Introducing "AI & How We Teach Writing"

a newsletter from W. W. Norton and Annette Vee

How we got to ChatGPT and what to expect in this newsletter

On Valentine’s Day in 2019, a little-known company called OpenAI published a blog post announcing GPT-2, an AI language model that could do: “rudimentary reading comprehension, machine translation, question answering, and summarization.” Back then, GPT-2 was merely the next step in one AI research company’s agenda—a leap, to be clear! But it seems no one intended to disrupt higher education, at least not at first. Yet less than three years later, OpenAI launched ChatGPT, which, among educators in 2025, needs no introduction.

What was remarkable about GPT-2 wasn’t that it could perform those reading, writing, and translation tasks, but that it could do them without specific training. Previous models were trained to answer specific kinds of questions or to comprehend narrow domains. GPT-2 was, instead, given 40GB of text from 8 million webpages on the internet and then “trained simply to predict the next word” in that text. The result was a general-purpose language model that could handle new texts and tasks, even ones its designers had never thought of. That’s why it’s called a GPT: generative pre-trained transformer. Using transformer technology, the model generates text based on training that precedes any specific task. Put another way, GPT-2 was the first language model that reads and writes pretty much anything you want.

What’s interesting, or perhaps harrowing depending on how you feel about it, is that this announcement of GPT-2 offers an alternative history to our current timeline. Just after OpenAI describes the model’s capabilities, they write:

Due to our concerns about malicious applications of the technology, we are not releasing the trained model. As an experiment in responsible disclosure, we are instead releasing a much smaller model for researchers to experiment with, as well as a technical paper.

OpenAI feared what GPT-2 could do in the wrong hands: generate fake news; impersonate others; automate spam or abuse through social media. They knew, in other words, that this powerful language model was potentially dangerous, and so they weren’t going to release it.

As a writing researcher interested in technology, I was excited by this choice. Look! The technologists see how powerful writing can be! Automated writing through AI language models could be used maliciously, but OpenAI also noted that this technology could improve dialogue assistants and speech recognition for accessibility or could provide translation between languages. I saw that potential, too.

It all seems kind of quaint now. Because eight months after OpenAI said they weren’t releasing the full model, they released the full model. They’d been under intense pressure to do so, with critics saying that they were overstating the risks, stifling innovation, and wasn’t their name OpenAI?? The release of GPT-2 didn’t make the splash that ChatGPT did two years later because you still needed to go through some extra steps to access the model. But it was a public launch of a powerful language model that carried significant risk—and promise.

OpenAI titled GPT-2’s launch post “Better language models and their implications.” Since this newsletter explores the implications of AI for teaching writing, I figured this was a good place to begin.

About this newsletter

I’m not an AI enthusiast. When I run a workshop on AI and teaching for writing faculty, I often say I wish I had a magic wand that would take us back to the time before ChatGPT. As a writing researcher, I find it thrilling to watch this writing technology roll out. What a time to be alive! AI large language models are the most rapidly adopted writing technologies in the history of the world! But as a writing teacher, I’m much less enthusiastic. These technologies promise efficiency and ease—which are not values I emphasize in a writing class. AI forces us to rethink and revise so much of our teaching. And I’m still tired from pandemic virtual teaching.

But here we are. We didn’t ask for this disruption, but we’re all in it together. One of the things I did as a Writing Program Administrator (WPA) during the pandemic was to call colleagues together to tackle our collective challenge. My hope is that this newsletter similarly helps us to navigate AI and how we teach writing. I’ve been living with AI and writing for a few years. I’m glad to share what I’ve learned, plus learn more from you in this space.

Twice a month, I’ll write about some contemporary aspect of AI that’s relevant to teaching college writing. Topics might include critical AI literacy, citing AI, how students are using AI, and how we might resist AI in our classes. I’ll always include some references and ideas for how you might bring the conversation into your classroom. I teach graduate and undergraduate classes at a large, state-supported, research-intensive university—University of Pittsburgh—and I started my teaching career in high schools. That’s my context, but of course, yours will be different. You are welcome to adapt anything I share here to your own teaching.

I’ve been writing on Substack for a year now, and I like the format for publishing timely pieces on AI. (You can check out my posts here.) In the middle of my career and deep into my second book, these posts remind me that I love to write. Writing is a mode of thinking—and isn’t that why we’re here? So, when Norton asked me about launching a newsletter on AI and writing to support teachers and their students, I couldn’t say yes fast enough! I’m excited to work with Norton because I get to connect with more teachers about AI. I hope that you feel welcome to share your experiences and ask questions in the comments. Please reach out to me directly if you have a specific question or an idea for a topic you feel should be covered here.

In the classroom

And here’s the part of the newsletter where I offer a few ideas for how to take this conversation about AI into the classroom.

In my upper-level writing classes focused on writing technologies, I’ve assigned the OpenAI blog post introducing GPT-2 in 2019, “Better Language Models and their Implications.” Students find it an interesting historical document for a technology they’ve all used. The post gives examples of how the model completes text-based prompts, comparisons to other contemporary models, and examples of how this model could be used as assistive technology or for malicious purposes.

Here are some questions I’ve asked in discussion:

• Was it responsible for OpenAI to release this smaller model and hold back the more powerful one?

• What if OpenAI had continued to keep the more powerful model private?

• Which implications of “better language models” does the post get right, and which ones does it miss?

Much of our current moment is foreshadowed in this post, but OpenAI misses some things, too. For instance, they mention nothing about education in their original post in Feb 2019, or the follow-up in Nov 2019 that announced the full model release.

If you’d like to extend the discussion into hands-on work with current language models, students can input versions of these questions into publicly accessible language models such as ChatGPT, Gemini, and Claude. I asked these three models, “In 2019, was it responsible for OpenAI to release the smaller model of GPT-2 and hold back the more powerful one?” The differences in their answers reveal variances that could spark conversation about the styles and values they express.

Students who want to learn more about GPT-2 can read its explanatory model card. Those who want to work on their technical skills can access GPT-2 on Hugging Face. An article from the Verge from Feb 14, 2019, “OpenAI’s new multitalented AI writes, translates, and slanders,” highlights the significance of GPT-2’s release and provides contemporary context.

Showing the history and contingency of these technologies helps students to see that they aren’t inevitable and that humans are behind their design and deployment. Knowing this about language models is part of critical AI literacy.