This issue came up in a User Fiction thread and Squishy suggested making a post in Staff Communications about it, so I figured I would bite the bullet.
It's, well, exactly what it says on the tin. From an ethical standpoint, both in terms of its training on plagiarized material and the recently revealed environmental impact, I would propose Sufficient Velocity institute a ban on content created via ChatGPT, Midjourney, or any other LLM/"AI" models. I see them as an affront to creatives and actively harmful on multiple levels.
What is the environmental impact? Sorry, I'm profoundly uninformed on this topic in general- I gave up trying to follow the technical side because I just don't have the background to properly judge whether what the learning alogorithm does really 'counts' as plagerisim, but this is the first time I've heard environmentalism brought up around the topic of AI content.
What is the environmental impact? Sorry, I'm profoundly uninformed on this topic in general- I gave up trying to follow the technical side because I just don't have the background to properly judge whether what the learning alogorithm does really 'counts' as plagerisim, but this is the first time I've heard environmentalism brought up around the topic of AI content.
Personally I don't see any problem with allowing such content if people want to post it and people want to read it. Banning content so generated from being posted on this specific website won't really put a dent in its actual generation, and thus the resource use, I don't think. I could be wrong?
I don't have much particular insight or expertise to bring to a debate on the topic. It simply seems worth saying that, as a creative on this website, I personally feel pretty unaffronted by AI-generated text and images. Since, presumably, the split of attitudes overall in the userbase would be relevant data to take into account for any potential decision, yeah?
The use of AI on the platform is a topic that can be both exciting and worrisome. It's hard to decide a policy for it as all sides are controversial or unrealistically idealistic; laws are not set yet; and the development of AI is not done yet either – tomorrow there might be something new to consider.
But, we feel the need to provide an AI-text policy at least to start with, to allow users to navigate the options. However, please note that this is a dynamic policy and as the world around us changes, we are going to adapt to it rather than fight it.
On one side, we want our authors to be able to use any tool they want. On the other side, the readers will decide if they like or dislike a certain type of content. After all, we have great lists that filter out what the majority of the readers do not want to read. However, it's going to be important to provide transparency in order to allow readers to decide quickly.
This policy will get into some rules about the use of text-AI, and where we stand as a platform.
But first, we are going to clarify how AI is used, down to a very basic form, and then our different outlooks on them.
In this policy, we are going to separate the use of AI for text, into 3 categories.
General Assistive Technologies
AI-Assisted
AI-Generated
1. General Assistive Technologies
This is where a tool takes a sentence and proofreads it. From spell-checking to suggesting a slightly different word or structure. This is the most common type that is built into (or is being built into) writing programs.
For example, this falls under the same concept as what Grammarly provides (including Grammarly Go). Where it takes the text that you wrote, suggests a couple of changes, but keeps the author's voice present.
For a more detailed example: I used the following text in Grammarly:
I want to use this text as an example of the different services provided by AI enhancement, ai assistance, and ai generated text. I'm writing this text without knowing what the results will be, but I hope that each of the different tools I'm going to use will give me something that is enough to explain my intention.
This was changed to the following text after I agreed to all the suggestions I received:
I want to use this text as an example of the different services provided by AI enhancement, ai assistance and ai generated text. I'm writing this text without knowing the results, but I hope that each of the different tools I will use will give me enough to explain my intention.
My writing and my voice remained very consistent. Which is why we want to clarify that this type of AI enhancement, or general assistive technology, does not fall under this policy. This will continue to be something that all users on the platform can use, without any sort of restriction or tagging.
In fact, we encourage using the Grammarly plugin.
2. AI Assistance
We are defining this level of using an AI tool as where it's the author's creativity and structure but it uses the AI's voice and tone.
To put it simply, where you take a paragraph that is already written and ask an AI to improve it for you.
As an example, using the same text as before, but asking ChatGPT to
Please refine the following:
"I want to use this text as an example of the different services provided by AI enhancement, ai assistance, and ai generated text. I'm writing this text without knowing what the results will be, but I hope that each of the different tools I'm going to use will give me something that is enough to explain my intention."
I would like to utilize this text as an illustration of the diverse range of services offered by AI enhancement, AI assistance, and AI-generated text. I am composing this text without prior knowledge of the outcomes, but I am optimistic that each of the distinct tools I am employing will produce content that sufficiently elucidates my intention.
This result keeps the same idea and structure, however, the voice and the words of the original author are gone.
The Rule:
We do not view this as either good or bad. However, the writing or reading process, which is unique to the author's voice, will now be somewhat uniform across all stories that use AI-assisted writing. So we ask that if you use this type of enhancement excessively – for example as your personal proof-reader for every sentence – to please tag your story as "AI-assisted".
Note that this allowance is only for authors within their chapter pages. We simply wish to provide the choice, but we do not want user interactions such as comments, reviews, forum posts, or private messages to lose the human voice.
Update: For novels that contain a negligible amount of AI-generated snippets, you may tag them "AI-assisted" rather than AI-generated content. But, please add a warning in the author notes on the chapter with the AI-Generated content.
3. AI-Generated
We are defining this level of using an AI tool as where the author is prompting the AI to write the content and is merely directing it. Where the author is assisting the AI rather than the AI assisting the author.
As an example,
Prompt: "Please write a text that I can use as an example, to test the results by different AI tool types"\
I am seeking to obtain a sample text that can serve as a practical illustration for evaluating the outcomes generated by various types of AI tools. By employing these diverse AI tool types, I aim to gauge their effectiveness and compare the outputs they produce. With an open mindset, I approach this task, uncertain of the specific results each tool will yield. Nonetheless, I anticipate that this experimental exercise will provide valuable insights into the capabilities and limitations of each AI tool, facilitating a comprehensive evaluation of their performance.
In this case, we feel that the content is not being authored by the author. However, we are restricted in what we can do about it. We do not want to start a witch hunt and assume an author is lying when they say that their text is not AI-generated; there have been enough examples in the wild to show how this would be a bad idea.
Until now, we have attempted to ban stories that are mainly written by AI but directed by the user. However, it's clear that automated systems that detect AI in text can not detect it correctly 100% of the time. Instead, according to research done by Scribbr, even the most sophisticated commercial software is only 84% accurate, with others being even less so. They will also become less accurate with time, as AI generators are being improved adversarially by using those systems to see what is considered AI and what is considered human-created content; not to mention the continued advances in the field including dozens of new models each month and the existence of LoRAs (lightweight modifications that may change the tone or style of the result, such as a sarcasm LoRA).
Based on that, we do not want to include rules that we can not reliably enforce. And while it's still possible to detect AI content, even with the human eye, in this case, we believe that AI will only get harder to detect. Which is why we will allow this type of text as well.
While you do not need to mention any 'back-office' assistance by an AI, such as brainstorming or research, you do need to mention it when you copy the text. When you do, please keep in mind the following rules and restrictions:
The Rules:
1. A high quality must be retained. AI-generated content must be moderated and refined by the author to ensure quality, continuity, and readability. Avoid any low-effort text generation.
2. Reviews, comments, forum posts, or any type of user interaction in general can not include AI-generated content unless it's a quote.
3. The AI-generated content can not violate any laws and can not violate our rules and guidelines. As this is a new topic, the laws might change in the future - any use of AI-generated content is at your own risk.
4. You must tag your story as "AI-Generated". But, for now, even if the story is not tagged as one, readers may say it is one in their reviews.
The new tags are as such:
AI-Assisted
Updated: The author has used an AI tool for editing or proofreading. The story thus reflects the author's creativity and structure, but it may use the AI's voice and tone. There may be some negligible amount of snippets generated by AI.
AI-Generated
The story was generated using an AI tool; the author prompted and directed the process, and edited the result.
5. If the AI is used for translating a fiction, the novel must be read and proofread by someone who knows the language to guarantee the accuracy.
(Update: No tag is required for this, as long as a human goes over the translation. As this is likely going to be the new translation standard.
If no human reads, proofreads, and edits the translation, then it must be tagged as "AI-assisted" as it will retain the AI's tone, and voice.)
In essence, they propose that there's degrees of AI and request that authors tag their fics appropriately. They also acknowledge that it's unrealistic to attempt to ban it and trying will result in harm if/when non-AI works are incorrectly flagged.
Considering the last time a thread about this was raised less than a year ago it went for 17 pages before being locked with an answer of 'it is being moderated on a case by case basis', I suspect that we're not going to see any change in the stance of the directors since then.
I've used AI assistance in my writing for the last... six or seven years, approximately. I don't believe anyone would be able to tell if I didn't tell them, but I'm also not the sort to lie. Not to mention that Squishy and Xon already know, so if such a ban were to be created, I'd obviously need to leave the forum.
Just speaking for myself, I'm getting tired of relitigating this every couple of months. It might be easier to give up on writing and just go play computer games for the rest of eternity, but I have stories I want to write, dangit. This is where my readers are. Shards of a Broken Sun is way too bizarre a piece of sci-fi to bother posting it elsewhere.
...I can feel myself getting stressed again. I think I'll stay out of this thread, and just hope people don't do something insane.
Yeah, I can see ground for recommending that AI assistance be tagged but...really who's enforcing it without first generating a lot of capacity for abusing censorship?
This issue came up in a User Fiction thread and Squishy suggested making a post in Staff Communications about it, so I figured I would bite the bullet.
It's, well, exactly what it says on the tin. From an ethical standpoint, both in terms of its training on plagiarized material and the recently revealed environmental impact, I would propose Sufficient Velocity institute a ban on content created via ChatGPT, Midjourney, or any other LLM/"AI" models. I see them as an affront to creatives and actively harmful on multiple levels.
I rather enjoy AI-generated content. AI art reminds me of dreaming, for example. Everything both makes sense on the surface but the more you look at it the less it makes sense. All the details are wrong. The writing isn't real words, hands have extra and weird fingers, patterns are nonsensical. And the generation method is similar, your brain semi-randomly combining previous experiences and ideas. Using ChatGPT is similar from a written instead of a visual perspective, though written words are more constrained in their communication compared to visual art. So I think removing it also removes part of our ability to look at our own subconscious we haven't been able to really examine previously.
This claim struck me as physically nonsensical, since it seemed to me that these numbers would only make sense if they were running tap water thru a cooler and directly into a drain, and the article didn't go into specifics, so I tracked down the actual paper. Here it is.
This result is based on losses arising from the use of cooling towers, which expose their working fluid to air, and so have evaporative losses. (This is partly a good thing because it means evaporation cooling.) So it seems that I'd underestimated the amount of consumed energy.
However, they are also including secondary losses from power plants that supply the electricity, which in many of their analyzed cases is many times larger than on-site consumption. The paper itself is completely straight-forward about this, but it seems to be getting reported on in a way that doesn't mention this point.
This issue came up in a User Fiction thread and Squishy suggested making a post in Staff Communications about it, so I figured I would bite the bullet.
It's, well, exactly what it says on the tin. From an ethical standpoint, both in terms of its training on plagiarized material and the recently revealed environmental impact, I would propose Sufficient Velocity institute a ban on content created via ChatGPT, Midjourney, or any other LLM/"AI" models. I see them as an affront to creatives and actively harmful on multiple levels.
I think asking people to tag it so it can be avoided by people who want nothing to do with it is all that can reasonably be expected. I hate AI tech and everything about it, I wish it hadn't been invented, and I wish people could be prohibited from using anything but public domain or explicitly donated works on AI, but sweeping bans like this are just going to hurt people who aren't using AI and just sort of seem like it, and people who wanna post dreamy weird AI stuff. If someone is caught plagiarizing directly, hammer that like you would any other plagiarist, emphasize tagging AI stuff so it can be filtered, but more than that will just hurt people.
Yeah I'm pretty firmly in the anti-AI content camp, but it's basically impossible to enforce and I think strict labels/rules on how it's deployed is the best course of action. If you use AI to assist writing your fic/quest/whatever, tell people up front - it's honest and it also opens you up to more constructive criticism since people can now give you advice on what prompts/whatever to use. If you use an AI generated picture, label them the same way you would attribute an image you scraped off Google image search or DeviantArt or something.
I can't begin to imagine you'd ever get decent writing from anything you need to prompt. Actual AI assistance (for writing, anyhow) is a mixture of feeding it story text to complete—uni or bidirectional generation—attention hacks, LoRA or token biasing schemes, or occasionally something as simple as injecting fake sentences in the context window to make it think it's trying to do something other than what it is, plus of course a metric ton of editing.
Unless you count the story-as-it-exists as the 'prompt', I suppose.
Instructional models are absolutely awful at this. RLHF and instructional fine-tuning is, by and large, directly opposed to good quality output.
Don't be. You found something cool, and wanted to show it off. There's no way you could have known the sort of acid bath you were jumping into.
Though for what it's worth, I'm serious when I say you'll never get good results with that sort of generator. If you want to change that, you should probably look at NovelAI instead. Yes, it'll ask you to do a lot more of your own work—it provides AI assistance, not AI writing—but that's also the only way you can get better at writing.
I don't think anyone reasonable thinks you intended harm. AI has a lot of issues and a lot of divisions. For my part, I hope the tone of my post doesn't suggest any animosity for you or your thread. I dislike what the big AI companies are doing and think it's a huge problem, but also, bluntly, I don't think it was very cool of the OP to do a callout thread like this, under the circumstances. A lot of people are pissed at the AI tech companies, but being mad about something else and charging recklessly into sweeping statements or callouts is a great way to hurt people who have nothing to do with what is actually the thing upsetting you.
Speaking as someone who has done that sort of thing a lot in the past and horribly regrets it. Just, seriously. You didn't do anything wrong, and you don't need to beat yourself up over this. It helps no one if this ruins your enthusiasm for writing or leaves you upset and guilty like you actually hurt people. You didn't hurt anyone, you just stumbled into a controversial new technology and were an example for people worried about what actual harm might end up happening to jump off on to propose a rules change. I hope you write your fic and have fun, even if you end up using an AI assistant. I hope you'll learn to write yourself, because learning to write yourself is really rewarding, but even if you don't, in the end, you're one of many people playing with a new toy. That's not something you deserve to feel awful about.
Don't be. You found something cool, and wanted to show it off. There's no way you could have known the sort of acid bath you were jumping into.
Though for what it's worth, I'm serious when I say you'll never get good results with that sort of generator. If you want to change that, you should probably look at NovelAI instead. Yes, it'll ask you to do a lot more of your own work—it provides AI assistance, not AI writing—but that's also the only way you can get better at writing.
The goal should still ultimately be to wean off AI entirely. Been writing fiction for almost 20 years and sportswriting professionally for over a decade and I'd be a fraction of what I am today without all the times I tore my hair out looking for the right word.
I have decided to discontinue the fic and work on a different fic idea that I have been developing before I discovered Character.ai and decided to enlist it for writing purposes. And I am very thankful for everyone's thoughts and opinions on the subject.
The goal should still ultimately be to wean off AI entirely. Been writing fiction for almost 20 years and sportswriting professionally for over a decade and I'd be a fraction of what I am today without all the times I tore my hair out looking for the right word.
The most common use I get from novelai is just... to get something written down, which I can then rewrite. Call it a quirk, but I find it much easier to see the flaws in a piece of text than in a blank page.
Besides that, funnily enough, I've used it to get better at english. If I write a paragraph, erase the middle bit, and repeatedly ask the AI to rewrite it... half the time it does something nonsensical, a quarter of the time it writes something worse than my own writing, but sometimes it invents turns of phrase that I really like. Which I can then make use of later, next time something similar comes up.
It's nowhere near as good as having a cowriter would be, of course, but that's not always an option.
The most common use I get from novelai is just... to get something written down, which I can then rewrite. Call it a quirk, but I find it much easier to see the flaws in a piece of text than in a blank page.
Besides that, funnily enough, I've used it to get better at english. If I write a paragraph, erase the middle bit, and repeatedly ask the AI to rewrite it... half the time it does something nonsensical, a quarter of the time it writes something worse than my own writing, but sometimes it invents turns of phrase that I really like. Which I can then make use of later, next time something similar comes up.
It's nowhere near as good as having a cowriter would be, of course, but that's not always an option.
"I have a trick that makes things easier for me. Since writing is very hard and rewriting is comparatively easy and rather fun, I always write my scripts all the way through as fast as I can, the first day, if possible, putting in crap jokes and pattern dialogue—"Homer, I don't want you to do that." "Then I won't do it." Then the next day, when I get up, the script's been written. It's lousy, but it's a script. The hard part is done. It's like a crappy little elf has snuck into my office and badly done all my work for me, and then left with a tip of his crappy hat. All I have to do from that point on is fix it. So I've taken a very hard job, writing, and turned it into an easy one, rewriting, overnight. I advise all writers to do their scripts and other writing this way. And be sure to send me a small royalty every time you do it."
Bespoke crap jokes and pattern dialogue will always be better than store-bought.
Same with the turns of phrase. It's never going to get easier for you if your solution to "this doesn't feel quite right" is to slap it into the plagiarism machine and press the big red button until it monkeys-with-typewriters its way into emulating Terry Pratchett. Writing more and reading more will always, always be better for your skills in the long run.
Not to interrupt the AI circlehug or anything but...
If you use AI, for text or images or whatever else it does nowadays, it should be required to be tagged as such. I'm not sure you could meaningfully enforce that outside of the really, really obvious stuff, the stakes for fanfic and such just aren't high enough for me to try and figure out how, and by the same token I don't see any real reason for someone NOT to disclose it.
That way people who like AI stuff can have their AI stuff and everyone else is saved a few minutes per thread.
