The Ghosts of Clipart
Flash Fiction, Alt Text, and Steganographic Narrative

In January 2025 I presented on a panel called “Alt-Text, Alt-Image: Multimodal Scholarship” at the annual Modern Language Association conference. My talk wasn’t so much about multimodal scholarship, but it did take the Alt Text part of the title quite literally. Here’s that talk, along with key slides.

I want to talk today about alt text—actual alt text—and its creative uses, and arguably, misuses. I am going to describe a creative project of my own that I see as an act of resistance, and maybe even a model for resistance, on several levels, to the rise of AI-generated content.

But before getting to my project, let’s talk briefly about the origins of alt text. 

In June 1993 computer scientists Tim Berners-Lee and Daniel Connolly, working on behalf of the Internet Engineering Task Force, published a report called simply, “Hypertext Markup Language (HTML).” The pair proposed—and by doing so actually invented—the fundamentals of a textual encoding protocol that all of us use every day. Among the various HTML tags they proposed is the image tag for displaying images inline. The image tag requires a source attribute, which points to the URL or relative location of the image. Another attribute was optional: alt. The purpose of this “alt” attribute was to display optional text in place of graphics in “text-only environments” (Berners-Lee and D. Connolly 18) . 

<img src="cat.jpg" alt="cartoon of an orange tabby">

At the time, “text-only environments” described most of the  Internet. Given low bandwidth and the prevalence of Unix terminals that couldn’t readily display images, text was all there was. 

Two years later, the Internet Engineering Task Force formalized the specifications for HTML in a document called “Hypertext Markup Language – 2.0,” once again authored by Berners-Lee and Connolly. According to these official specifications for HTML, ALT text was “text to use in place of the referenced image resource, for example due to processing constraints or user preference” (Berners-Lee and D. W. Connolly 35).

Keep these two rationales for alt text in mind as we move on. The first is technical—“text-only environments” and “processing constraints.” The second is “user preference.” User necessity doesn’t come into play. The idea of adding image descriptions to accommodate low-vision users only came about a few years later. From the beginning, alt text was seen as a choice.

Today it’s the the World Wide Web Consortium that designs and sets the standards for HTML. The consortium outlines seven types of images found online—such as “decorative images,” “functional images,” and “image maps”—and describes how alt text should be used in each case. 

While many websites adhere to these guidelines, casual posters on social media may not. Twitter didn’t add the ability to include alt text until in 2016. A study by Carnegie Mellon researchers in 2019 found that out of one million tweets with images, a scarce 0.1% included alt text (Gleason et al. 550). 

The culture of alt text is arguably stronger on what internet scholar Robert Gehl calls  “alternative social media”—which Gehl characterizes as media platforms that (1) oppose “the commercialization of speech”; (2) offer users “more access to shape the underlying technical infrastructure”; and (3) which resist “surveillance regimes” (Gehl 2). Mastodon—or rather, the fediverse, of which Mastodon is just one instantiation—exemplifies alternative social media. The first call to add alt text capabilities to Mastodon came in 2016, not long after the first working version of Mastodon was released. And because Mastodon is open source we can see that discussion unfold on its Github repository. Developers explicitly describe the need for alt text in terms of accessibility: “it’d be cool to see this become a thing so we aren’t excluding blind/visually impaired people” says one contributor. In relatively short order for a sprawling open source project, the ability to attach alt text to uploaded images was added to Mastodon in September 2017. 

Early users on Mastodon fostered a strong culture of accessibility, with many insisting that images should have alt text. Others avoid third-party apps that don’t accommodate alt text. And many users will not “boost”—Mastodon’s equivalent of retweeting—image posts that don’t include alt text. 

And this is where my project comes in. On the surface it’s a silly little bot that posts old clipart to social media.

The clipart comes from a CD-ROM that I purchased back in 1995, the same year the HTML 2.0 specifications were finalized. I still have the original CD-ROM and the accompanying book, and at one point in 2017 it occurred to me that sharing the old clipart on Twitter might be cool.

It wasn’t easy. The clipart on the CD-ROM is a defunct proprietary format. Assuming your computer even has a CD drive, no modern operating system will recognize the files. In order to access the clipart I had to emulate Windows XP in a virtual machine, install the software from the CD-ROM on an external drive, and export the images one by one into a usable image format. 10,000 images, one by one. 

It was tedious work but finally on August 9, 2017, I released a bot that automatically posts random clipart images to Twitter, one every two hours. 

The bot simply posted the file name and the image. There was no alt text. Once it had cycled through all 10,000 pieces of clipart it would restart, posting them again, one-by-one, but in a different order. 

With its chunky graphics and frequent references to 80s and 90s technology and celebrities (like, say, Vanilla Ice), the bot tapped into the deep well of nostalgia that hides just under the surface of any part of the Internet. The bot racked up nearly 30,000 followers before Twitter became X and Elon Musk shut down access to the Twitter API. I unceremoniously deleted the account forever from Twitter in October 2022.

It wasn’t long before I reconstituted the bot on Mastodon. There was a problem, though. As I’ve mentioned, the ethos on Mastodon prioritizes accessibility. Fans of the bot told me they were happy I migrated it to Mastodon, but they wouldn’t be boosting its posts because there was no alt text.

I wanted to include alt-text, but how? I considered several possibilities. The first that occurred to me was somehow crowdsourcing the alt text. I’m a digital humanist after all. What if I made some sort of platform where people could adopt a piece of clipart and write the alt text for it? I decided against this though; crowdsourcing seemed like a possible content moderation nightmare, especially in the mid 2020s.

The next idea that occurred to me was to use AI to generate alt text. Remember, this is late October 2022. ChatGPT hadn’t been released yet. The options at the time were all clunky GPT-2 class LLMs. And the resulting alt text lacked crucial context or affective nuance. The feeling among the disability community and accessibility advocates at the time was that AI-generated image descriptions were often worse than no image descriptions at all.

Now here is where things get a bit complicated. It turns out, and this is something most people who write or use alt text don’t realize, that alt text and generative AI are already deeply entangled with one other. The reason a tool like Dall-e or Midjourney knows that the word banana equals “curved yellow thing with tapered ends” is because they were trained on vast datasets of images paired with textual descriptions of those images, which is to say, alt text. Quite simply, AI image generators would not exist without alt text.

Let me highlight a few features of the data infrastructure of image generators. 

LAION is a non-profit organization that releases vast datasets of image/text pairs. In March 2022, LAION released a dataset of 5 billion image-text pairs. That is, 5 billion images, each with a supposedly accurate image description. These image descriptions primarily came from the alt text that accompanied the images wherever they were found online. Of course, whatever biases, stereotypes, and harmful misconceptions that were in the original alt text leak downstream to the LAION dataset. AI researcher Abeba Birhane has identified many troubling image-text pairs in LAION. Writing about an earlier LAION dataset released in 2021, Birhane found “even the weakest link to womanhood or some aspect of what is traditionally conceived as feminine returned pornographic imagery” (Birhane 4).

And it gets worse. About a year after LAION-5B was released, researchers at the Stanford Internet Observatory discovered hundreds of instances of child sexual abuse material in the dataset (Thiel 7)—images that a tool like Stable Diffusion could use to generate realistic pornographic material. About eight months later, LAION released an updated version of its 5 billion image dataset, this time, purged of any known links to child sexual abuse material.

How did such material end up in the LAION datasets in the first place? Birhane critiques the dominant method of building AI training datasets as “crawl over curate” (Birhane 5), in which such mind-bogglingly vast datasets are accumulated without discernment under the presumption that “scale beats noise” (6)—in other words, that the mistakes, the biases, the child porn, are “somehow ‘averaged out’” (6) because of the massive size of the datasets. This is a fantasy in Birhane’s point of view, a fantasy that underwrites the entire AI techno-industrial complex.

Circling back to my clipart project, something felt wrong about outsourcing my work to a large language model that itself was built on the outsourced labor of unrecognized and uncompensated writers of alt text, especially if that alt text was going to be problematic.

So that’s when I decided I would write my own alt text. The first pieces were rote, functional image descriptions that would have made the World Wide Web Consortium proud. But I quickly realized this process could become mind-numbingly tedious. And I also quickly realized it didn’t have to be. Inspired by Randall Monroe’s use of alt-text on his webcomic XKCD, where hovering your mouse over one of his comics reveals an additional joke or piece of commentary, I wondered if I could hide, so to speak, a tiny self-contained story in the alt text. Flash fiction.

After a few days of intense writing, I began reusing some of the characters who appeared in the early pieces. Suddenly I wasn’t just alt text, I was writing flash fiction, and I wasn’t just writing flash fiction, I was weaving a web of interconnected pieces of flash fiction.

So Vanilla Ice isn’t just Vanilla Ice. There’s a story behind him. This one mentions the character of Phil, who, in one of many meta moments of the project, makes clipart for Totem Graphics, one of the, in real life, independent graphic design shops in the early 1990s that licensed its content—clipart—to Corel.

Behind Tina Turner’s clipart is a glimpse of a character named Ann, who appears to be a tradwife but who is in an unhappy marriage.

Then there’s Dr. Slife, a creepy dentist who collects patients’ teeth.

All told, there about 30 recurring characters and many other one-off characters. The stories jump between past, present, and future. There’s apparently some sort of war with the fae folk going on. In the future, narrated by a character who speaks in the first person “I”—who is not myself—the world is in ruin. My inspirations include Jeff VanderMeer, Carmen Maria Machado, Shelley Jackson, and Stuart Moulthrop, and many others.

I began this project, which I call “The Ghosts of Clipart,”in December 2022. For months I averaged writing about 550 words of alt text a day. I kept up that pace until the summer of 2023, which I took off, having amassed about 80,000 words. My pace has slowed down since then. Right now I’m at 150,000 words of interconnected flash fiction for 1,800 pieces of clipart. The shortest story is 7 words and the longest is 282.

This is my Scrivener workspace:

It’s how I write all the pieces and keep track of everything. Scrivener’s built-in stats tool tells me that if this were a paperback book, it’d be about 400 pages long.

The work rewards what Jason Mittell calls “forensic fandom”—it invites its audience to drill into the text and find connections under its surface (Mittell). This means many followers of the bot—which now posts to Bluesky as well as Mastodon—don’t read the flash fiction, or may not even know they’re there. My fancy name for this kind of hidden writing is steganographic writing, steganography referring to the technique of hiding a message in another medium, such as replacing part of an image file’s hex code with plain text in such a way that the image appears unaltered to the human eye. The hidden nature of this writing is amplified by the fact that the bot periodically deletes its old posts, leaving no archival traces of the flash fiction stories.

Am I bothered that most of the thousands of people who follow the bot don’t read these stories I have spent hundreds of hours writing? Not really. As an academic, I’m used to people not reading what I write. And, I love the moment when somebody does realize there’s something going on in the alt text. There’s the sense of wonder that is such a joy to see.

It is interesting to me that most people who comment on the alt text seem to think the alt text comes out of nowhere, as if the text is always already there.

There’s some sort of short-circuit with the author-function, even though the bot’s profile page is clear that there’s a person writing the text. I wonder, is this erasure is a symptom of a broader crisis of authorship online, which is to say, a crisis of labor online, in which the actual workers who make possible all the content we consume, AI-generated and human-produced alike, disappear?

Before I wrap up, I want to mention the future of this project, which I suspect will take at least a decade to complete. While multimodal LLMs were not up to the task of generating meaningful image descriptions in 2022, they have made remarkable progress.

We’ve reached the point where you can run LLAVA, the Large Language and Vision Assistant LLM, on a consumer laptop on the command line, point it to an image on your computer, and ask for a description. Seconds later, you’ll get a fairly accurate, if uninspired piece of alt text. You can even do this at scale, asking LLaVA to generate image descriptions for an entire folder’s worth of images, hundreds of images in a matter of minutes. Here’s LLaVA’s suggested alt text for a particularly niche piece of clipart showing an engine and some diagrams.

The image appears to be a graphical representation or illustration, possibly of an assembly process. 

It includes two primary components: 

1. A blueprint or diagram on the left side, which shows a top-down view of what looks like a mechanical structure with various parts labeled "A," "B," and so forth. The labeling suggests that these are components that might be part of an engine or another complex machinery. 

2. On the right side, there is an image of an assembled structure, which could represent the completed product from the left diagram. It has a purple outline and is connected to two other objects: one on top, which seems to be a cover or cap, and another at the bottom that appears to be a base or support. In the center of the image, there is a red circle with a small black plus sign inside it, which could symbolize an "add" or "include" function in the context of assembling or designing the machinery. Above and to the right of this central icon, there's a car, which might indicate that the assembled structure (on the right) is related to automotive technology or mechanics. The car is partially visible, showing only the front wheel and a small portion of the vehicle's interior. 
Overall, the image seems to be a conceptual diagram used for instructional purposes, likely in an educational setting or for engineering illustration.

It’s…accurate. But like much of what AI produces, somewhat…uninformed by the human experience. Here’s my alt text, featuring a recurring character named Matt. 

A diagram of a car engine, with a graph in the top left corner of horsepower and engine speed. Figures missing from the graph include: (a) force of impact; (b) intensity of scream; (c) number of shards into which the windshield shattered; and (d) the unbearable weight of Matt’s grief.

“The Ghosts of Clipart” pushes the limits of accessibility, visibility, and narrative—not to mention creative endurance. It appears to be an automated bot (and it is), but it is also a resolutely human-authored text, a doubling down on human labor in the age of generative AI.

Suffice it to say, I won’t be using AI to complete this project. Even though it’s a grind, and perhaps even impossible, and even though most who encounter it will have no idea what’s going on, it is, for me, an humbling, grounding writing experience, an utterly, utterly, human experience.

Works Cited

• Berners-Lee, Tim, and Daniel Connolly. Hypertext Markup Language (HTML). Internet Engineering Task Force, June 1993, https://www.w3.org/MarkUp/draft-ietf-iiir-html-01.txt.
• Berners-Lee, Tim, and Daniel W. Connolly. Hypertext Markup Language – 2.0. Request for Comments, RFC 1866, Internet Engineering Task Force, Nov. 1995. IETF, https://doi.org/10.17487/RFC1866.
• Birhane, Abeba, et al. “Multimodal Datasets: Misogyny, Pornography, and Malignant Stereotypes.” arXiv:2110.01963, Oct. 2021. arXiv.org, http://arxiv.org/abs/2110.01963.
• Gehl, Robert W. “The Case for Alternative Social Media.” Social Media + Society, vol. 1, no. 2, July 2015, https://doi.org/10.1177/2056305115604338.
• Gleason, Cole, et al. “‘It’s Almost like They’re Trying to Hide It’: How User-Provided Image Descriptions Have Failed to Make Twitter Accessible.” The World Wide Web Conference, Association for Computing Machinery, 2019, pp. 549–59. ACM Digital Library, https://doi.org/10.1145/3308558.3313605.
• Mittell, Jason. “Forensic Fandom and the Drillable Text.” Spreadable Media, 8 Dec. 2018, https://web.archive.org/web/20181208130547/http://spreadablemedia.org/essays/mittell.
• Thiel, David. Identifying and Eliminating CSAM in Generative ML Training Data and Models. 2023. purl.stanford.edu, https://doi.org/10.25740/kh752sm9123.