I follow the writing gossip across the internet, or at least on Substack, WordPress, and Bluesky. All of these platforms have the classic downsides of social media discussions: they are full of drama and heated opinions, everyone thinks a little too highly of themselves, and topics tend to blow up and fade quickly. However, AI is taking up a lot of space in the conversation, and it doesn’t seem likely to die down anytime soon.
There is always the risk of groupthink, but I see an overwhelmingly negative attitude toward AI-generated writing, especially in fiction. This is across pretty much every axis: amateur and pro authors, editors, agents, publishers, and readers. It mirrors the steadily declining sentiment toward LLMs in the general population, but is notably more extreme and vociferous. A lot of people do not want AI-written fiction. Almost every publication where I’m submitting short fiction has a “no AI” clause in their submission guidelines and contracts.
Fiction is big business in some ways, but most people at ground level are still in it for the love of human stories. Even for more cynical publishers, the prevailing opinions make accepting AI-generated work a huge reputational risk. We have already seen scandals and backlash and seven-figure deals exploded, and there will no doubt be more.
AI is affecting even those authors who aren’t using it. Not only does it clog the slush piles at publications everywhere. Publishers are incentivized (intrinsically or extrinsically) to identify and reject AI-generated work—to give their limited slots to real people, and to avoid the bad reputation. And authors now have a brand new worry: what if my totally human work is flagged incorrectly as AI?
The Detection Problem
The current state of business is that most AI detection is done by editorial eyeballing and tech-powered AI detectors, and both of these methods are fallible. Just about everyone agrees that detector tools give too many false positives and false negatives to be conclusive. Some people still maintain that well-trained humans can recognize AI, but this has not generally been borne out by rigorous experiments, and different models have different foibles.
As usual, the EU is the first mover on regulation and has recently mandated more stringent labeling of AI generated output.
Anthropic announced that it will support this by including a form of watermarking on the content generated by its popular Claude models. Other major providers are likely to follow suit.
While many people have speculated that these watermarks would be as simple (and useless) as hidden characters, they are actually more interesting and harder to circumvent. The models weight their word choices in such a way that the raw text itself can be used for detection. Since this is statistical weighting, it is unclear is how effective it will be for small pieces of text or heavily mixed human and AI output.
It has also been pointed out that this kind of watermark is useful to AI companies regardless of regulation. It can help them identify AI text in the now heavily polluted internet content that they scrape to feed their models, avoiding the Habsburg AI loop.
Regardless of the efficacy of these kinds of watermarks, most LLM models do not currently have them. Users who want to avoid detection can simply move to unmarked models. This all implies that we can’t rely on any kind of known detection to sort out what is or isn’t AI.
So some authors and companies are exploring the opposite approach: proving the human providence of text.
Proof of Humanity
Bona Books recently published the excellent article, The Machines are Coming for your Masthead – Small Press Publishing in the Age of AI.
As a small indie publisher, they developed their system for dealing with AI submissions when confronted with the possibility that they had already accepted AI generated work for publication, despite a clear no AI policy. Their solution hinges partly on human relationships: talking to the author who submitted the work in an effort to ascertain whether the person really had a deep understanding of what they had submitted. This seems like a good way to weed out those looking for a quick buck (if such a thing is even possible in indie publishing). But it’s effort-intensive and tough to scale. It also becomes complicated for work where the author treats AI more as a collaborator or editor.
There are upsides and downsides to an approach like this. The positive aspect is that it encourages truly close collaboration between publisher and author. However, there is also a deeply unfortunate distrust inherent in the process. The publisher has to assume some possibility that the author has misled them, and the author has to accept a level of distrust and scrutiny that may be uncomfortable, especially in a notoriously introverted profession.
Author Vera Kurian has recently proposed a different method of verification: cryptographically verified provenance for a work of art. She doesn’t necessarily suggest this is the be-all end-all solution. Her articles on the subject are more of an eloquent and irritable throwing hands up, saying “if nobody else is going to solve this, I’ll try something.”
Kurian’s homemade tool is designed to be used at every step of the writing process, tracking changes in documents and storing cryptographic signatures to match archived copies. It builds on some ideas popularized by cryptocurrency, and it is (perhaps ironically) a vibe-coded solution to an AI problem, but it’s the kind of idea that makes some sense to the part of me who has been building software for half my life.
However, as an author, it’s a depressing solution. It puts the onus entirely on the writer to prove they are not a bad actor, and it’s fairly inconvenient. I use two phone apps and at least three programs on PC for different aspects of my writing. Heck, I use pen and paper sometimes. Trying to build an authoritative record of all this sounds like a bad time, and a significant distraction from the already hard work of trying to write something people want to read.
It also relies on faking the paper trail being onerous, not impossible. This is the sort of thing that works when it’s not popular. Unfortunately, if it were a tool that everyone adopted, there would then be an incentive to circumvent it, and clever automation could probably do so.
Even More Tools
Kurian is far from the only one considering solutions an author can use to protect their reputation. Writing tools with history features—like Scrivener and Google Docs—often come up in these conversations. To me, these are in the same class as Kurian’s solution, offering weaker, more easily fakeable proof but more ease of use (especially since many authors already use them daily).
There are newer contenders in the space as well.
Encribe is a word processor and “live streaming” writing platform in early development, whose strategy is to train AI to recognize natural human writing patterns in order to verify the authenticity of work by “proving” that it was written by a human. Unfortunately, in my limited interactions with the tool it has been pretty hit or miss on recognizing my input accurately.
Ellipsus is a word processor for writers, in much the same vein as Scrivener, and they recently announced a “plus” option that includes a feature called Emboss, packaging writing stats and snapshots specifically to use as a proof of authorship tool.
At this point, I think all of these tools have a potential fakeability problem. They aren’t popular enough to impact the people trying to sneak AI writing into publications. As soon as any of them reach that threshold, there will be an incentive for adversarial attacks, and this is exactly how software security arms races start.
However, it’s a big “if” to suggest that any of these tools can even reach that level of popularity. That would require a lot of authors shelling out subscriptions to a particular company, and a lot of publishers signaling that this kind of proof of provenance is something they want. The people making these tools see an opportunity in their market. Time will tell how big it is.
Silver Linings
If there’s anything we can take from the current state of play, it’s that a lot of readers, writers, and publishers want AI-generated content out of their fiction. But a lot of AI-generated submissions are being sent in anyway. Some are being caught. Some are certainly getting through. And if there haven’t been false accusations against authors yet, there certainly will be.
There are no clean or easy solutions today, and I don’t see any on the horizon either. It’s always possible that some tech breakthrough will make AI detection reliable, but I wouldn’t count on it. It’s far more likely that we’ll continue collectively muddling through, at least for the next few years.
However, I think there are some silver linings. In an age when novels and stories have to compete with all sorts of other media, where art has to battle with monetized, algorithmic “content,” and people regularly claim that literature is dead, it’s clear that there are still a ton of people who really care about human stories. As long as that’s the case, it seems inevitable that we’ll find ways to continue sharing our art.
That said, it’d still be nice to have the chance to get paid for it too.
Words Deferred 