AI and Data Scraping on the Archive

Post Header

Published:

2023-05-13 21:58:58 UTC

Translations:

Tags:

With the proliferation of AI tools in recent months, many fans have voiced concerns regarding data scraping and AI-generated works, and how these developments can affect AO3. We share your concerns. We'd like to share what we've been doing to combat data scraping and what our current policies on the subject of AI are.

Data scraping and AO3 fanworks

We've put in place certain technical measures to hinder large-scale data scraping on AO3, such as rate limiting, and we're constantly monitoring our traffic for signs of abusive data collection. We do not make exceptions for researchers or those wishing to create datasets. However, we don't have a policy against responsible data collection — such as those done by academic researchers, fans backing up works to Wayback Machine or Google's search indexing. Putting systems in place that attempt to block all scraping would be difficult or impossible without also blocking legitimate uses of the site.

With that said, it is an unfortunate reality that anything that is publicly available online can be used for reasons other than its initial intended purposes. In many cases, AI data collection traffic relies on the same techniques as the legitimate use cases above.

Once we became aware that data from AO3 was being included in the Common Crawl dataset — which is used to train AI such as ChatGPT — we put code in place in December 2022 requesting Common Crawl not scrape the Archive again.

We cannot go back in time to stop data collection that already occurred, or remove AO3's content from existing datasets, as much as we may dislike that it happened. All we can do is attempt to reduce such collection in the future. The Archive's development team will continue to be on the lookout for individual scrapers collecting AO3 data, and to take action as needed.

Likewise, our Legal committee has and will continue to serve the OTW mission of protecting fanworks from legal challenge and commercial exploitation. This includes their position that users should be allowed to opt out from having their works incorporated into AI training sets, a position that they have presented to the U.S. Copyright Office. They, too, will continue to keep pace with this developing field.

What can I do to avoid data scraping?

You may want to restrict your work to Archive users only. While this will not block every potential scraper, it should provide some protection against large-scale scraping.

AI-generated works and AO3 policies

At the moment, there is nothing in our Terms of Service that prohibits fanworks that are fully or partly generated with AI tools from being posted to the AO3, if they otherwise qualify as fanworks.

Our goals as an organization include maximum inclusivity of fanworks. This means not only the best fanworks, or the most popular fanworks, but all the fanworks that we can preserve. If fans are using AI to generate fanworks, then our current position is that this is also a type of work that is within our mandate to preserve.

Depending on the circumstances, AI-generated works could violate our anti-spam policies (e.g. if a creator posts a significant number in a short time). If you're uncertain whether a work violates our Terms of Service, you may always report it to our Policy & Abuse team using the link at the bottom of any page, and they can investigate.

This statement reflects AO3’s policy at the time of writing, as we wanted to be transparent with our users about what our current stance is and what can be done – and is being done – to mitigate scraping for AI datasets. However, these policies are also under discussion internally among AO3 volunteers. If we agree on changes to these in the future, those will be announced publicly; additionally, if there are any proposed changes to the AO3 Terms of Service, they will be made available for public comment as is required of any and all changes to our Terms of Service.

We hope that this helps to make things more clear – this is a complicated situation, and we’re doing our very best to address it in a way that doesn’t compromise AO3’s principles of maximum fanwork inclusivity or legitimate uses of the site. As discussions and approaches evolve, we will keep our users updated.

Actions

Comments

Sorry, this news post doesn't allow comments.

Pages Navigation

bonchCronch Tue 16 May 2023 01:20AM UTC

no ai generated things at all.
if you are going to let scraping happen make it an opt in not an opt out.
an opt out situation is predatory and designed to take advantage of site users.

Comment Actions
- Thread
1. Kereudio Tue 16 May 2023 03:04AM UTC
  
  They worded the bit about privating your works poorly: it is not an opt-anything system, the archive is not doing, or allowing, any scraping, they explicitly blocked the most common one (aptly named common crawl) that LLM's like chatgpt use to scrape data.
  Unfortunately, some scrapers, be they for AI purposes or not, will ignore that request, and there is no way to 100% prevent them from doing so: rule breakers will break the rules, etc. They suggested privating your fics if you're concerned about that, same way they suggest turning comments to users only if you're being trolled/harassed/spammed by anonymous users/bots. It is against the rules to do so, but there is no way to 100% prevent it.
  
  Comment Actions
  - Thread
  - Parent Thread
Hunter Nightblood (kah0922) Tue 16 May 2023 01:42AM UTC

Just as a PSA, AI story generation isn't a new phenomenon. In fact, you can find AI generated stories on this very site as early as 2019, probably even earlier. AIDungeon's been a thing since 2019, and GPT's been around since 2018.

Comment Actions
- Thread
Midnight_Stars_Moon Tue 16 May 2023 01:42AM UTC

AI-generated works should not count as stories by fan-writers, which is what this site is for. Using AI does not make a person a writer, but more like a programmer. It's like commissioning a book and then taking credit for the writer's work. Except AI is not a person.
This is very sad to learn. One of the fun parts of this website was reading stories by real people. I foresee myself reading less and commenting WAY less. There is no need to praise a machine.

Comment Actions
- Thread
MossyOak Tue 16 May 2023 01:45AM UTC

Thanks for the news!

Comment Actions
- Thread
Cloudsu Tue 16 May 2023 02:04AM UTC

Okay, but since we know that AI-generated fanworks are stealing from someone, like, some other writer, doesn't that go against OTW/AO3's general want to protect its creators? I mean you can't say in the terms that you're anti-plagiarism and then also in the same breath say you're okay with AI-generated works because it represents a new form of fanwork. I understand and am thankful for the code protections, but if you're allowing poisoned fruit into the site then you're not really protecting us fully or appreciating the actual work that your writers do on this site. Because AI-generated works represent something much larger than "maximum inclusivity," and goes against everything that OTW has stood for in terms of celebrating actual creators of fanworks.

Comment Actions
- Thread
Rainbow_Foxes Tue 16 May 2023 02:09AM UTC

I'd like to implore the Archive to reconsider their stance on AI generated works on the basis that they're plagiarized. We know, as was stated in the first half of this announcement, that the Archive was and will likely continue to be scraped for AI training — thus it's a fair assumption that any AI generated work posted to the Archive is in fact plagiarizing form other works on the Archive. This is a violation of AO3's own plagiarism policies.
This choice to allow AI generated works also shows a lack of understanding for the reasons why creators hate it and would like to see it banned — it's not just an act of theft, it is also actively devaluing the work of artists and writers alike and disrespecting the craft. AI has no place anywhere that claims to be a safe haven for creators, the Archive included.
If we can support the WGA and their work to severely limit AI in Hollywood writer's rooms, then I think it is hypocritical to allow AI works on the Archive, especially when efforts are being made to prevent AI from using the Archive to train.

Comment Actions
- Thread
1. Hunter Nightblood (kah0922) Tue 16 May 2023 02:21AM UTC
  
  It actually isn't a fair assumption at all. People are under the misconception that AI Generated works are a recent thing, but they aren't. You can find AI generated works on this very site as early as 2019 (possibly earlier), when AIDungeon was created. Can you say for sure that those fics are plagiarized from AO3 works?
  
  Last Edited Tue 16 May 2023 02:23AM UTC
  
  Comment Actions
  - Thread
  - Parent Thread
2. 1. Kereudio Tue 16 May 2023 03:19AM UTC
    
    AIDungeon uses GPT-2, which was trained on pages linked on Reddit posts that had at least 3 upvotes. Considering there is a few fanfiction subreddits, that absolutely do link to AO3 works, yes, it absolutely was trained on data from AO3 pre 2017, and therefore is absolutely plagiarized.
    Source: https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
    Relevant quote: "[...] we scraped all outbound links from Reddit, a social media platform, which received at least 3 karma. [...] The resulting dataset, WebText, contains the text subset of these 45 million links."
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. Kereudio Tue 16 May 2023 03:26AM UTC
      
      Like. I don't see how pointing out that AI and GPT aren't new is making matters any better. The reason it's a conversation now is because ChatGPT and other LLMs were so indiscriminate of their scraping, and so public and marketed that every major tech company jumped on the bandwagon, that it made everyone else aware of the problem, especially when people started losing their jobs to it. But GPT having been around for longer than the 2020s just means that unethical data scraping has been around for longer than most realize, which makes it worse, actually.
      
      Comment Actions
      - Thread
      - Parent Thread
    2. 1. stormyseasons Tue 16 May 2023 09:54AM UTC
        
        This. Quietly scraped and collated under the name of research if I remember some of those data collections correctly.
        And then, now, commercially exploited across the world, often to the detriment of others. What a chilling effect on the urge to create anything - and that's assuming one still has the physical resources to create anything.
        
        Comment Actions
        
        Thread
        
        Parent Thread
3. ofthesummer Wed 17 May 2023 01:19AM UTC
  
  If you don't want your work to be co-opted and overshadowed by people who can churn out hundreds of longfics in minutes...
  Or if you don't want the AO3 to get caught up in a costly legal battle on the side of AI, using OUR donations...
  Then write to the OTW and tell them to get rid of Betsy Rosenblatt. Every voice matters, so do the right thing!
  
  Comment Actions
  - Thread
  - Parent Thread
piedrafundamental Tue 16 May 2023 02:16AM UTC

I do not want to read AI-generated work; nonetheless I thank you all for your support in protecting existing works from AI scrapers.

Comment Actions
- Thread
Wirewolf Tue 16 May 2023 04:32AM UTC

Oh this explains some of the stuff when I was going through my subs. If I opened too many at once I'd get a blank page and some text (which I can't recall). Great job ^^ ( I went from over 100+ pages of subs to 11 so I can see why I wouldve been triggering something against the AI)

Comment Actions
- Thread
Nightfall_1409 Tue 16 May 2023 04:54AM UTC

I strongly disagree with calling AI generated anything a 'fan' work. Allowing it to exist side by side with works that actual human authors painstakingly wrote, especially /knowing/ that they were almost certainly written off the back of the data scraping of this very website is an egregious mistake.
The maximum inclusivity of fan works should stop just there. At /fan/ works. as in a fan made it. Not a computer off the back of the rest of us.
Fanfiction is not merely content to be consumed, they are not generated. They're written! By people who are fans! AI aren't fans and don't deserve a place on this platform. Nothing's stopping these people from generating these works but posting them? for what purpose? They're not made by a fan...they're generated. I understand that there's nothing stopping people from posting it if they make it and just not saying anything, but that's on their conscience. This- pretending that they're the same- is on the Archive's.

Comment Actions
- Thread
1. the_black_rose Tue 16 May 2023 05:41AM UTC
  
  Kinda makes you wonder why someone would bother using an AI to generate fanfiction - when they could generate something saleable for Vella or Radish, Webtoon, online magazines, a Substack... the list goes on. The US Copyright office is having to try to figure out where to draw lines; a traditional publisher just got caught / called out for using an AI-generated image on the cover. Also rife with copyright issues.
  I think people will vote with their engagement. If someone can write 100 fics in a month and doesn't get much engagement because people here don't want to read AI-generated content, they'll eventually understand they won't get whatever it is that they're after....
  Maybe as the detectors get better, the answer isn't whether the content is banned or permitted, but the site provides a "likelihood of AI content generation" score - and people can choose which types of content to engage with.
  I'm not a fan of how these things were trained. I do not support them taking content creators' jobs. I think there's an eventual logical fallacy (covered in a story by the Atlantic about the coming Textpocalypse), which we saw in a more confined scale with the Facebook AI chatbots that created their own language - when it was just the 2 of them conversing. And I agree with the many people who signed the letter asking for these things to be tabled while regulations and society tries to catch up with all the potential ramifications. But, once out of the bag....
  At the end of the day, we still have a say in what we consume and engage with - whether that's content on Ao3 or NetFlix or anywhere else.
  And I think we should be informed, just like companies have to put nutrition labels on packaged food, so should our content be labeled with who and what generated it.
  Only then can anyone say they're being transparent.
  
  Comment Actions
  - Thread
  - Parent Thread
2. 1. butterflyslinky Tue 16 May 2023 11:52AM UTC
    
    As others have pointed out, having the site decide what is or isn't AI generated wouldn't be feasible, as there would be a lot of false positives, especially on new or ESL fic writers. It would have to be a voluntary, self-identified tag, which many people would probably decline to use.
    I'm not a fan of AI-generated content as a concept, but I think the issue is being blown way out of proportion. The imaginary Internet clout someone might get from putting out a lot of fanfics isn't sufficient motivation for most people to stop writing for a hobby site.
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. the_black_rose Tue 16 May 2023 01:10PM UTC
      
      Well, it may not be feasible today. My point was “as the mechanisms get better”. Someone could be coding a plug-in that people could easily use - like the PDF maker - that could be integrated into the page. We shouldn’t necessarily assume that this isn’t feasible. May not be today. Could be in 6 months at the rate of technological change.
      Like I said, I don’t understand why people would bother. But, as it begins to invade and pervade into normal ways of working, it has the potential to become something else. And then what, people will turn it off when writing their fanfics? No. They’ll use it the same as they began to adopt pushing a “like” button or composing tweets in (the original) 140 characters back when all of that was brand new….
      Will we wake up tomorrow to a sudden deluge of generated by AI fanfic? Not likely. The people who are looking to capitalize on it today are looking for $$ payoffs.
      But, as we find day to day uses for the technology, where’s the line drawn for “generated” vs “human created” going to be?
      And whatever it is today isn’t the same as it will be in a year, 5, 10.
      
      Comment Actions
      - Thread
      - Parent Thread
  3. Nightfall_1409 Tue 16 May 2023 01:30PM UTC
    
    The line already exists, it just needs to deploy it. Anything created without the significant contribution of a human is not eligible for copyright. A prompt is not considered a 'significant contribution' to the end product. If it is wholly or mostly generated by a computer it will never be eligible for copyright.
    And frankly, my level of hatred on all levels to the technofetishists who decided the first thing to try and truly automate with algorithms were the creative spheres of art and writing... I deeply wish them all a very "history will hate you forever."
    I don't doubt that one of the stop gaps will be labelling AI works, but again, my central issue is whether or not it belongs here on this platform. Much in the same way that copyright cannot be granted to something without significant human contribution, I don't think that the classification of fan work should be granted either.
    Posting it here- the earnest love of a story website- feels like a pathetically sad act of desperation for attention combined with a lack of self confidence in one's own abilities. It does not belong here.
    
    Comment Actions
    - Thread
    - Parent Thread
  4. 1. the_black_rose Tue 16 May 2023 02:16PM UTC
      
      Agree with you. Believe me. I still think we'll generally find that the thing works for (the average human attempting to use it): search engine optimization, automation, and some coding - the way it is today, anyway. Who knows what tomorrow will bring....
      But, ultimately, using it to generate end work products isn't super useful without more advanced application and knowledge / coding / practice. And still, it can only mimic what its been 'trained' on, it can't create.
      It can, however, hallucinate. And that's a problem.
      
      Comment Actions
      - Thread
      - Parent Thread
    2. ofthesummer Wed 17 May 2023 01:20AM UTC
      
      If you don't want your work to be co-opted and overshadowed by people who can churn out hundreds of longfics in minutes...
      Or if you don't want the AO3 to get caught up in a costly legal battle on the side of AI, using OUR donations...
      Then write to the OTW and tell them to get rid of Betsy Rosenblatt. Every voice matters, so do the right thing!
      
      Comment Actions
      - Thread
      - Parent Thread
Sciamachy (ShadowedMoonlight) Tue 16 May 2023 06:20AM UTC

I support the efforts to protect the archive from scraping.
However, since I believe using scraped data to generate fiction is close to plagiarism, I am not in favour of entirely AI-generated fanworks.

Comment Actions
- Thread
1. ofthesummer Wed 17 May 2023 01:20AM UTC
  
  If you don't want your work to be co-opted and overshadowed by people who can churn out hundreds of longfics in minutes...
  Or if you don't want the AO3 to get caught up in a costly legal battle on the side of AI, using OUR donations...
  Then write to the OTW and tell them to get rid of Betsy Rosenblatt. Every voice matters, so do the right thing!
  
  Comment Actions
  - Thread
  - Parent Thread
Blueicyeyes Tue 16 May 2023 06:33AM UTC

You need certain topics the generate the fics. If you know nothing about fandom then this fic going basic and will not get the desired number of likes and hits.

Last Edited Tue 16 May 2023 06:33AM UTC

Comment Actions
- Thread
LeeMorrigan Tue 16 May 2023 06:36AM UTC

Thank you

Comment Actions
- Thread
Pistols_at_Dawn Tue 16 May 2023 08:26AM UTC

Do any of the people saying how much they hate this have a PRACTICAL suggestion other than the current rule?
There is no button that they can press to ban AI generated fic and only AI generated fic. Banning it in theory but not enforcing the rule just means it will be untagged and unfilterable.
"AI fic is bad and I dont want it on this site :(" may be true, and a reasonable thing to feel, but it is not a policy suggestion!!

Last Edited Tue 16 May 2023 08:27AM UTC

Comment Actions
- Thread
1. stormyseasons Tue 16 May 2023 10:02AM UTC
  
  Tag, if we must tolerate it. I would like to suggest that IF it has gone through substantial human action - e.g. first drafts or ides prompts or similar that the human writer worked on after that - then it's a fanwork. The human fan worked on it, thought about it, had their own interpretation and so on. Have that distinction in the TOS. Honour code system, for practical reasons.
  But the raw stuff the prompt got the 'generator' to 'make', that isn't a fanwork. Likely plagiarised too, as people experimenting with very specific prompts found they could get the various popular algorithms to practically recreate what they were trained on. And that should not be officially welcomed here, nor supported by the organisation that says it was, and hopefully still is, by fans, for fans.
  
  Comment Actions
  - Thread
  - Parent Thread
2. 1. Pistols_at_Dawn Tue 16 May 2023 11:27AM UTC
    
    sure, yea, its not really a fanwork, agree 👍 But what does "not officially welcomed" mean in actual action, in terms of what the site does about it? if someone posts ai generated content, tagged "made by ChatGPT", in your ideal world, would that work be removed? would it be allowed to stay, just "not officially allowed"?
    my concern is that if the tag gets works removed, people will still post the AI content, they just wont tag it as AI generated. if theyre going to be here (which they ARE-- again, theres no reliable way to remove them), i want them clearly labled!
    and if they DONT do anything about then... how is that different from just allowing it?
    
    Last Edited Tue 16 May 2023 11:37AM UTC
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. stormyseasons Tue 16 May 2023 12:40PM UTC
      
      In my ideal world, there wouldn't be such a work at all because the person would embrace the delight of writing for themselves. In my slightly less ideal world, fics that are 'ai generated' should be tagged as such, in the manner of the major archive warnings we now have.
      (I would not personally read such a thing, I am interested in what a fellow fan might write, not what an algorithm smashed together based on whatever it has been programmed to output.)
      Even more ideally, I would like it to be strongly encouraged that only writing pieces which posters are willing to vouch or otherwise state have been substantially worked on to be allowed. Raw chatgpt or similar output would not be. Raw 'generated' work which happens to be found that is highly similar to a posted work is possibly plagiarism and can be dealt with that way.
      But to be honest, my beef is this - it is not pragmatically possible to remove - too many posted works, too few volunteers better tasked elsewhere. In any case, there are significant parts of fandom who do not want such works considered parts of fanwork anyway, for a wide variety of reasons included but not limited to plagiarism concerns, not wishing to contribute to commercial exploitation, and so on.
      But ideologically, I do not wish AO3 and the OTW to lock themselves - and by extension, every fan that uses their services but who hasn't the resources left to migrate to yet another service again - into defending 'AI' algorithms being used to carry out even more commercial exploitation of people for the benefit of a small group of private billionaires, 'techbros', or whatever term is appropriate to describe here. Which is what the OTW may find itself led to defend, if it decides to defend the computational processes of such algorithms as inherently transformative.
      I have not personally seen any flood of 'ai generated' fanwork on ao3, nor do I think there will be. The hustlers are crowding the places where one might be paid for writing, following dreams of wealth and choking out others who live by the words they write, or art they make. But that is not the true danger. The true risk, is buying into the notion that there is nothing we can do but meekly submit to being exploited utterly, to have our creativity redefined and misused by those who think only of their private profit, and blithely claim that any jobs or livelihoods lost can be regained through development of new industries. (Who will be left to buy or use those things, by that time?) And for the OTW and the Archive it tries to run for fans, by fans? The danger is of drifting from its purpose, of abetting those who would commercially exploit fans, defending the means by which they would exploit the world in the belief that their methods, their algorithms, are inherently transformative.
      
      Comment Actions
      - Thread
      - Parent Thread
    2. 1. Sherloqued Tue 16 May 2023 02:14PM UTC
        
        This is such a great discussion, so many great comments here.
        I'm not thrilled about what's happened, but I could accept tagging. I don't consider grammar tools, visual aids, etc. to be the problem. I'm certainly not against AI in and of itself. It's the flat out lifting text, maybe changing a word here or there, and calling it original. It doesn't seem like much now, but we know that technology improves by leaps and bounds in a very short time. People who want to take part in helping to 'teach' language and the fine tuning of it is an admirable thing. But there's also the other side of the coin too.
        Well, at least we're aware now - taking steps to minimize exposure to scraping and tagging your AI work I guess is the best we can do.
        
        Comment Actions
        
        Thread
        
        Parent Thread
      2. KrakenRiderEmma Thu 18 May 2023 03:57PM UTC
        
        Since you've made some thoughtful posts about how to address this issue practically, I'd love your thoughts on a dilemma I'm facing. If I've made works that start off with a lot of AI-generated text (say, 50%) and then "have gone through substantial human action," as you put it, and been transformed (by human writing and editing) into a majority-human work... should those works be tagged as AI-generated, along with the "made by ChatGPT" outputs that were produced with very little human effort in a minute? Or do they need different tags?
        
        Comment Actions
        
        Thread
        
        Parent Thread
      3. stormyseasons Fri 19 May 2023 01:40AM UTC
        
        From the sound of it, for whatever reason you have... first drafts and fic prompt ideas as generated via the algorithm? And then you developed and edited the drafts yourself? Then I feel it wouldn't warrant a 'ai- generated' or 'made by chatgpt' tag. Perhaps more of hmm.... 'beginning prompt was ai-generated' or similar.
        Because, well, for whatever reason you have utilised the algorithms to churn out something of a starter for you, and then you developed it the rest of the way. You did the work, as some people would say. [ I personally feel that it's more tiresome to edit and shape something with that kind of start, and it might even make finding your own writing voice harder because of how the algorithms just present the most likely options for your query as derived from their data set. The most common answer, as it were. ]
        
        Comment Actions
        
        Thread
        
        Parent Thread
Deuterosis Tue 16 May 2023 10:10AM UTC
Since I jumped in thoughtlessly at first, I felt I should clarify my thoughts on AI (or rather, SALAMIs).
Regarding whether SALAMIs steal: while generations are supposedly new texts based on probabilities and not cobbling pieces of the works trained on, another user here mentioned seeing generations that were "clearly lifted from erotic Attack on Titan fanfic". The argument is probably that this is still the result of probabilities from training on many of those fics. I would like to see the generations myself, but they probably weren't kept.
If at least some SALAMIs are actually cobbling constructions of fics outright, then those ones are not merely making probable texts. Given generators that aren't open-source, it would be hard to know which one they're really doing in gray cases; the creator's word is the only thing that "tells us". (And cobbling would certainly be easier to program and get running as intended.) For that reason, I don't think wariness on that matter is necessarily a position of ignorance in light of constant corporate exploitation.

Regarding the datasets, it is true that ethically-sourced SALAMI is not objectionable based on theft or plagiarism. But I still see a tremendous difference between the act of planning out and writing a story by hand and using SALAMI to generate one, regardless of the finessing involved unless the results are dramatically rewritten. (Even then, it's the SALAMI's output's beats that are being followed; not what the fan might actually have invented alone.)
It's hard to figure out how to explain why, but I'll try: Even if a human makes the prompt and massages the output, the results from the SALAMI are really unlikely to match what that person would have come up with on their own.
(This is the strongest reason I wouldn't agree that a fan "made" a SALAMI work. They got the machine to make something based on their idea, which is a different thing and more, as I've said, a form of programming. While getting cohesion out of a SALAMI is not a simple task, it is still not an identical task; without heavy edits the user needs do very minimal actual writing. It's like generating a fractal image, which is very pretty but not the same act as drawing one.)
The mere fact of it generating "likely text" seems to preclude genuinely surprising developments in the plot, sentence construction, and use of simile and metaphor. I know there've been many times the story I intended to write and the story I actually wrote are very different, because I had a new, completely unconnected idea that worked well enough to fit in. Sometimes the entire story is transformed into a new, different work from the original intent as I progress and the story shapes in unexpected ways. Sometimes my first go at a bridge was a bad one and I discover a new, better way of expressing what I wanted, that also mutates the work around to be richer.
SALAMIs would have to simulate this completely unpredictable process by managing to select otherwise-unrelated concepts that would work with the progression or wording it is generating. I am unsure they could possibly become that sophisticated; by definition, the process I described isn't rational or predictable.

Regarding minor examples like spell-checking, machine translation, and summary generation:
1. I don't see how the first can be classed as SALAMI generation; it's really just a computerized auto-dictionary.
2. Hmm. If the output only comes from the author's fic and the languages' syntax rules, nothing has technically been created that wasn't there already as it's conveying the author's own words into their equivalents in another tongue. But it does cut out the need for a human translator and I'm not sure how well it translates idioms, particularly those that don't translate.
3. This type of thing has existed for a very long time, although I'm not sure if the types of generation are different; the summarizers I'm thinking of simply took sentences from the work and concatenated them. Because it's only the summary, I am feeling it's different from generating a full text but I'm not sure that's a reasonable position.
Ultimately, whether it's feasible to ban generated texts is a separate question from whether they should count as fanworks. As I stated above, I believe they don't. A pragmatic decision on this matter is different from an ideological one that states there is no difference between hand- (or voice-)writing, however fast and simple for the author, and partially-edited generated text from a prompt.
An additional thought: I wonder where Mad Libs fits into this?
Comment Actions
- Thread
1. stormyseasons Tue 16 May 2023 10:21AM UTC
  
  About your machine translation point... poorly. It does nuance and idiom and other idiosyncratic bits of language poorly.
  I read faster in English than Chinese, and sometimes on AO3 there's an absolutely lovely Chinese fic I'd like to read faster or to recommend to someone who doesn't read Chinese, so I copy some text to say, DeepL to see how intelligible it is... and it fares poorly. Names, idioms, figures of speech, and anything subtle... no. (If you just need an extremely literal rendering, it's tolerable. If you don’t mind spending a lot of time editing the output to match the original language's tone, precision, beauty, or authorial voice... then it's a tolerable starting point.)
  And that is for two languages which though quite different in grammar and syntax and vocabulary, are spoken and used by enormous numbers of people. Plenty of reference corpora... but still not that good. And doesn't keep up with slang, references, and changes in dialect abd language.
  For fun I sometimes put English sentences through DeepL to change to German, to ask my friend who speaks that language if it makes sense. Closer langauage... you'd think it would work, but the result, she told me, is often very literal, stilted, and not how an actual human would use words.
  
  Comment Actions
  - Thread
  - Parent Thread
2. 1. Deuterosis Tue 16 May 2023 10:35AM UTC
    
    Thank you for the answer; I had a sense that machine translation still isn't a true substitute for human translation, but wasn't sure how much the technology had improved since I myself used it to read a Chinese fic, and didn't want to accidentally say something foolish again. (I also didn't have anything to compare the translation to, of course, to know exactly how bad it was.) It's interesting that machine translators even struggle with English to German to this day.
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. stormyseasons Tue 16 May 2023 11:00AM UTC
      
      I was surprised by that too. I'm fluent enough in both English and Chinese to know how I'd have to change sentences and words to get same meanings across, let alone nuance, so that part didn't surprise me, but English to German? Still?
      And those two languages are fairly close-ish as far as grammar goes, as well as some part of the vocabulary!
      
      Comment Actions
      - Thread
      - Parent Thread
3. stormyseasons Tue 16 May 2023 11:02AM UTC
  
  But honestly I do agree with you on that idealogical stance. It may be pragmatically unfeasible to ban... but as you say, that is different from ideologically choosing to accept it all!
  
  Comment Actions
  - Thread
  - Parent Thread
4. 1. Deuterosis Tue 16 May 2023 11:17AM UTC
    
    Indeed! That's really my main concern, and it's pretty alarming.
    
    Comment Actions
    - Thread
    - Parent Thread
5. (Previous comment deleted.)
6. 1. Deuterosis Tue 16 May 2023 12:33PM UTC
    
    Ahh, thank you very much.
    The first sentence in the output isn't even a sentence, nor does it make sense that way... that's funny.
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. Hunter Nightblood (kah0922) Tue 16 May 2023 03:06PM UTC
      
      It's kind of how AI work. As someone who's experimented with both AIDungeon and AI art, AI is like a disobedient dog. You can tell it what you want it to do all you want, but sometimes it just flat out won't do it.
      In regards to AIDungeon (and I believe Novel AI as well), you can't just tell it to write you a glorious piece of degenerate smut and call it a day. You have to guide it every step of the way. And sometimes when you tell it to go right it says "screw it" and goes left.
      To put this another way, writing with AI in that fashion is like telling a person who has barely any idea what the subject matter is what you want it to write. It unironically is a collaborative effort. Just replace the second person with a program.
      
      Last Edited Tue 16 May 2023 03:10PM UTC
      
      Comment Actions
      - Thread
      - Parent Thread
7. Deuterosis Tue 16 May 2023 05:42PM UTC
  
  Regarding "SALAMIs would have to simulate this completely unpredictable process by managing to select otherwise-unrelated concepts that would work with the progression or wording it is generating."
  This does happen to a degree, as pointed out by Hunter Nightblood (kah0922) (I haven't replied because I can't think of one, and am not sure a reply was expected anyway). However, the unrelated concepts I had in mind involve such things as one's memories and personal experiences; generated text certainly can't include things like that (nor can any person who isn't the writer in question). In other words, those things which truly are wholly unique to each person and can't be replicated, can't be simulated by SALAMI in abstract either. Each passage it generates that isn't rewritten, therefore, removes the possibility the writer may there tap into their unique insights to produce passages only they could make.
  This is another reason generated text can't and really shouldn't qualify as "a fan's writing".
  I respect that with enough changes the text does become more the author's original creation. However, to get there significant rewrites are mandatory, and what I mentioned about following the generated beats still applies unless portions of that are also discarded in favor of the author's concepts.
  
  Comment Actions
  - Thread
  - Parent Thread
8. 1. ofthesummer Wed 17 May 2023 01:21AM UTC
    
    If you don't want your work to be co-opted and overshadowed by people who can churn out hundreds of longfics in minutes...
    Or if you don't want the AO3 to get caught up in a costly legal battle on the side of AI, using OUR donations...
    Then write to the OTW and tell them to get rid of Betsy Rosenblatt. Every voice matters, so do the right thing!
    
    Comment Actions
    - Thread
    - Parent Thread
9. KrakenRiderEmma Thu 18 May 2023 04:25PM UTC
  
  "The mere fact of it generating "likely text" seems to preclude genuinely surprising developments in the plot, sentence construction, and use of simile and metaphor. I know there've been many times the story I intended to write and the story I actually wrote are very different, because I had a new, completely unconnected idea that worked well enough to fit in. Sometimes the entire story is transformed into a new, different work from the original intent as I progress and the story shapes in unexpected ways. Sometimes my first go at a bridge was a bad one and I discover a new, better way of expressing what I wanted, that also mutates the work around to be richer.
  SALAMIs would have to simulate this completely unpredictable process by managing to select otherwise-unrelated concepts that would work with the progression or wording it is generating. I am unsure they could possibly become that sophisticated; by definition, the process I described isn't rational or predictable."
  Thank you for this thoughtful analysis of SALAMIs. This part that you wrote is worth discussing -- because it very much depends on how you use the tools. Most large language models, "under the hood," have a setting called "Temperature" or something similar which can be turned up to make the outputs more unpredictable and more "unrelated." If you turn it up too high, it can lead to nonsensical or surreal writing. However, because it makes for less predictable results, a lot of people who are making fiction with AI, especially if they are co-writing in a "ping-pong" back-and-forth way, turn up the temperature as high as possible to get the kind of effect you're talking about, a surprising development that works well enough to fit in. It's not the same as what you might have come up with on your own, but I think it's a little bit like say, doing a tarot-reading to help determine what might happen next in a story -- a randomized process outside of yourself that you can decide whether it makes sense, how to interpret it, etc.
  It is definitely possible to get a LLM to reproduce some text exactly -- if you prime it to do so by basically requesting that it finish an existing sentence from a work it has analyzed, for instance. I feel that this is an obvious case of plagiarism, but it can be avoided.
  
  Comment Actions
  - Thread
  - Parent Thread
10. 1. Deuterosis Thu 18 May 2023 04:59PM UTC
    
    You're welcome, and thank you for this informative reply.
    After it was too late to edit my comment, I realized I put this poorly, because I have heard of generated text spiraling out of control and going odd places. I wasn't able to form a satisfactory wording of what I really meant, an example of which is applying a term of art from one's own field of expertise in a place where it's unexpected, yet is fitting. That's the sort of thing I meant by something a SALAMI can't necessarily replicate. Only just now was I able to word that example in a way that satisfied me.
    However, I'm not sure if that clarification plus the one I replied to my comment with necessarily dispute your point.
    
    Comment Actions
    - Thread
    - Parent Thread
darkwlightb Tue 16 May 2023 10:54AM UTC

fanworks are by implicit definition human made, a machine/program cant be a fan of anything so why are you including them in your protection

Comment Actions
- Thread
synoeketeMora Tue 16 May 2023 01:25PM UTC

thank you for doing this!

Comment Actions
- Thread
impertinence Tue 16 May 2023 01:38PM UTC

A brave attempt to deflect from the fact that your legal chair made shit up to the US Copyright Office, engaging in credulous "AI" boosterism based on vibes and ? people she talked to ?
You should have a survey mechanism to gather feedback from stakeholders prior to claiming you've gathered feedback before, again, the United States Copyright Office.
This includes their position that users should be allowed to opt out from having their works incorporated into AI training sets, a position that they have presented to the U.S. Copyright Office. is a gross misrepresentation of Rosenblatt's comments.
"Maximum inclusivity of fanworks" that includes machine-created nonsense is a ridiculous interpretation of your mandate. Why allow LLM fic and not fic prompts? Why allow a genre of fiction which by definition is created by a for-profit tool with minimal human involvement? Who does this decision serve? Based on Rosenblatt's comments about DALL-E/ChatGPT fic, the answer is, again, "credulous staffers with bad taste", but why precisely does that take precedent over the needs and interests of your now-massive user base?
Shameful day.

Last Edited Tue 16 May 2023 01:40PM UTC

Comment Actions
- Thread
1. Sherloqued Tue 16 May 2023 02:27PM UTC
  
  Yes, absolutely. It's a disappointment.
  And no doubt, before we think just how magnanimous all of this can be, these companies have insured that they have patent/copyright protection!!!!
  
  Last Edited Tue 16 May 2023 02:35PM UTC
  
  Comment Actions
  - Thread
  - Parent Thread
2. 1. impertinence Tue 16 May 2023 02:49PM UTC
    
    Crucially, the companies themselves have patents and expect to profit from the technology but the state of copyright protections for system outputs and indeed the training dataset itself are entirely up in the air. All the more reason to respect your users and not functionally lie to regulatory bodies about what feedback you've gathered. And all the more reason to be cautious about what content is permitted on the site using novel technology, rather than declaring machine-extruded prose "fanworks" worth preserving.
    
    Last Edited Tue 16 May 2023 02:52PM UTC
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. Sherloqued Tue 16 May 2023 03:09PM UTC
      
      Yes, absolutely.
      
      Comment Actions
      - Thread
      - Parent Thread
3. gildedfrost Tue 16 May 2023 09:45PM UTC
  
  What comments has Rosenblatt made to the US Copyright Office?
  The comments that I have seen were from an interview with ARL.org, however if Rosenblatt has made similar comments to the Copyright Office, then that would be much more concerning.
  
  Comment Actions
  - Thread
  - Parent Thread
4. 1. Kereudio Tue 16 May 2023 11:22PM UTC
    
    I don't think the comments are in text format, but this thread breaks some of them down, as well as provides timestamps.
    
    Comment Actions
    - Thread
    - Parent Thread
  2. 1. impertinence Tue 16 May 2023 11:45PM UTC
      
      Yep, this is what I was thinking of.
      
      Comment Actions
      - Thread
      - Parent Thread
5. ofthesummer Wed 17 May 2023 01:16AM UTC
  
  Real fans need to team up and get her removed from her position ASAP.
  
  Comment Actions
  - Thread
  - Parent Thread
6. lurrel Wed 17 May 2023 02:50PM UTC
  
  Whew I really do not like the idea of OTW going to bat for AI to the danged US Copyright Office -- how much work has gone in to coming up with that as an official stance? I do not think the paying members could be considered at a consensus about this at all.
  
  Comment Actions
  - Thread
  - Parent Thread
7. 1. impertinence Wed 17 May 2023 03:44PM UTC
    
    With the full disclosure that i'm not a paying member because I frankly do not trust the OTW to look out for fan interests, yeah, as far as I know there was no effort to collect feedback nor communication with membership (or site users -- who are also stakeholders, since the OTW is a nonprofit serving fandom) and there really should have been. It's disrespectful at best to present yourself as representative of fan interests and make statements about fan perceptions of AI without actually consulting with those fans beyond your immediate social circle or doing any outreach or education of your community. If you don't want to talk to strangers, you shouldn't represent an org that serves millions in front of the government, it's that simple.
    
    Comment Actions
    - Thread
    - Parent Thread
8. Unicron Sun 21 May 2023 01:50AM UTC
  
  For me that is as big an issue as anything, and I thank you bringing it up when many people aren't. The fact the chair spoke in an offical capacity in favor of aI for the organization when she is set to profit off of AI work.
  Even in a nonprofit you NEED to have standards of conduct for your volunteers. It was such an obvious conflict of interests I am disgusted there isn't anything in OTW's bylaws to kick people out for things like this.
  
  Comment Actions
  - Thread
  - Parent Thread
DaytimeSteakGrinder Tue 16 May 2023 02:24PM UTC

Disappointed to see this is what's being focused on with this site. It's pretty obvious people don't even understand how these artificial intelligence things work; just using a kneejerk reaction. There are still legitimate issues with both the site and its userbase that deserve more attention. I won't be putting my own work here again. I'm not commenting further.

Comment Actions
- Thread
BlackPanther1412 Tue 16 May 2023 03:23PM UTC

Taking measures against scraping is good but you cannot condemn the tool and allow the result of that same tool to exist. AI fanfiction is NOT fanfiction and should not be welcomed on this site. As another comment mentioned, no one is expecting ground breaking code to be developed that would identify AI works, but at least don't welcome AI as a legitimate form of fanwork in site policy.
Please don't destroy a beloved website for millions of fans just because AI became a trend. Especially with the current political climate and the WGA strike taking place.
AI FANFICTION IS THEFT, NOT ART. PROTECT THE WRITERS!

Comment Actions
- Thread
KitePiper Tue 16 May 2023 03:54PM UTC

I'm just.... so tired of being devalued and unprotected.
I started writing out a long comment with arguments and examples and citations but that's what it boils down to. I'm so tired of being devalued and unprotected and exploited on a site that was supposed to be safe for writers, just for being stupid enough to post stories I was excited to share with other fans.

Comment Actions
- Thread
1. Sherloqued Tue 16 May 2023 03:59PM UTC
  
  I know. I've been going over in my head 'if I had it to do all over again . . .' too.
  But, we've still got the opportunity to share with other like-minded fans, it's sad that this kind of thing has to spoil it. I hope the situation can be addressed.
  
  Last Edited Tue 16 May 2023 03:59PM UTC
  
  Comment Actions
  - Thread
  - Parent Thread

AO3 News