Faculty Bibliography

Released in 2025, Institutional Books: Harvard Library (IB-HL) is a collection of 983,004 volumes (242B o200k_base tokens), originally digitized through Harvard Library's participation in the Google Books Library project. As researchers and developers have begun to use IB-HL, a tension has emerged between standard large-scale preprocessing practices and the goals of careful information stewardship. Many existing pipelines optimize for web text: as a result, they tend to aggressively filter, deduplicate, restrict by language, and sometimes discard meaningful metadata. Meanwhile, researchers seeking to use IB-HL duplicate effort while performing similar processing and analysis. We describe an approach that we call Enriched Text. Instead of producing a single 'complete' stream of tokens, we normalize the text while preserving metadata through annotations. We separate endmatter, detect per-paragraph language, identify clusters of duplicate paragraphs, and compute per-paragraph bits-per-byte scores. We provide this information through HTML-like annotations layered on top of the text. By parsing these annotations, users can tailor the output to their own needs instead of accepting a global editorial decision on content. The pipeline applies to all ≈250 languages in the collection. This report describes this project's goals, implementation, and design rationale. The release includes IB-HL-ET (an enriched-text version of IB-HL containing 217B o200k_base tokens across 983,003 volumes, organized into 1.39B annotated subtopic paragraphs) and the pipeline that produced it. These serve to make the collection easier for machines to parse and for humans to study.

Historical book collections contain rich visual elements - such as illustrations, photographs, engravings, and decorative art - that are frequently under-explored in large-scale digitization projects. While Optical Character Recognition (OCR) has standardized the extraction of textual content, these visual components offer a layer of nuance and context that remains largely untapped by automated text extraction workflows. This technical report introduces Institutional Books - Visual Elements, an open-source end-to-end pipeline for detecting, classifying, deduplicating, and captioning visual elements from historical book collections. Alongside this pipeline, we release an initial dataset of 22.6 million visual elements extracted from the 983,004 scanned volumes that comprise the Institutional Books: Harvard Library dataset. This work contributes to ongoing, community-wide efforts to enable new use cases for digitized library collections through computational access, from artificial intelligence model training to digital humanities research.

Historical newspapers are an abundant record of public life, but their dense, irregular and sometimes noisy layouts make computational access to these materials both challenging and limited. We present the Institutional Newspapers Pipeline, a modular system we jointly designed with Boston Public Library to extract high-quality, structured datasets from historical newspaper scans. It was architected so that each step remains interpretable and customizable, and so that the pipeline as a whole remains computationally frugal enough to run on workstation-level hardware. The pipeline runs each scan through a multi-step process: it segments scans into individual type-agnostic crops and performs OCR on each resulting segment before then performing text analysis, type classification, reading order detection, named entities recognition, subject classification, language detection, and pre-computed embeddings generation on every crop. We ran this pipeline against a portion of Boston Public Library's holdings and released the results as an open dataset. The optical character recognition (OCR) output represents 16.3 billion o200k_base tokens across 83.1 million individual crops, extracted from 1,473,635 public domain newspaper scans published between 1795 and 1930. This report describes our methods for each processing step, the small models we trained, as well as the evaluation results and dataset-scale measurements we collected in the process. It accompanies the release of the pipeline, models, and dataset. We position this work as a substantial step towards unlocking high-quality data from tens of millions of newspaper scans.

We introduce a structured dataset and accompanying processing pipeline for nearly 200 years of decision making, derived from the Proceedings of the Board of Regents of the University of Michigan. This English-language collection holds official decision making records of one of the oldest public research universities in the United States, covering the period of 1837-2023. These records span 55,177 pages across 71 volumes, containing roughly 36 million o200k_base text tokens. We describe how our pipeline makes use of open-source reasoning Large Language Models (LLMs) to extract insights from these records, at meeting, volume and collection level. We separate each volume into individual Board meetings—1,755 in total—and then extract meeting-level metadata records (date, attendees, location, presiding officer) and governance events in several categories: degree programs, leadership transitions, organizational units, fundraising, buildings, and strategic plans. The event categories we extract were defined together with University of Michigan librarians, reflecting the kinds of requests for hard-to-reach information they have received about these records over the years. We also include a derivative reasoning dataset that pairs each meeting with a model-generated summary, chain-of-though traces, and decade-level synthesis. Finally, we include EPUB exports for each meeting to increase the accessibility of the information for human readers. These datasets are of particular interest for research on historical governance and decision-making and, as a narrow-domain corpus paired with reasoning traces, may also contribute to AI training efforts. None of this data processing would be possible without clearly established ground truths. To that end, we evaluate each stage of the pipeline against hand-annotated reference data.

Alignment of artificial intelligence (AI) encompasses the normative problem of specifying how AI systems should act and the technical problem of ensuring AI systems comply with those specifications. To date, AI alignment has generally overlooked an important source of knowledge and practice for grappling with these problems: law. In this paper, we aim to fill this gap by exploring how legal rules, principles, and methods can be leveraged to address problems of alignment and inform the design of AI systems that operate safely and ethically. This emerging field -- legal alignment -- focuses on three research directions: (1) designing AI systems to comply with the content of legal rules developed through legitimate institutions and processes, (2) adapting methods from legal interpretation to guide how AI systems reason and make decisions, and (3) harnessing legal concepts as a structural blueprint for confronting challenges of reliability, trust, and cooperation in AI systems. These research directions present new conceptual, empirical, and institutional questions, which include examining the specific set of laws that particular AI systems should follow, creating evaluations to assess their legal compliance in real-world settings, and developing governance frameworks to support the implementation of legal alignment in practice. Tackling these questions requires expertise across law, computer science, and other disciplines, offering these communities the opportunity to collaborate in designing AI for the better.

Publicly launched in 2004, the Google Books project has scanned tens of millions of items in partnership with libraries around the world. As part of this project, Google created the Google Return Interface (GRIN). Through this platform, libraries can access their scanned collections, the associated metadata, and the ongoing OCR and metadata improvements that become available as Google reprocesses these collections using new technologies. When downloading the Harvard Library Google Books collection from GRIN to develop the Institutional Books dataset, we encountered several challenges related to rate-limiting and atomized metadata within the GRIN platform. To overcome these challenges and help other libraries make more robust use of their Google Books collections, this technical report introduces the initial release of GRIN Transfer. This open-source and production-ready Python pipeline allows partner libraries to efficiently retrieve their Google Books collections from GRIN. This report also introduces an updated version of our Institutional Books 1.0 pipeline, initially used to analyze, augment, and assemble the Institutional Books 1.0 dataset. We have revised this pipeline for compatibility with the output format of GRIN Transfer. A library could pair these two tools to create an end-to-end processing pipeline for their Google Books collection to retrieve, structure, and enhance data available from GRIN. This report gives an overview of how GRIN Transfer was designed to optimize for reliability and usability in different environments, as well as guidance on configuration for various use cases.

The innovations emerging at the frontier of artificial intelligence (AI) are poised to create historic opportunities for humanity but also raise complex policy challenges. Continued progress in frontier AI carries the potential for profound advances in scientific discovery, economic productivity, and broader social well-being. As the epicenter of global AI innovation, California has a unique opportunity to continue supporting developments in frontier AI while addressing substantial risks that could have far reaching consequences for the state and beyond. This report leverages broad evidence, including empirical research, historical analysis, and modeling and simulations, to provide a framework for policymaking on the frontier of AI development. Building on this multidisciplinary approach, this report derives policy principles that can inform how California approaches the use, assessment, and governance of frontier AI: principles rooted in an ethos of trust but verify. This approach takes into account the importance of innovation while establishing appropriate strategies to reduce material risks.

Large language models (LLMs) use data to learn about the world in order to produce meaningful correlations and predictions. As such, the nature, scale, quality, and diversity of the datasets used to train these models, or to support their work at inference time, have a direct impact on their quality. The rapid development and adoption of LLMs of varying quality has brought into focus the scarcity of publicly available, high-quality training data and revealed an urgent need to ground the stewardship of these datasets in sustainable practices with clear provenance chains. To that end, this technical report introduces Institutional Books 1.0, a large collection of public domain books originally digitized through Harvard Library's participation in the Google Books project, beginning in 2006. Working with Harvard Library, we extracted, analyzed, and processed these volumes into an extensively-documented dataset of historic texts. This analysis covers the entirety of Harvard Library's collection scanned as part of that project, originally spanning 1,075,899 volumes written in over 250 different languages for a total of approximately 250 billion tokens. As part of this initial release, the OCR-extracted text (original and post-processed) as well as the metadata (bibliographic, source, and generated) of the 983,004 volumes, or 242B tokens, identified as being in the public domain have been made available. This report describes this project's goals and methods as well as the results of the analyses we performed, all in service of making this historical collection more accessible and easier for humans and machines alike to filter, read and use.

The next round of cyberattacks might come from a discussion about knitting.

In this chapter, law and technology scholar Jonathan Zittrain warns of the danger of relying on answers for which we have no explanations. There are benefits to utilising solutions discovered through trial and error rather than rigorous proof: though aspirin was discovered in the late 19th century, it was not until the late 20th century that scientists were able to explain how it worked. But doing so accrues ‘intellectual debt’. This intellectual debt is compounding quickly in the realm of AI, especially in the subfield of machine learning. Whereas we know that ML models can create efficient, effective answers, we don’t always know why the models come to the conclusions they do. This makes it difficult to detect when they are malfunctioning, being manipulated, or producing unreliable results. When several systems interact, the ledger moves further to the red. Society’s movement from basic science towards applied technology that bypasses rigorous investigative research inches us closer to a world in which we are reliant on an oracle AI, one in which we trust regardless of our ability to audit its trustworthiness. Zittrain concludes that we must create an intellectual debt ‘balance sheet’ by allowing academics to scrutinise the systems.

Archivists regularly contend with a wide range of security threats, including data breaches, inadvertent loss, and legal action by those hoping to make sealed records public. These threats are particularly salient when sensitive materials are donated with delayed-release conditions. Trust in archivists’ ability to enforce such conditions gives donors the confidence to enter into the historical record materials that they might otherwise destroy. But as these materials are increasingly born-digital (and therefore hackable, convenient to exfiltrate en masse, and more easily corrupted), and as governments and private parties become ever more aggressive in their efforts to secure early releases, we must innovate in order to stand still. To compensate for these new dynamics, we propose Strong Dark Archives (SDA), a blended legal and technical protocol for securing delayed-released archival materials among a network of libraries. SDA leverages modern cryptography and institutional agreements to coordinate access-control across multiple accredited archival organizations, providing broad resilience to data breaches, technical failures, and legal process. Through this distributed approach to security, SDA imposes meaningful friction on efforts to force the early disclosure of archival records.

"What’s a tort? It’s a wrong that a court is prepared to recognize, usually in the form of ordering the transfer of money ("damages") from the wrongdoer to the wronged. The court is usually alerted to wrong by the filing of a lawsuit: anyone can walk through the courthouse doors and, subject to the limits explored in civil procedure, call someone else (or, if a company, some-thing) to account. We’ll discuss the sources that courts turn to in order to answer such questions. Rarely, in tort cases, are those sources the ones laypeople expect: statutes passed by legislatures. Without statutes to guide them, what are courts left with?"– Provided by publisher.

Jonathan Zittrain’s testimony before Hearing of the Subcommittee on Competition Policy, Antitrust, and Consumer Rights on the Internet of Things.

The Internet, and the Web built on top of it, were intended to support an “entropic” physical and logical network map (Zittrain, 2013). That is, they have been designed to allow servers to be spread anywhere in the world in an ad hoc and evolving fashion, rather than a centralized one. Arbitrary distance among, and number of, servers causes no particular architectural problems and indeed ensures that problems experienced by one data source remain unlinked to others. A Web page can be assembled from any number of upstream sources, through the use of various URLs, each pointing to a different location. To a user, the page looks unified. Over time, however, there are signs that the hosting and finding of Internet services has become more centralized. We explore and document one possible dimension of this centralization. We analyze the extent to which the Internet’s global domain name resolution (DNS) system has preserved its distributed resilience given the rise of cloud-based hosting and infrastructure. We offer evidence of the dramatic concentration of the DNS hosting market in the hands of a small number of cloud service providers over a period spanning from 2011-2018. In addition, we examine changes in domains’ tendency to “diversify” their pool of nameservers – how frequently domains employ DNS management services from multiple providers rather than just one provider. Throughout the paper, we use the catastrophic October 2016 attack on Dyn, a major DNS hosting provider, to illustrate the cybersecurity consequences of our analysis.

Preserving records of what user content is taken down—and why—could make platforms more accountable and transparent.

Hyperlinks are a powerful tool for journalists and their readers. Diving deep into the context of an article is just a click away. But hyperlinks are a double-edged sword; for all of the internet’s boundlessness, what’s found on the web can also be modified, moved, or entirely disappeared. This often-irreversible decay of web content is commonly known as linkrot. It comes with a similar problem of content drift, or the often-unannounced changes––retractions, additions, replacement––to the content at a particular URL. Our team of researchers at Harvard Law School has undertaken a project to gain insight into the extent and characteristics of journalistic linkrot and content drift. We examined hyperlinks in New York Times articles starting with the launch of the Times website in 1996 up through mid-2019, developed on the basis of a dataset provided to us by the Times. We focus on the Times not because it is an influential publication whose archives are often used to help form a historical record. Rather, the substantial linkrot and content drift we find here across the New York Times corpus accurately reflects the inherent difficulties of long-term linking to pieces of a volatile web. Results show a near linear increase of linkrot over time, with interesting patterns emerging within certain sections of the paper or across top level domains. Over half of articles containing at least one URL also contained a dead link. Additionally, of the ostensibly “healthy” links existing in articles, a hand review revealed additional erosion to citations via content drift.

People across America and the world remain under strong advisories or outright orders to shelter in place, and economies largely shut down, as part of an ongoing effort to flatten the curve of the most virulent pandemic since 1918. The economic effects have been predictably staggering, with no clear end in sight. Until a vaccine or other transformative medical intervention is developed, the broad consensus of experts is that the only way out of mass sheltering in place, if hospital occupancy curves are to remain flattened, entails waiting for most of the current cases to resolve, and then cautiously and incrementally reopening. That would mean a sequence of allowing people out; promptly testing anyone showing symptoms — and even some who are not; identifying recent proximate contacts of those who test positive; and then getting in touch with those contacts and, if circumstances dictate, asking or demanding that they individually shelter until the disease either manifests or not. The idea is to promptly prune branches of further disease transmission in order to keep its reproductive factor non-exponential.

The governance of online platforms has unfolded across three eras – the era of Rights (which stretched from the early 1990s to about 2010), the era of Public Health (from 2010 through the present), and the era of Process (of which we are now seeing the first stirrings). Rights-era conversations and initiatives amongst regulators and the public at large centered dominantly on protecting nascent spaces for online discourse against external coercion. The values and doctrine developed in the Rights era have been vigorously contested in the Public Health era, during which regulators and advocates have focused (with minimal success) on establishing accountability for concrete harms arising from online content, even where addressing those harms would mean limiting speech. In the era of Process, platforms, regulators, and users must transcend this stalemate between competing values frameworks, not necessarily by uprooting Rights-era cornerstones like CDA 230, but rather by working towards platform governance processes capable of building broad consensus around how policy decisions are made and implemented. Some promising steps in this direction could include delegating certain key policymaking decisions to entities outside of the platforms themselves; making platforms “information” or “content” fiduciaries; and systematically archiving data and metadata about disinformation detected and addressed by platforms.

To understand where digital governance is going, we must take stock of where it’s been, because the timbre of mainstream thinking around digital governance today is dramatically different than it was when study of “Internet governance” coalesced in the late 1990s. Perhaps the most obvious change has been from emphasizing networked technologies’ positive effects and promise – couched around concepts like connectivity, innovation, and, by this author, “generativity” – to pointing out their harms and threats. It’s not that threats weren’t previously recognized, but rather that they were more often seen in external clamps on technological development and upon the corresponding new freedoms for users, whether government intervention to block VOIP services like Skype to protect incumbent telco revenues, or in the shaping of technology to effect undue surveillance, whether for government or corporate purposes. The shift in emphasis from positive to negative corresponds to a change in the overarching frameworks for talking about regulating information technology. We have moved from a discourse around rights – particularly those of end-users, and the ways in which abstention by intermediaries is important to facilitate citizen flourishing – to one of public health, which naturally asks for a weighing of the systemic benefits or harms of a technology, and to think about what systemic interventions might curtail its apparent excesses. Each framework captures important values around the use of technology that can both empower and limit individual freedom of action, including to engage in harmful conduct. Our goal today should be to identify where competing values frameworks themselves preclude understanding of others’ positions about regulation, and to see if we can map a path forward that, if not reconciling the frameworks, allows for satisfying, if ever-evolving, resolutions to immediate questions of public and private governance.

With public and academic attention increasingly focused on the new role of machine learning in the health information economy, an unusual and no-longer-esoteric category of vulnerabilities in machine-learning systems could prove important. These vulnerabilities allow a small, carefully designed change in how inputs are presented to a system to completely alter its output, causing it to confidently arrive at manifestly wrong conclusions. These advanced techniques to subvert otherwise-reliable machine-learning systems—so-called adversarial attacks—have, to date, been of interest primarily to computer science researchers (1). However, the landscape of often-competing interests within health care, and billions of dollars at stake in systems' outputs, implies considerable problems. We outline motivations that various players in the health care system may have to use adversarial attacks and begin a discussion of what to do about them. Far from discouraging continued innovation with medical machine learning, we call for active engagement of medical, technical, legal, and ethical experts in pursuit of efficient, broadly available, and effective health care that machine learning will enable.

John Perry Barlow's insights were inseparable from his lyrical way of conveying them. Barlow's expression mates joy and canniness, and one of his talents in writing about new technologies was to flip our conception of the status quo in order to correct it. In 1994, the conventional sense was that the Internet and its champions were heedlessly upsetting a longstanding set of relationships and legal entitlements, with copyright as a signal example. And while that was superficially true, it wasn't the whole story. This brief essay examines the legacy of Barlow's work from the vantage point of today's markedly different digital world.

In a book chartered to demonstrate intellectual property in objects, what concrete thing can represent the Internet, a phenomenon that exists only as a well-elaborated idea? Perhaps the best physical representation of the genius of the Internet—and in particular, “Internet Protocol”—is found in an hourglass.

Jonathan L. Zittrain, the George Bemis Professor of International Law and Director of the Law Library at Harvard Law School, delivers the annual David L. Lange Lecture in Intellectual Property Law.