Data has a story before it becomes a dataset.

When the Web Becomes Evidence

My first Digital Humanities reflection

One of the first ideas I am carrying into Digital Humanities is that the internet is not the same thing as a permanent record. It feels permanent because information is easy to copy, search, and share, but the readings for Weeks 1, 2, and 3 show how quickly digital evidence can disappear. A page can vanish when a company closes, a platform changes its policies, a server is not renewed, or a legal dispute changes what is allowed to remain online.

A history that can disappear

Chris Stokel-Walker’s BBC article begins with a statistic that made the problem concrete: research found that 25 percent of web pages posted between 2013 and 2023 had vanished, with older pages disappearing at an even higher rate. The article presents the Internet Archive and its Wayback Machine as an important response. Volunteers and archivists have built an enormous collection of snapshots, books, videos, and other digital materials so that researchers can still encounter pieces of the web that no longer exist in their original form.

What surprised me was how fragile that solution is. The Internet Archive is doing public-facing historical work, but it still depends on money, storage, technical infrastructure, volunteers, and legal permission. Cyberattacks and lawsuits can threaten the archive itself. That tension changed the way I think about “the cloud.” Digital information may look weightless, but it depends on physical data centers, institutions, and people willing to maintain it.

The article also made me think about what future historians will not be able to see. Much of our everyday life now happens through websites, social media, online communities, and digital services. If those spaces disappear without being preserved, the loss is not only technical. It changes the evidence available for understanding how people lived, communicated, organized, and made culture.

From scarcity to abundance

Roy Rosenzweig’s “Scarcity or Abundance? Preserving the Past in a Digital Era” approaches the same problem through history. His example of the “Bert Is Evil” website is memorable because it shows that even a strange piece of internet culture can become historical evidence. The site disappeared after its creator deleted it, yet copies and mirrors survived in other places. That story raises a difficult question: what counts as worth preserving, and who gets to decide?

Rosenzweig describes a shift from an older age of scarcity to a digital age of abundance. In the past, historians often worried that too few documents survived. In the digital era, we may face the opposite problem: too much information is produced too quickly, and much of it is poorly organized, undocumented, or dependent on software that may not exist in the future. Abundance does not automatically create better history. Researchers still need ways to select, describe, preserve, interpret, and teach from the record.

Rosenzweig’s argument also makes preservation feel less like a purely technical assignment. File formats and storage systems matter, but so do copyright, funding, public policy, institutional responsibility, and access. The Internet Archive is powerful precisely because it makes digital history available to the public, but its private and legally uncertain position shows why preservation cannot depend on one organization or one passionate founder alone.

Why this matters to a Data Science major

These readings connected directly to the way I think about datasets. A dataset can appear complete while still reflecting decisions about what was collected, labeled, excluded, or lost. A digital archive works the same way. Its search box may make a collection feel neutral, but the collection is shaped by crawl schedules, metadata, file formats, language support, permissions, and the choices of the people who built it.

The biggest takeaway for me is that preservation is also a question of representation. If a community’s websites, images, posts, or local histories are not archived, future researchers may mistake absence for unimportance. That is why Digital Humanities asks both technical and humanistic questions: How do we store a record, and whose record are we storing?

What I am carrying forward

Weeks 1, 2, and 3 taught me that the digital past is threatened by both scarcity and abundance. Some records disappear before anyone saves them, while other records accumulate so quickly that future researchers may struggle to understand them. My responsibility as a student is not to assume that digital information will take care of itself. I need to ask how a source was created, how it was preserved, what context it lost, and who can still access it.

That question will stay with me as I continue studying Digital Humanities: when I find a digital source, am I only using it, or am I also helping understand its history?

The Search Box Is Not Neutral

My first Digital Humanities reflection

Before this week, I usually thought of digital tools as ways to make research faster. Search engines find information, databases organize it, maps display it, and code helps us process it. As a Data Science major, I am used to thinking about tools in terms of efficiency: How much information can I collect? How quickly can I find a pattern? How can I make a result easier to interpret?

The Week 4 readings made me slow down and question that mindset. Digital Humanities is not just about applying technology to humanities material. It is also about asking what technology changes, what it hides, and who gets represented when information becomes digital.

Digital Humanities as intervention

In the introduction to New Digital Worlds, Roopika Risam describes postcolonial Digital Humanities as an intervention in the “digital cultural record.” Her example of the Puerto Rico Mapathon made this idea concrete. After Hurricane Maria, students, university staff, and community members worked with OpenStreetMap to improve geographic information that could support disaster relief. This was not technology for technology’s sake. Mapping became a form of public service, collaboration, and teaching.

What stood out to me was the connection between theory, praxis, and pedagogy. Theory helps us understand the historical and political problems inside digital systems. Praxis means acting on those problems by building or improving tools, archives, maps, and workflows. Pedagogy means teaching other people how to participate. The Mapathon brought all three together: participants thought critically, performed technical work, and helped others learn the process.

Risam also argues that the digital cultural record contains gaps, omissions, and older colonial patterns. Digital projects can preserve and share knowledge, but they can also repeat the same inequalities found in older archives. If some languages, communities, and histories were underrepresented in the original cultural record, digitizing that record does not automatically make it fair. The process of choosing what to scan, how to describe it, how to organize it, and who can access it all matters.

That point connected strongly to my Data Science background. A dataset is never simply “the truth.” It reflects decisions about collection, categories, missing values, labels, and access. A digital archive has similar decisions built into it. The database may look objective because it is organized and searchable, but its structure still reflects human choices and institutional power.

The opportunities and blind spots of search

Lara Putnam’s article, “The Transnational and the Text-Searchable,” focuses on how digitized sources and full-text search have changed historical research. Search makes it possible to discover people, places, words, and connections across large collections without physically traveling to every archive. Putnam calls attention to the value of “side-glancing”: looking beyond the boundaries of a familiar topic or location to notice unexpected connections.

This is exciting to me as someone who works with data. Search increases the scale and speed of discovery. It can reveal patterns that would have been difficult or impossible to find through older research methods. Putnam explains that digital search has helped historians think across national borders because information is no longer as tightly connected to one physical place.

At the same time, the article warns that easier discovery can create new forms of ignorance. Search results are shaped by what has been digitized, what can be recognized by optical character recognition, what language the system handles well, and what terms a researcher already knows to enter. The sources that are easiest to find can begin to look like the most important sources, even when they are only the most searchable ones.

Digital search does not simply open a window onto the past. It creates a particular window, with a particular frame.

Putnam’s argument changed the way I think about a search box. It can help us see connections, but it can also make us overlook local context, people who left fewer written records, and communities whose materials were never digitized. In data terms, search is powerful, but it still has coverage problems, bias, and blind spots.

What I am taking forward

My main takeaway from this week is that Digital Humanities is both technical and ethical. It asks us to build, search, map, annotate, and analyze, but it also asks us to examine the assumptions inside those actions. The goal is not simply to add more information to the internet. The goal is to create digital knowledge more carefully, with attention to context, representation, access, and power.

As a Data Science major, I am interested in the possibilities of computation. This week reminded me that computational skill is strongest when it is paired with humanistic questions. Before trusting a dataset, archive, search result, or visualization, I should ask: Who created it? What was left out? Which communities are represented? What would I fail to see if I only followed the easiest results?

That is the question I want to carry into the rest of this course and into the rest of my work with data: not only “What can this tool show me?” but also “What is this tool making harder to see?”

Searchable does not mean complete.

My reflection on language, digital archives, and handwritten text recognition

I usually think of a digital tool as something that helps me get past a problem. These readings made me think more about the problems a tool can create, or carry forward, without making them obvious. Language support, the way an archive is organized, and the examples used to train a model all affect what becomes available to a reader. A searchable result is useful, but I still need to ask how it got there.

What happens when English becomes the default?

Paul Spence and Lorella Viola’s introduction to Multilingual Digital Humanities made me look differently at something I normally take for granted. English is so common in software, documentation, and academic work that it can start to feel like a neutral starting point. Their discussion shows why it is not. Researchers working with other languages may have fewer tools, less training material, or systems that do not handle their scripts well. That can limit the questions they are able to ask before the research even begins.

What stood out to me was that multilingual work means more than translating a page or adding another language to a menu. A tool also needs to work with the material, and the people using it need a way to contribute. If someone has to change their sources to fit the software, it is worth asking whether the software is supporting their work or making them compromise it.

I also appreciated that the authors acknowledge the contradiction of discussing multilingual DH in an English-language book. They are not claiming that including several languages solves every problem. They point out that support can still concentrate around a small group of well-resourced languages. Calling a project global does not tell me whose language it actually supports.

As a Data Science major, this makes me more careful about what a successful model result means. Good performance on English material does not automatically show that the method works equally well elsewhere. As an international student, I also want to notice when participation depends on adapting to one language or one way of working. I would rather check those assumptions at the beginning of a project than discover them after choosing all my tools.

Can something be in the archive and still be invisible?

The Global Debates in the Digital Humanities extract helped me see that missing information is not always information that was never collected. In Sayan Bhattacharyya’s discussion, words from non-Western languages can appear in several romanized spellings. A tool that counts each spelling separately may treat every version as too uncommon to include. The words are there, but the assumptions behind the analysis make them harder to see.

This is a useful challenge to how I think about cleaning data. My first instinct might be to standardize the spellings and move on. That could help in some cases, but I need to understand the variation before deciding that it is just noise. Bhattacharyya’s suggestion of keeping a record of the invisibilities students discover also appealed to me. Explaining what a method leaves out can be as useful as showing what it finds.

Puthiya Purayil Sneha brings a wider view through archival work in India. Her examples of institutional, community, and personal archives show why the history of DH cannot be reduced to one familiar story about American and European universities. I was especially interested in the work that continues after a collection is digitized. Describing materials, maintaining access, handling ownership and privacy, and keeping a project running are part of the archive too. An online collection does not sustain itself.

Ernesto Priego’s reflection connects this to who gets to participate. Blogging and inexpensive tools can open opportunities, but they do not give everyone the same time, funding, infrastructure, or recognition. His point about building a community alongside a resource feels relevant to this blog. Publishing something is not the same as making it meaningful to other people. Together, these chapters make me ask who helped create a digital collection, whose work gets recognized, and which experiences its categories might leave out.

What can AI tell us about an old handwritten page?

Aleksej Tikhonov and his coauthors begin the MultiHTR chapter with the possibility of finding old family letters and not being able to read their script. I liked this example because it gives the technology a purpose beyond producing a better accuracy score. Recognizing handwriting could help someone reconnect with a family history that is physically present but difficult to access. The project works with German, Russian, Serbian, and Ottoman Turkish, where changes in scripts and writing practices create different challenges.

The process also makes clear how much human work supports the AI. The researchers use carefully prepared transcriptions, called ground truth, to train models in Transkribus. They then compare the output with manually transcribed material held aside for validation. To me, that makes the result easier to understand: the model has learned from particular examples, not developed a complete understanding of every handwritten document.

One example that stayed with me was the Serbian chemistry notebook. The squared paper caused problems for layout analysis, and the model struggled with formulas it had not seen during training. Recognizing a script was not enough to handle a different kind of page. This connects directly to the way I think about testing models. Before trusting a result on new material, I need to check how that material differs from the training examples.

The discussion of Ottoman Turkish adds another important distinction. Someone trying to understand a family document broadly may need a different transcription from a linguist studying the exact historical characters. There is not one useful output for every reader. I need to know what the result will be used for before deciding whether its accuracy is good enough.

My takeaway is not that AI transcription should be avoided. It can make difficult material easier to approach, but the original page and the choices behind the transcription still matter. If I used a model for this kind of work, I would keep its output connected to the source and make uncertainty clear. That brings me back to the questions running through these readings: who can use the tool, what does it make visible, and what do I still need to check for myself?

Automatic does not mean human-free.

My reflection on human-centred AI and the hidden labour behind ChatGPT

So far this semester I have mostly asked what digital tools make visible and what they leave out. This week the readings turned that question toward the people inside the system. One article argues that humanities students should help design AI. The other shows workers who already shape AI every day, in ways most users never see. The tool for the week, Voyant, sits in between: it gives more people a way to analyse text, while still depending on choices someone made before the analysis starts.

Who should learn how AI works?

Jon Chun and Katherine Elkins start from crisis: humanities enrolments falling, students using ChatGPT to write, and AI moving into white-collar work. Their answer is not retreat. They built an AI curriculum inside digital humanities at Kenyon College so that humanities students can help create what they call human-centred AI. One question runs through the article: will we program humanity into our tools, or will we cede our humanity to AI?

What convinced me is that they do not stop at critique. They warn that critique alone sidelines humanists from decisions being made without them. Their courses teach real tools, including Python, Pandas, scikit-learn, and Google Colab. Students examine bias at every stage of the data science pipeline, then test competing definitions of fairness on the COMPAS parole algorithm using the What-If Tool. Fairness cannot be settled by one equation. Someone has to choose which definition to use, and that is a humanities question as much as a technical one.

As a Data Science major, I am usually on the other side of this classroom. My courses go deep on how models work and lighter on who decides what they are for. Their enrolment figures stayed with me: 61 percent women and more than 90 percent non-STEM students, in courses built entirely on free tools with no textbook costs. Access is a design choice. The important questions are shifting from how a model works to who it serves and why, and I want both kinds of fluency.

Who reads the worst of the internet so a chatbot stays polite?

Billy Perrigo’s TIME investigation shows the other side of that classroom. To make ChatGPT less toxic, OpenAI sent tens of thousands of snippets of text to an outsourcing firm, Sama, in Kenya, beginning in November 2021. Workers read and labelled graphic descriptions of sexual abuse, hate speech, and violence so that a safety system could learn to detect toxic language and filter it out.

The numbers are hard to sit with. Labelers took home between about $1.32 and $2 per hour, while OpenAI paid Sama $12.50 per hour for the project. Workers told TIME they were expected to read and label 150 to 250 passages in a nine-hour shift; Sama says the figure was 70. All four workers interviewed described being mentally scarred by the material, and the wellness support they were promised was hard to reach in practice. In February 2022, Sama cancelled the work eight months early, after a related image project raised alarms about illegal content. Ending the contract ended the exposure. It also cost workers income, and some of them their jobs.

In my field, labels like these are called ground truth, the same term from last week’s handwriting recognition project. A safety model is trained on human judgment, made quickly, under pressure, by people far from the product launch. When I evaluate a classifier I usually ask about its accuracy. This article makes me ask about the conditions its labels were produced under, and whose judgment the model has actually learned. Less toxic output rested on people absorbing the toxicity first.

Every map is making a claim

My reflection on what maps show, what they hide, and who gets to draw them

This week the class worked through three documents side by side in a document lab, and together they changed how I look at maps. I used to treat a map the way I treat a chart in my data classes, as a neutral picture of whatever the data says. These readings argue something less comfortable: a map is a claim somebody built, out of choices about what to show, what to group, what to shrink, and what to leave blank. Once you see the choices, you cannot go back to seeing just the picture.

The little lies add up

"7. Lying With Maps" opens with Mark Monmonier's idea that an accurate map must tell white lies. Flattening a round Earth onto paper forces distortion, so the Mercator projection keeps shapes and stretches sizes until Greenland looks as large as Africa, when it is really only 7 percent of it. The Gall-Peters projection fixes the sizes and stretches the shapes instead. Neither is the true one. What stayed with me is how far the small choices travel: the same poverty data, drawn with different class breaks, tells three different stories, and an average calculated for a whole block can describe nobody who actually lives there. In my own coursework I have accepted default settings in mapping software without thinking about them. This chapter made me realise the defaults are arguments too, they are just arguments nobody has to defend out loud.

A city built by cotton and slavery

The Guardian's "Cotton Capital" investigation tells the history of Manchester through a visual tour of its own maps, from a small manorial town of about 9,000 people in 1700 to Cottonopolis, the world's first industrial city. Its argument is that slavery was not beside this growth, it was inside it: Manchester cloth sold into the slave trade, mills running on slave grown American cotton, a bank funded by slave trading profits, and, after abolition in 1833, GBP 20 million in compensation paid to the enslavers and reinvested, while the formerly enslaved received nothing. I keep thinking about the source notes at the end. The map of slave trade linked businesses shows 31 addresses, and then admits there are five more known suppliers whose addresses were not documented well enough to map. The map looks complete, but it is really a map of what got written down. Anyone who works with data knows that gap well. Missing records are not the same as missing businesses.

Rebuilding a city the archive tried to forget

Ademide Adelusi-Adeluyi's "Mapping Old Lagos" asks what you do when the written record was mostly produced by colonizers and the indigenous city survives mainly in oral tradition and in the ground itself. Her project, New Maps of Old Lagos, georeferences British colonial maps and layers them over the modern city to recover what was paved over: the original shoreline of Lagos Island, its filled in lagoons, and the steam tramway that ran from 1902 to 1933 and then vanished so completely that its records were destroyed. Two details stayed with me. First, she is honest that colonial maps cannot really present indigenous understandings of space, so the method only works if you read those maps against the grain. Second, the project's fancy interactive map tool never worked reliably in Lagos itself because of bandwidth, so the maps are also offered as simple GIFs. Designing for the audience you actually have turns out to be part of the scholarship, not a technical footnote.

What the three say together

Read together, the three documents describe a ladder. The chapter shows that every map distorts, innocently or not. Cotton Capital shows a map series making a moral and historical argument about a city and an empire. Mapping Old Lagos shows maps being used to argue back, to recover people and places that the official record left anonymous. For everyone who makes or reads maps, and that is everyone now, the working habit is the same set of questions the chapter opens with. Who made this map and why? What is included, and what is excluded? Where did the data come from? I would only add the data science version of the same habit: show your settings, name your missing records, and never trust an average to speak for the individuals inside it.

The transcript is not the point.

My reflection on crowdsourcing, participation, and who does the work

Every crowdsourcing project I have ever seen judges itself the same way: pages transcribed, records tagged, items identified. Output. Trevor Owens, who ran these projects at the Library of Congress, says that way of counting misses the point so completely that he calls it upside down. In his 2012 post "Crowdsourcing Cultural Heritage: The Objectives Are Upside Down," he argues that the transcript was never the real product of crowdsourcing. The participation is. This week we read the post and annotated it together as a class, and it rearranged how I think about every project that asks the public for help, including some of my own.

Participation is the product

Owens starts from the mission. A library or archive exists so that people can explore the past and feel connected to it, and a person who spends an evening transcribing a Civil War diary gets more of that mission than a hundred people who skim the finished transcript later. The transcriber follows the story day by day, learns the writer's handwriting, starts to care what happens next. In the diaries project Owens describes, use jumped from about a thousand visitors a day to more than seventy thousand, donors showed up, and one volunteer said the people in the diaries had become almost an extended part of his family. None of that shows up if you only count pages. The data still gets made, transcripts are a real and useful by-product, but it is the by-product. The experience is the product.

Meaningful work, or don't bother

Not any task earns that kind of engagement. Owens puts meaningful activity at the top of his pyramid, above the gimmicks projects use to keep people clicking. The difference is whether the work asks something real of you. Transcribing a difficult hand teaches you the hand. Sorting something a machine could have read teaches you nothing, and volunteers can tell. One regular in Owens's project rationed himself to two or three hours a day so the work would last longer. Nobody rations busywork. The lesson I wrote in my notes this week: if the task could be automated without losing anything, asking a person to do it is not participation. It is just unpaid data entry wearing a friendlier name.

The dark side has a name

Owens is just as clear about how this goes wrong. When institutions treat volunteers the way Amazon treats Mechanical Turk workers, as the cheapest possible source of data, crowdsourcing turns into what he calls exploitationware: points, badges, and leaderboards designed to trick people into free labour. The costs are real even when no money changes hands, because trust is the currency these projects run on, and a burnt volunteer does not come back. Reading this two weeks after our unit on AI training data gave it an edge for me. The same tasks Owens describes are now done quietly by paid workers in places like Kenya so that chatbots stay polite, which our Week 7 reading covered. The question is the same in both rooms: would the person doing this work call it participation, or would they call it a shift?

What I am taking from it

I run projects that ask people for their time and attention, this blog included, so I cannot read Owens as theory. His test is simple and a little uncomfortable: judge the project by what the participant gets, not by what the institution collects. A project that leaves people knowing more, feeling connected to something bigger, and wanting to return has done crowdsourcing right, even if its page count is modest. A project with huge numbers and bored, tricked, or invisible workers has failed at the only objective that was ever really on the table.