#ai training

[dduane]

For those who also have this problem: info on the class action suits, and where to start taking action.

(Here's what they took without asking me, BTW.)

[oldschoolfrp]

It just takes a second to turn on "Prevent third-party sharing" for AI training:

On Desktop: "Blog Settings" then scroll almost all the way down

On Mobile: Click the Settings gear icon, scroll down to "Visibility"

Repeat for each main or secondary blog

[offical-firefox-nightly]

Hey guys, uh I need y'all to do me a favor-

Go to each of your blogs' settings and turn the highlighted settings on- it'll opt you out of being shared to third parties, including being trained for AI. Unless you want to be used for AI idk up to you. Sorry not sorry for the ping btw.

@claires-unofficial @nasa-real @totally-dollar-tree @totally-official-yahoo @totally-official-goodwill @the-true-internet-explorer @realsafari @im-pandora-i-promise @the-tumblur-searchbar @definitely-wikipedia @firefox-official @official-fedex @its-target-official @hot-topic-unofficial @spotify-official @walmart-the-official @the-real-google @the-real-firefox @totally-ikea @yahoo-official @mcdonalds-official-verified @official-opera-gx @official-chai @theveryrealvine @bingle-official @intothegimmickverseofficial @spotify-kids-real @same-pic-of-pluto-everyday @sams-club-official @same-pic-of-neptune-everyday @same-pic-of-a-dictionary-daily @the-one-and-only-pornhub @post-uwuifer @definitely-spirit-halloween @definitely-youtube @definitely-tor-browser-official @scottishslenderman @slenderdawg @robloxian-slenderman @the-fake-list-of-ominous-threats @bitchlessslenderman @gimmick-blog-stimboards @are-there-grammar-errors @grammarly-unofficial @important-question-anon uh I probably forgot a gazillion people but that's okay

[tinyperson00]

ATTENTION ALL ARTISTS I THINK I FOUND A SOLUTION TO THE AI STUFF

I was going thru the blog setting cause I wanted to change something then I can across THIS

TURN THIS ON. PREVENT AI TRAING FROM YOUR BLOG!!

Hopefully this works ;-;

REBLOG THIS TO LET OTHER KNOW!

@aceofstars0 @larz-barz @tokito-dulya20@frostburn-shoto @silliestsakura @muitsuri @shycroissanti @kiyokatokito @kimetsu-chan @ta-ni-ya @your-local-demon-slayer-nerd @cloudymistedskies @demon-slayer-manga-simp @boo-simplified @night-mince0@pinkwisteria @sweetstarryeyedgirl @giyubabe @axolotl321 @1julak1 @ghostbite0

[mariluphoto]

And things get even more terrifying!! I saw someone mention this in Jon's insta comments.

Everyone is at risk and not just artists...

Here's the full article!

[savage-rhi]

**Just an FYI, there's nothing concrete as of yet, and I'd encourage people to draw their own conclusions.

[27paperlilies]

Tumblr is selling it's user data to train AI tools. So to keep your blog content from not being used, go to blog settings > Visibility > Prevent third-party sharing.

Enable this for all your blogs and sideblogs. Spread the word please <3

thank you for the info anon! Spread the word!

[winterthebeau]

fuck microsoft

[starwikia]

wait i just realize something hilarious about the ai thing.

do you guys realize how many bots there are on this site?

do you guys realize how many porn bots are on this site?

there’s no way they’re going to be able to accurately identify even a percent of them, especially give the sheer scale. my god. if this somehow poisons the data bases of whatever got the data, it’s fucked. this is going to end up horrifically backfire on the people who want this website.

that’s not even going into how people are making side blogs already to purposely mess with the data gathered

tech bros severely underestimated how users are just completely willing to fuck over the staff on this

[avaantares]

Prevent AI from feeding on your blog!

There is now a setting on Tumblr to toggle whether or not your blog content can be sold, and by default it allows your content to be shared with third parties and potentially used for AI training.

Unless you want to be a maverick and deliberately poison datasets (which... honestly sounds kinda fun and I would probably do it myself if I had more time), go to Blog Settings -> Visibility for each blog you have and toggle this setting to On:

Again, do this for all of your blogs separately. Just doing your main will NOT change the settings on your sideblogs.

Spread the word, and death to AI content creation.

[dduane]

Found in the ask box this AM

”I've been playing with [a major AI-driven chat bot] and I asked it to write a young wizards fanfic that contains dialogue. ...I wish i could paste it here, but the character limit won't let me.”

(hiding eyes)

(a) Once more, the reminder: please, please don't send me story ideas or fics set in my universes. Reading them would expose me legally in ways I can't permit. If such things turn up in my ask box, I delete them unread. If through a lapse of concentration or some similar error I find myself mistakenly reading a story idea or fic, I am required to make a note of it for legal purposes, and then can never use those ideas in my own writing, ever, anywhere.

So please understand that sending me a story idea or a fic set in my universes (or others where I write) will guarantee that the idea never happens in my work. It may seem paradoxical and/or counterintuitive, but if you have an idea that you wish I'd do something with, the odds of that happening are significantly increased by you making sure that I never see it.

(This, BTW, is one of the reasons I had to close my ask box to anon asks; some repeat offenders were making it impossible for me to protect myself... as when people refuse to stop sending me such things, I block them. My apologies to all the well-intentioned would-be nonnies out there—but the actions of some of you have destroyed an avenue of access for everybody else.)

(b) Please do not purposefully expose my IPs to the attention of any AI/LLM-based chat platform, anywhere. Unscrupulous whole-book AI-oriented scraping has stolen enough of my stuff already this year. (As my search of the Books3 pirated books database, widely used for AI training, shows here.) It's infuriating enough that bad actors have taken my writing and will be using the expertise inherent in it to put other living, breathing writers out of work. Please don't help them do more of that.

(sigh) Thanks.

[henchmanintraining]

Are they really training AI on the website that gave us "car door hook hand?"

[ebookporn]

An interesting development in dealing with the AI "training data" issue. Though supported by the AAP Fairly Trained seems to be weighted in the music industry. Let's see if this gains some teeth.

[pearwaldorf]

content notes: worker exploitation, exposure to traumatizing content, descriptions of sexual assault

We knew that ChatGPT had scraped AO3 but this TIME article is the first confirmation from a news organization I know of that has explicitly confirmed a fic was used in training.

So if you're going to release an AI out into the world, it needs safety tools. And you need data to show it what it needs to look out for. This is all basic machine learning stuff. Things this tool need to learn to recognize include:

child sexual abuse, bestiality, murder, suicide, torture, self harm, and incest.

And one of the things mentioned was a non-con fic in which a superhero sexually assaults his sidekick. The article does not mention which site you can find the fic on, but it is entirely possible it could be from AO3.

(Don't you dare use this as a justification for saying this is why this sort of thing shouldn't exist on the internet. That's bullshit and not the discussion we're having here.)

Most authors are not sadists (to real people anyways). If this was on AO3 a warning would be required at the very least. This is not the sort of thing you want somebody to read if they don't want to, and it makes me feel ill knowing somebody was forced to do so. I think the author would be horrified as well.

The thing about a lot of AI is it is obscured manual labor by humans, a la the Mechanical Turk, who are usually not treated well. ChatGPT is in a long line of companies (including Facebook and Twitter) who exploit workers, usually in the Global South, to do unpleasant and often harmful work for cheap.

I also don't think we as fanfiction authors (I certainly wasn't) were cognizant of how our work, decontextualized from safeguards like warnings and tags, could be used to inflict harm without our knowledge or consent.

This does not mean we should stop writing things things like that, but we need to think about ways to limit access to unauthorized collection/scraping of our work on a systemic level. I don't think we can do anything about what's already in ChatGPT's training data, but there may be ways to prevent them and other companies like them from getting more; and to force removal if it comes to it.

I don't have definitive answers, and I'm not sure I'm the person who should be coming up with them, but it's something I'm happy to discuss with people and/or groups who are better suited for it.

[wixed]

Hey so you know those antiAI posts that are just nonsense words so the AI that is getting trained on posts learns nonsense? What... what would happen if we blazed them? Like... does the AI scraping take blazing into account?

I don't want to spend money on it, but also God it would be so funny if the blazing actually helped kill the AI.

oh and for good measure. I talked about hair ran out to grocery carton of eggs for table legs at dinner for us to enjoy. Over next days we gathered with smile to the camera. 😁

[melchinafan]

Remember to opt out of the ai update in your blog settings so tumblr cant steal your art

((Not a joke, its above mass post editor))

Aw, thanks for looking out and giving a heads up, anon. I have seen some posts about it (and will reblog some of them in a moment), but I appreciate you making sure I knew.

For the record, the relevant setting is under Blog Settings -> Visibility -> Prevent third-party sharing. It has to be done individually for every blog/sideblog, and (currently) doesn't seem to be available on mobile, from what I've seen.

There are a number of questions on the actual efficacy of the setting (it's possible that data may have already been sent? why is it fucking opt-out instead of opt-in, yet again, when there were huge uproars and really damned solid points made on why that's shitty when places like deviantart did it, like "what about folks who have left/died"?)...but better to do it than not, just in case it actually does something.

I have been...struggling very hard not to fall to the nihilism in *gestures vaguely* all of this, for a long while now. But boy howdy, does it keep getting harder.

It may also be worth it for folks to try editing old posts to update their images with a Glazed and/or Nightshaded version, in the hopes that those are what get trained on. (Glaze is intended to protect from style imitation, and Nightshade poisons training data so incorrect outputs are given relative to their word prompts. If you use both, the dev team recommends Nightshade first, Glaze second.)

Also, I was gonna wait until I got an email response to a few questions I had about the site (mostly like "hey can we ramble in tags or are they purely archival haha"), and then maybe poke around myself to see how I felt about it, before bringing it up. But if you're looking for a new combo art portfolio/social media site, there's one called Cara. Their small team has worked with the Glaze team to incorporate a webglaze version into their site (and will eventually include Nightshade, when a web version of that is also made available). It seems like it could be a good option for folks who are interested in using Glaze, but don't have a computer able to run it themselves.