I haven’t paid much attention to NaNoWriMo–the organization behind National Novel Writing Month, aka “write 50,000 words of a novel in the month of November”–for several years. My wife participated in 2003, and I followed her in 2009; our experiences have grown into WriteTrack, though I’ve expanded it to support any size writing goal and user-defined start and end dates. I believe the last time I did anything with NaNoWriMo’s website was several years ago, when I found out that they had removed the API that allowed WriteTrack to automatically pull your word counts.
This week, I’ve noticed a surge of people registering for WriteTrack, many stating that they’re “Looking for a NaNoWriMo alternative”. This has been caused by the release of a statement by the NaNoWriMo staff condemning people who condemn AI, stating that “to categorically condemn AI would be to ignore classist and ableist issues”. They’ve updated the statement, but honestly, the update doesn’t really seem to help much, largely (in my opinion) because they don’t give any indication when they think the use of AI “conflict[s] with [their] values”, leaving the general impression that they are supportive of the use of AI for writing, with maybe one or two small exceptions.
My opinion: Modern generative text tools, like ChatGPT, are not “intelligent”, and authors should not ask them to write their books for them. This is not ableist or classist; this is a sober judgement from a career computer scientist. Large language models (LLMs) are effectively statistical machines, predicting what the “best” next word would be based on all of the writings they have been trained on. While their abilities to understand requests and generate well-written outputs are frequently impressive, they don’t actually understand the world the way we do. Note: Because I believe that LLMs are not intelligent, I try not to use the term “AI” when describing them.
Lack of Intelligence: There have been many demonstrations of the fact that LLMs aren’t applying any actual intelligence to their outputs. One example is asking “how many R’s are there in the word ‘strawberry’?“; ChatGPT’s default answer is that there are two. While there’s a charitable explanation (generally, people ask this sort of question when they’re not sure if they should use one or two of a letter), it’s at best lacking. More significantly, however, are the many examples of “jailbreaking” LLMs to get them to generate content that their creators don’t want them to. The simplest jailbreaks have the LLM “mask” their outputs, which make it clear that the LLM itself is often able to generate the output, but that there’s a secondary system which scans the generated output for inappropriate content which needs to be blocked. A truly intelligent system would not have that vulnerability. Finally, truly intelligent systems would not have the issues with “hallucinations” that LLMs are notorious for (such as the lawyers who have been sanctioned for citing non-existent cases in briefs that were, in large part, generated via LLM).
Training: A huge problem, when thinking about using LLMs for writing, is that the outputs they generate are based on the texts they’ve been trained on. This is different from the old “in order to write well, you should read as much as possible” advice, because while a human who reads extensively develops a sense of good writing, understanding of allegory, and determines for themselves what is beautiful, LLMs simply know what words are most commonly used together. LLMs can be “tuned” by tweaking their learning sets, but their very nature leads to the conclusion that they are ultimately just really powerful plagiarism engines. I know that’s a sweeping, controversial statement, but LLMs generate their text based on many sources, without the ability to show attribution. And that’s before we even get into the fact that many, if not all, LLMs are trained on texts without the original authors’ express permissions, including from pirated eBooks, scraping websites, and so on.
Acceptable use for LLMs: Given the above, you might think that I’m totally opposed to using LLMs. And the answer is, not really. There are use cases for them; as I noted above, their ability to understand questions and generate reasonable outputs is often impressive. In my day job, we use an LLM to extract data from large documents, note patterns in data, and, yes, create a first draft of communications. I’m proud of the fact, however, that our company emphasizes a “human hands on the steering wheel” approach–while the LLM makes suggestions, a human reviews, revises, and fixes the output. I’d recommend a similar approach to applying a LLM to writing: work with an LLM to flesh out a general story outline; pose problems from your story to the LLM and ask it to suggest a list of possible solutions; ask an LLM to suggest reasons your main character has a specific flaw. Then take those outputs, think about them, and take your story in a direction that speaks to you. But don’t ask the LLM to write a scene for you–that’s not going to be your work, and won’t reflect you as a writer.
Having gone through NaNoWriMo, I can attest that the challenge of putting as many words on paper as I could each day forced me to turn off my internal editor, freeing my creative side. There were genuine moments where I paused and said, “I don’t know where that came from, but I like it.” Using an LLM would prevent that from happening. Is this ableist or classist? Hardly! It’s not a question of are you able to write without an LLM; the only thing that using an LLM would help with is your ability to write a specific number of words in a given timeframe. There’s nothing magical about “50,000 words in 30 days”; some people need a lower (or higher) target, and they need to find what works for them to push the internal editor out of the way. And NaNoWriMo’s “classism” argument seems to be aimed at the review/editing process, rather than the writing process, which is really odd given that NaNoWriMo’s focus is supposed to be on, well, writing. (Incidentally, the statement that some communities would need to pay someone “for feedback and review” is striking; do they assume that there are communities where no one is able or willing to read a friend’s work for free?)
In summary: NaNoWriMo was wrong to call protests against “AI” “classist and ableist”. Large Language Models can be useful, but are no substitute for the creative genius that a human can provide when they turn off their inner editor. Letting a LLM write entire scenes (or more!) for you is not very different from plagiarism, and should not be encouraged by an organization aimed at encouraging everyone to write.
Edit: Since the time of my post, they updated the statement again, so I’ve updated the link to an archived copy for context. I feel that the newly updated statement is as weak as you can get–“AI is simply too big and too varied to categorically support or condemn”….”We made mistakes in our initial expression of this position”, with no boundaries or examples at all. There’s also an undated “letter to our community” that tries to walk back much of what they said, including the language about ableism and classism, but I find it unimpressive and unpersuasive. You can judge for yourself, of course.