by Kevin Breen

This article is the first in a biweekly series that will examine AI and its implications for publishers and authors.

In 2025, the term “artificial intelligence” is as ubiquitous as it is vague. It seems as though every organization and industry has begun to adopt AI technology, helping users drive cars, streamline rote work tasks, and even draft fan mail to Olympians. But what gets identified as “AI” continues to shift. Meanwhile, the technology itself continues to change rapidly, doubling the confusion.

As publishers and literarians, we know this issue well: language adapts over time, often complicating the meaning of certain words. (Literally!) But knowing the origin of the words we use helps us more deeply understand their meaning. Likewise, understanding what AI has meant historically, how we use the term today, and what aspects of AI are relevant to the world of publishing will help us determine the technology’s potential impact on our work. 

History of AI

What AI might have invoked twenty years ago is different from what it means today. In 2025, we don’t typically visualize the same outcomes that Alan Turing, John McCarthy, and other “founders” of AI imagined. Instead, we think of tools like ChatGPT, DeepSeek, DALL·E, and other software integrations that answer questions, draft content, and iterate as requested.

The term “artificial intelligence,” first coined in Alan Turing’s research, has existed since the 1950s. Over the following decades, artificial intelligence has been used to describe Deep Blue, a chess-playing computer from the 90s, as well as anthropomorphic robots, “smart” appliances, and natural language processors like IBM’s Watson from the 2010s.

The contemporary inception of AI

Alan Turing was a twentieth century British computer pioneer who grappled with the question of artificial intelligence throughout his career. As early as in his 1948 writings (which became Intelligent Machinery), Turing was discussing concepts like neural networks and potential biases programmers might encounter while building a computer’s “mind.” Turing’s 1950 publication of Computing Machinery and Intelligence included a description of the famous “Turing Test,” used to gauge a machine’s relative intelligence. 

Importantly, “intelligence” has always been the core debate within the field of artificial intelligence. In 1997, this debate manifested in the form of IBM’s Deep Blue, a computer “intelligent” enough to beat world chess champion Gary Kasparov. In the 2006 paper “The History of Artificial Intelligence,” authors Ting Wei Huang and Christopher Smith write, “Few [were] surprised by a computer beating a world chess champion.” And still, the response to such a technological achievement was mixed. That’s because Deep Blue was a Type-A, or brute-force, computer. Armed with massive branching decision trees of potential chess moves, the computer would examine thousands of options and select at the best scenario. This was seen as less intelligent or even interesting to the general public, compared to Type-B programs which use “strategic” intelligence to examine a narrower, scenario-specific set of potential chess moves. In other words, Deep Blue’s achievement lacked finesse in the eyes of many, since it was playing a memory-matching game and not a strategic one.

A more human “intelligence”

This has been the contemporary focus of AI technology. Nowadays, AI is an invocation for technology that mimics human intelligence and communication. Before, AI usually meant a robust computer that could access pre-cached information almost instantly. In 2025, AI most often refers to computers that “learn” by ingesting data and then perform new tasks without being explicitly programmed to do so. Think of today’s AI outputs as music played by a pianist who can improvise, based on their extensive experience reading sheet music. When technology can go off-script in this manner, we refer to it as “generative AI.”

The specifics of generative AI

The hallmark of generative AI, or genAI, is the new content it produces. Of course, using terms like “produce” and “new content” to characterize genAI has been a hot-button issue since tools like ChatGPT hit the mainstream. Critics would argue that these terms anthropomorphize the technology and erroneously credit genAI for “creating” content that has been stolen from uncredited human authors. (We talk about the ethics of genAI later in another article.)

As mentioned earlier, the term AI was historically confined to predictive models based on finite information. It was a hugely complicated memory-match game, with specific inputs pointing the model to fixed outputs. But generative AI produces outputs not already provided to the model. How?

Generative AI models are trained on vast sets of data. Initially, this data is often labeled by humans, which is referred to as a supervised training period. Here are some types of data annotation important to an AI model’s success:

• Part-of-speech tagging: Indicating to the model which words are verbs, nouns, adjectives, etc.
• Sentiment analysis: Categorizing clauses or words by their underlying meaning and usage. Sentiment analysis applies labels like positive, negative, or neutral to input data.
• Name entity recognition: Labels that classify and identify proper nouns and other specific attributes.

Eventually, most genAI models go on to “unsupervised learning,” where they start to ingest unlabeled data without specific guidance and parameters from humans. To be as realistic and useful as possible, these models require immense quantities of data. According to estimates, GPT-3 was trained on around 45 terabytes of text data, or roughly one million feet of bookshelf space. And some estimates say that AI models could ingest all the internet’s available text data as soon as 2026.

GenAI versus LLMs

Most relevant for publishers is the technology of large language models (LLMs), a subset of genAI. Whereas genAI refers to machine-learning outputs across various mediums (images, video, audio, and more), LLMs produce text. Large language models are what we often think of when we picture genAI. These services include ChatGPT, customer-service chatbots, and the newer top-of-page AI overview suggestions from Google searches. In each case, a large language model is trained on input data (in supervised and unsupervised settings) to produce authentic-seeming outputs, be they short answers to questions, genre-specific short stories, free verse poems, or other outputs.

What does this mean for literature and publishing?

Understanding genAI and LLMs is crucial for publishers. Doing so helps us speak a shared language, identify instances of the technology’s use, and anticipate which areas of our working lives are most open to its influence. This foundation of knowledge can also help publishers of all sizes clarify their stances on the subject and articulate clear messaging to authors, readers, and other stakeholders.

Throughout this series, we’ll examine the present-day impact of genAI on publishing, examine its ethics, discuss the technology’s pros and cons, and consider the manifestations of genAI in our daily work: drafting, acquiring, editing, designing, producing, and sharing creative writing with the world.

Kevin Breen lives in Olympia, Washington, where he works as an editor. He is the founder of Madrona Books, a small press committed to place-based narratives from the Pacific Northwest and beyond.