Have some shame, NaNoWriMo.

Over the weekend, National Novel Writing Month kicked a hornet’s nest. They posted a statement claiming that any categorical dismissal or criticism of the use of “AI” tools in writing is classist & ableist, and implies some unrecognized privilege on the part of the critic. That’s a strong stance, and merited a robust response, which the community was happy to provide. Many published authors (Chuck Wendig¹, Daniel José Older², and others) jumped in to point out how the tools on the market are built by destroying copyright, but I haven’t seen an angle that addresses the biases inherent in the LLMs & neural networks that are being used for most of these “AI” products. (I’m drafting this longhand & hate writing quote marks, so just imagine quotes around AI anywhere I write it. The tone is going to need to be inferred as well.)

Who the fuck am I to be talking about this subject, anyways? I’m some writer, what do I know about machine learning and artificial intelligence? For some background on me, I work in software for my day job on network protocols, I’ve dabbled in the underlying technologies of LLMs and their predecessors, and my degree included work in applied mathematics. I’ve got a working knowledge of the underlying techniques and research. LLMs in their current form are effectively Harvard Architectures with no NX bit, and that’s a terrifying thing to have decided to ship at a global scale. While some number of readers will have already decided I do not have sufficient background to have an informed opinion on this topic, let’s assume I’m knowledgeable in the topic and move on.

Technology had historically had a blind spot around bias. More broadly speaking, scientific enquiry has historically had a similar gap in its recognition. The relationship between the observer & observed, and what information an observer includes, inherently relies on the discriminating filter of the scientist!

There’s an entire novel on the topic of identifying what facts are worth selecting, Zen and the Art of Motorcycle Maintenance. Pirsig had a lot to say on Quality, and bias plays a role in defining Quality. To cut a very long discursus short, LLMs are built on data sets that have been filtered for specific stances, and have incorporated some level of bias in their training. This can be seen very literally in an adjacent form of GenAI: when prompting image generating tools like DALL-E or Stable Diffusion with ungendered and racially ambiguous terms, users still got biased output (skin tones and genders for given jobs and social statuses, for instance.) The companies building these tools have done a lot of work to try and suppress this behavior―by making substitutions to users’ prompts to force diversity… which effectively proves the point. They haven’t fixed the bias, they’ve papered over it.

To put it mildly, pervasive bias problems in AI tech have negative downstream impacts on their applications, especially in creative arts. The dearth of material from minority artists, writers, etc. in the training set―much like documented issues in other technical areas―skews their output. Pervasive stereotypes in older works of literature, about any minority or marginalized community, are amplified by the “training” techniques used to tune and build these systems. Like an automatic suspect identification system that can’t distinguish Black men, a system trained on decades of misogynistic, racist books (stolen from their authors, even!) will perpetuate those stereotypes in generated prose and editorial suggestions.

Personally, as a queer writer of queer fiction, I’m concerned what the prudishness of Silicon Valley would do when presented with my writing! What adjectives would it seek to erase, what aspects of my experience would it stifle as a statistical anomaly, under the guise of being an editorial tool?

This is hardly an idle question. Speech to text tools routinely obfuscate swear words, because obviously we need to be shielded from profanity. “Fuck!” gets rendered F*ck! or F***!―neither of which is what I said. Tools like Copilot will often refuse to discuss politics, anything about sex, and sometimes even routine queer social activities or figures! Censorship of minority experiences is hardly new; resources on sex ed and LGBTQ+ experiences is routinely filtered by so-called “porn filters” as “adult” content, even after objections by some of the youth orgs that were hit.

Somehow, we’re expected to believe that tools built by some of the same teams and under the same direction as the people who built biased services, to be fair and correct in their handling of queer issues with GenAI? NO!

Beyond the facile argument in favor of using AI tools, NaNoWriMo’s counter, that my categorical dismissal of the tools is classist, ableist, and indicative of my unrecognized privilege, is downright insulting. This claim twists the language of the oppressed to support a venture capital-funded company’s defense³, who are also funding NaNoWriMo. The claim is clumsy, which is probably its greatest insult: it rejects some of the original goals of the organization and the project, and this leaves it a hollow statement. It wasn’t even made as a blog post, it was put up on their ZenDesk/help pages!

To claim that rejecting AI tools is classist is to ignore the actual cost of using those tools. Paying for a professional license to any of the AI editing tools is a not insignificant cost, for the limited utility of that program―and not a person you can maintain long-term interactions with, building up rapport and shared context! Those tools, too, are built using data taken from users and other sources, generally without or with minimal compensation for their effort. Using them won’t build your writing community, it’ll impoverish it by depriving you and others of connections and directing resources out of your community and into the pockets of the funders of those projects, indirectly enriching their investors to your community’s detriment.

Ableism… I’m not qualified to fully break this one down, but relying on a statistical model is hardly the only way to write, and certainly isn’t the aid these writers necessarily need. This argument is an attempt to claim an entire community as shields for corporate greed, when they don’t want the help.

Lastly, let’s talk about privilege. Again. I’m definitely privileged: my parents are both college educated, I was born in a state which prioritized education & had early exposure to a lot of culture. I first tried my hand at a NaNo in high school, even! I’ve had years of time to develop my craft, and I have a well-paying job that leaves me with leisure time to write. None of that detracts from my criticism, however, and I reject that my privilege is blinding me to the cost or the inaccessibility of writing and editing. Yes, hiring an editor is an expense. I’m still waiting to do that for my first novel, in part due to the cost! But part of the point of NaNo is to learn the tools & techniques of the trade. Yes, editing is a lot like work… isn’t that the point?

Ultimately, the statement from NaNoWriMo feels like the organization is trying to throw their support behind their AI tooling partner, to defend their sponsor from criticism from the community that they had to understand was coming. Unfortunately, the stance they took was so extreme and so directly contrary to the attitude of their community that it was bound to explode. We, artists and writers, are sick of being told how to think in furtherance of commercial interests, and are hardly going to be quiet now. When we’re seeing beloved internet groups like the Internet Archive attacked for improving access to information during an unprecedented geopolitical event and health crisis⁴, this support of commercial interests who are seeking to profit off of broad social goods & communical resources was definitely going to land poorly.

I hope the cash was worth it.