https://labs.loc.gov/static/I abs/work/reports/Cordell-LOC-ML-report.pdf?loclr=blogsig [last accessed: Mar. 29, 2021].

9 Thomas Padilla, Responsible Operations: Data Science, Machine Learning, and Al in Libraries, Dublin, OH, 2019, see 12, doi:10.25333/xk7z-9g97.

archives, the limitations of Google-type searching are becoming more evident.”!°

For Winters and Prescott, it can be useful to re-examine the work of mid-twentieth century precursors of the web such as Vannevar Bush and Ted Nelson, who viewed the recording of links and information as the most effective way of processing very large quantities of information.

In addition to textual records, digital pictures can be very difficult to find without adequate metadata and cataloguing. The issue of discoverability is central for the Frick Collection in New York City, an art museum that houses the collection of industrialist Henry Clay Frick (1849 — 1919). Digital images and documentation for hundreds of thousands of art works have already been made freely available on the institution's online digital archive. An ambitious program of digitization is underway for the Frick Art Reference Library’s Photoarchive — a research collection with more than 1.2 million reproductions of works of art in the Western tradition. As Ellen Prokop and her co-authors highlight in this edited volume, the rapid pace of digitization has led to a backlog of images that have been digitized but lack metadata and cataloguing information. In turn, this lack of information makes these digitized pictures unfindable and unusable. To solve this problem, the Frick Collection has worked with computer scientists to apply AI to the Photoarchive, automatically annotating images in the collection with the headings used in the archive’s classification system. The team harnessed the power of Convolutional Neural Networks (CNNs), a deep learning technique used for classification and computer vision tasks — including image classification and face recognition.

Handwritten manuscripts also fall under the category of hidden collections, difficult to find, search and analyze at scale. While Optical Character Recognition (OCR) can decipher machine-generated texts to make them fully searchable, the technology does not work well for human-generated texts in digital form. Unlike typed characters, no handwritten characters are identical. Handwritten Text Recognition (HTR) is still at an early stage, but significant progress has been made in the past few years using Convolutional Neural Networks. In this edited volume, Tobias Hodel draws on the example of the European project Transkribus, a comprehensive platform for the digitization, Al-powered recognition, transcription and searching of historical documents. He shows that large amounts of data are used to train algorithms to recognize hand-written characters. As a consequence, “the resulting models are highly biased by the material they are trained on,” as Hodel points out.

In recent projects on handwritten text recognition with deep learning, researchers often used the IAM Handwriting Dataset to train their models. This dataset contains 115K+ English words by 600+ authors. Since deep learning model

10 Jane Winters/Andrew Prescott, Negotiating the Born-digital: a Problem of Search, in: Archives and Manuscripts 47 (3/2019), 391-403, doi:10.1080/01576895.2019.1640753.

need at least 10° - 10° training examples in order to perform well, the IAM Handwriting Dataset meets those requirements. Using large sets of data is a prerequisite for HTR, but it can also lead to a “cycle of bias” as Hodel puts it. Understanding the source materials and the methods used are essential to engage critically with HTR and more broadly, with any Al-powered techniques applied to archival collections.

The case studies featured in this book will be useful to our intended audience - archivists, digital humanists and social scientists, computer scientists and anyone else interested in the issues faced by archival collections in the twenty-first century. In their 2019 article “More Human than Human? Artificial Intelligence in the Archive,” Gregory Rolan and his colleagues note the “barriers to the uptake of Al technology for recordkeeping knowledge work.” One explanation is “a lack of compelling case studies”: “there are not many real-world examples within the academic or professional literature.”” Intertwining practical case studies and theoretical insights, the edited collection aims to fill this gap in scholarship. It presents advanced work in Archives and AI, while also “zooming out” and looking at the big picture.

To solve the problem of access to digital archives, cross-disciplinary collaborations are absolutely essential. The big challenges of our time — from global warming to social inequalities —- cannot be solved within a single discipline. The same applies to the challenge of “dark” archives. We cannot expect archivists or digital humanities to find a magical solution that will instantly make digital records more accessible. And it is not enough to encourage collaborations between disciplines that are very close (for example, history and literary studies). Instead, we need to take a radical step outside our comfort zone and set up collaborations across disciplines that seldom talk to each other. This is the main goal of the AURA network (Archives in the United Kingdom/ Republic of Ireland and AI), funded by the Arts and Humanities Research Council in the UK and the Irish Research Council in 2020-2021.”

Led by a management team with expertise in digital humanities, archives and computer science, AURA organized three workshops to bring people together and offer a forum for discussion and future collaborations: “Open Data versus Privacy” (Workshop 1); ‘AI and Archives: Current Challenges and Prospects of Born-digital archives” (Workshop 2); “AI and Archives: What comes next?” (Workshop 3). While

11 Gregory Rolan et al., More Human than Human? Artificial Intelligence in the Archive, in: Archives and Manuscripts 47 (2/2019), 179-203, see 186, doi:10.1080/01576895.2018.1502088. See also: Basma Makhlouf Shabou et al., Algorithmic Methods to Explore the Automation of the Appraisal of Structured and Unstructured Digital Data, in: Records Management Journal 30 (2/2020), 175-200, doi:10.1108/RMJ-09-2019-0049; Tim Hutchinson, Natural Language Processing and Machine Learning as Practical Toolsets for Archival Processing, in: Records Management Journal 30 (2/2020), 155-174, doi:10.1108/RM]-09-2019-0055.

12 www.aura-network.net[last accessed: Mar. 29, 2021].

AURA focuses mostly on the UK and Ireland, AEOLIAN (Artificial Intelligence for Cultural Organizations) strengthens connections between British and American partners. Funded by the AHRC in the UK and the National Endowment for the Humanities in the US, AEOLIAN runs from 2021 to 2023 and consists in a series of meetings and case studies that bring together a team of experts to develop new approaches to improving access to and use of digital archives.”

Other initiatives have created partnerships between experts in various fields, benefiting cultural institutions in general and archival collections in particular. The AHRC-funded Computational Archival Science (CAS) research network and the Advanced Information Collaboratory explore the conjunction of big data methods and technologies with archival practice. In Germany, the Leipzig Computational Humanities group includes experts in the humanities and computer science. In the USA, the HathiTrust Research Centre and Stanford Literary Lab also foster computational research in the humanities. The global spread of these initiatives show that cross-disciplinary collaborations are not enough: we also need to bring the best people together, independently of their nationalities and professional affiliations.

Published by a German press and written by an international team of contributors, the six chapters of this book feature examples of collaborations between researchers in computer science and engineering, archivists and scholars in the digital humanities. Often, these collaborations are made possible thanks to external funding and are hosted in large cultural institutions in world cities. Applying AI to archives would be difficult for a small archival collection in, say, Loughborough (a market town in the North of Britain). The combination of prestigious metropolitan institutions, ground-breaking technology, and advanced expertise can be intimidating. But this does not have to be the case. AI is still a very imperfect technology, with applications to the archival sector at an early, experimental stage.

Ina 2019 article entitled “Artificial Intelligence — the Revolution hasn't happened yet,” Michael Jordan notes that when the term “AT” was coined in the late 1950s, it referred to the ambition to build hardware and software possessing human-level intelligence. He uses the phrase “human-imitative AI” to refer to this aspiration to create an entity that would resemble humans. AI was meant to focus on “the high-level or cognitive capability of humans to reason and to think,” Jordan points out. “Sixty years later, however, high-level reasoning and thought remain elusive. The developments now being called AI arose mostly in the engineering fields associated with low-level pattern recognition and movement control, as well as in the field of statistics, the discipline focused on finding patterns in data and on making wellfounded predictions, tests of hypotheses, and decisions.”

13 +www.aeolian-network.net[last accessed: Mar. 29, 2021]. 14 Michael |. Jordan, Artificial Intelligence -The Revolution Hasn't Happened Yet, in: Harvard Data Science Review 1 (1/2019), 1-9. doi:10.1162/99608f92.fo6c6e61.

If we look at the specific case of archival collections, it is certainly true that AI has been used for low-level tasks: identifying sensitive information such as credit card numbers in emails, tagging pictures, transcribing handwriting for examples. These tasks could be done by any normal teenager with a minimum of training. The value of AI is not its ability to perform complex high-level tasks that require contextualization, theorization or creativity. Instead, the value of AI comes from its capacity to process huge amounts of data very rapidly - something that no human can do single-handedly.

Michael Jordan argues that success in human-imitative AI has been quite limited. We are very far from having artificially intelligent systems that can compete with humans at the higher levels of intelligence. A focus on human-imitative Al can distract us from a key challenge of our times: making sure that AI works for humans, that it makes our human lives better rather than worse. For Jordan, we are witnessing the creation of a new discipline: human-centric engineering. “Whereas civil engineering and chemical engineering built upon physics and chemistry, this new engineering discipline will build on ideas that the preceding century gave substance to, such as information, algorithm, data, uncertainty, computing, inference, and optimization.”5 Since the new discipline will focus on data from and about humans, it will need the perspectives of social scientists and humanists.

Applied to archival collections, human-centric engineering will focus on building artifacts and designing processes to make archives more accessible. The new discipline will bring together not only engineers, data scientists and computer scientists, but also archivists and scholars in the humanities and social sciences. Working collaboratively, these interdisciplinary teams will pay close attention to issues of privacy and biases. Unlocking archives should not come at any price, and the hype surrounding Open Data and AI must not distract us from the need to comply with data protection regulations and to address bias associated with black-box algorithms. In short, we need to mitigate the risks of malicious AI in our quest to unlock archival records.

2. The Threat of Dark Al

Using AI to make dark archives accessible is risky: sensitive information in government archives could inadvertently be released and fall into the hands of criminals; private information in email archives could be leaked, leading to distress for individuals and breaches of data protection laws; pictures in digital collections could be mis-labelled, leading to embarrassment and damage to the cultural institution's “brand.” Automatic image labelling has a long history of producing embarrassing

15 Jordan, Artificial Intelligence, 3.

results — in 2015, Google Photos labelled a picture of two black people as “gorillas.” Google Maps and Flickr have also suffered from race-related problems. Training datasets that contain biases result in problematic, sometimes appalling outcomes. In 2016, after engaging with users on Twitter, a Microsoft chat box began sharing racist, genocidal and misogynistic messages. And in 2020, the researcher Timnit Gebru said she was fired from Google after co-writing a paper on Al-generated language, which replicates unsavory biases found in online text. To unlock dark archives, we cannot rely on dark AI - defined as AI that is making human lives worse, not better.

Bias in large datasets is at the center of Thomas Padillas and Ryan Cordell’s recent reports on ML applied to libraries and archives. Cordell gives the example of the Chronicling America newspaper collection, a searchable database of US newspapers with descriptive information and selected digitized pages. Chronicling America is produced by the National Digital Newspaper Program (NDNP), a partnership between the National Endowment for the Humanities and the Library of Congress. The NDNP relies on institutions in each state to select and digitize approximately 100,000 newspaper pages representing that state’s history and geographic coverage. The website gives the impression that digitized collections reflect the diversity of each state:

Participants are expected to digitize primarily from microfilm holdings for reasons of efficiency and cost, encouraging selection of technically suitable film, bibliographic completeness, diversity and “orphaned” newspapers (newspapers that have ceased publication and lack active ownership) in order to decrease the likelihood of duplicative digitization by other organizations."

Yet, far from reflecting diversity, “the data skews to newspapers serving the majority,” Cordell argues.” For example, the collection privileges newspapers read by white middle-class audiences in the nineteenth century, rather than black and other minority-run papers. As Benjamin Fagan points out, Chronicling America does currently list forty-six black newspapers in its digital archive (c. 2.5% of a total of 1,799), but all were printed in 1865 or later.'® There are no black newspapers among the digital copies of 215 newspapers published before 1865. NDNP participants prioritized geographic spread, inadvertently deemphasizing racial representation.

16 https://chroniclingamerica.loc.gov/about/[last accessed: Mar. 29, 2021].

18 Benjamin Fagan, Chronicling White America, in: American Periodicals: A Journal of History & Criticism 26 (1/2016), 10-13, see 11, https://muse.jhu.edu/article/613375 [last accessed: Mar. 29, 2021].

“Consequently any ML projects based on Chronicling America will reflect those same oversights and exclusions,” writes Cordell.”

Bias in large newspaper datasets is also at the center of the Oceanic Exchanges project (2017-2019). Melodee Beals and her co-investigators argue that the national focus of digitized newspapers collections obscures the fact that international news exchange was central to the nineteenth-century press. ML projects based on national newspapers risk missing important international links with other papers. To mitigate these risks, Oceanic Exchanges produced the Atlas of Digitised Newspapers and Metadata, an open access guide to digitized newspapers around the world. Highlighting the history of digitized newspapers and digitization choices, the atlas examines metadata available in these collections. In particular, it “explores how machine-readable information about an issue, volume, page, and author is stored in the digital file alongside the raw content or text”? This project sheds light on information that is often hidden. It invites researchers and other users to see digitized newspaper datasets as human constructions, rather than unproblematic data.

Managing bias is essential for libraries and archival collections. For Thomas Padilla, eliminating bias entirely is not an option: in the hope of cleaning the dataset, elimination risks introducing more bias. Instead, Padilla proposes a bias management strategy to reflect on and integrate bias within the cultural organization. One of the recommendations is to:

Hold symposia focused on surfacing historic and contemporary approaches to managing bias with an explicit social and technical focus. The symposium should gather contributions from individuals working across library organizations and focus critical attention on the challenges libraries faced in managing bias while adopting technologies like computation, the internet, and currently with data science, machine learning, and Al. Findings and potential next steps should be published openly.”

Symposia and other communication strategies would promote self-conscious attention and criticism at every stage of an ML project. But, as Cordell argues, this may not be enough. “To create ML projects that reflect data justice, [...] libraries cannot pretend to be neutral or objective in relationship to race, class, gender, sexuality, or culture, but instead must consciously strive to forefront marginalised voices.”””

The notion of data justice is closely related to the black box problem, explain

Catherine D’Ignazio and Lauren Klein in Data Feminism. Machine learning algo-

20 https://www.digitisednewspapers.net/dhawards/[last accessed: Mar. 29, 2021]. 21 Padilla, Responsible Operations, 10.

rithms are so complex that they are often described as incomprehensible black boxes: you put data in, and you get something out, but what happens inside the box is a mystery. For D’Ignazio and Klein, data justice aims to “ensure that past inequities are not distilled into black-boxed algorithms.” While terms such as ethics “locate the source of the problem in individuals or technical systems,” justice acknowledges “structural power differentials” and works “toward dismantling them.””? Data justice seeks to address the structural inequalities in the training datasets that lead algorithms to produce less-favorable outcomes for women and ethnic minorities.

In Algorithms of Oppression, Safiya Umoja Noble argues that Google search algorithms privilege white people and discriminate against ethnic minorities, particularly women. These algorithms in turn reinforce existing prejudices against women of color with search results presenting black women as “angry” or “sassy.” Noble points out that the search operations are invisible: users of Google and other search engines have no access to the algorithms and deep machine learning systems, developed to index masses of information and move some to the first page of results. While the process of searching is hidden, the information users see on their screens becomes a reality and has an impact on decision making. For Noble, Artificial Intelligence will become a major human rights issue in the twenty-first century. “We are only beginning to understand the long-term consequences of these decisionmaking tools in both masking and deepening social inequality.”*4

Noble, Cordell and others have stressed the role that libraries and other cultural institutions can play to balance the power of tech giants. For Noble, “the public is increasingly reliant on search engines in lieu of libraries, librarians, teachers, researchers, and other knowledge keepers and resources.” Yet, it does not make sense “to outsource all of our knowledge needs to commercial search engines” that will return biased results.” Librarians and other knowledge keepers can mitigate bias and offer an alternative to commercial interests. Likewise, Cordell argues that “by centering ethics, transparency, diversity, privacy and inclusion, libraries can take a leadership role in one of the central cultural debates of the twenty-first century.””° Many entrepreneurs in the tech industry still live by Facebook’s early motto: “Move fast and break things.” But the same motto cannot apply to cultural institutions that value continuity over disruption.

23 Catherine D’lgnazio/Lauren F. Klein, Data Feminism, Cambridge, MA, 2020, https://data-femi nism.mitpress.mit.edu/ [last accessed: Mar. 29, 2021].

24 Safiya Umoja Noble, Algorithms of Oppression: How Search Engines Reinforce Racism, New York City, 2018, 1.

25 Noble, Algorithms of Oppression, 16.

As institutions of memory and community, libraries cannot be bound to destructive ideologies of technological implementation but must instead model alternative engagements with ML focused on building rather than breaking. Libraries can become ideal sites for cultivating responsible and responsive ML, as that term describes a constellation of technologies that explain data within the contexts of its collection, aggregation, and association.”

This debate over the place of libraries in the digital age is not new. In his 1994 article “Electronic Records, Paper Minds,” Terry Cook reminded information professionals that their role was to guide users from masses of information onto specific knowledge. At a time of big digital data, this role had become particularly challenging. If librarians and other knowledge keepers failed, they will “be replaced by software packages that can handle facts, and data, and information very efficiently, without any mediation by archivists or anyone else.””®