Karla Ortiz is a leader. Her professional work inspires colleagues and fans. She advocates for creative rights. It's not just her talent that makes her an icon. She is art in action.
You probably know her concept art images for Dr. Strange...
And this often shared (rarely credited) concept art for the famous Game of Thrones character Rhaegar Targaryen
Karla is a popular panelist at conventions. Her online and in-person workhops are packed with industry peers as well as art students and collectors. Karla’s using her platforms, experience and contacts to start important conversations about issues impacting indie art life.
Between July and October 2022, her Facebook page became a remarkable series of posts, and informative comment threads, about Copyright and Creative Rights issues unaddressed as AI art platforms (Midjourney etc) evolve.
With Karla’s permission, I’ve collected her posts here for reference. Facebook posts can be hard to retrieve. My blog post here makes Karla’s July – Sept 2022 AI art posts accessible. Her writings document the rise of awareness…. and alarm… in the artist community. Karla's posts are a remarkable record of the AI art issue at an early point of its impact on the community it thrives on, but refuses to credit and compensate.
AI
art issues arrived on my Radar via an announcement on Karla’s page about the 8.31.22
Concept Art Association Town Hall, hosted by Karla, Nicole Hendrix Herman and
Rachel Meinerding (UPDATE 11.22 -- this turned out to be the 1st Town Hall Meeting)
“AI
& Industry: Impact, Issues and Solutions Townhall.”
Hosted
by the Concept Art Association (and Karla), with guest Christian Alzmann,
Lauren Panepinto , Craig Mullins, Frank Gonnello and Abhishek Gupta.
Thank
goodness they made it available as a You Tube video. Here’s the link:
More
about the Concept Art Association here:
https://www.conceptartassociation.com/about
Social media platforms profit from image sharing. Online artWORK posted there gets shared without credit or compensation to the artists. AI art, with its practice of “scraping” images from sites like Pinterest, takes copyright protected art and spits out derivative versions based on prompts. US and Intl copyright law will be playing catch-up in the courts for a long time. Meanwhile….it’s up to artists (again) to be raising awareness for fans, clients and colleagues on these infringement practices.
Educated fans and peers are the front lines for protecting creative rights. We need artists like Karla who are not only top in their fields, but also articulate and passionate about creative rights. We can’t let technology benefit from artWORK without compensating artists.
Online
art is never free. Art is copyright protected creative property. Tech companies
are trophy hunters turning artists into endangered species. They will kill what
we all love… if we let them. Get educated. Get activated.
I’ve started this series of Karla's posts with her 10.7.22 post about the AI art shared after the sudden death of Kim Jung Gi. Her post shows the visceral impact of others taking an artist's work and running it through an AI grinder. Especially right on the heels of a painful loss still resonating through the art community.
After the Kim Jung Gi post, I’ve put Karla’s posts on AI in chronological order. From her 7.30.22 “call to action” post… through her 9.30.22 post, where she was “taking a break” from social media (for a few days).
Her posts on her Facebook page from July – Sept included in depth comments threads with other professional artists. (There is a sample one in the 9.29.22 post. ) I can’t include all of the comments thread exchanges here, but they can be found (for now) on her Facebook page. As we all know, over time Facebook deletes posts and it gets hard to scroll back to find earlier entries. So search now for those while you can.
Karla's posts also included links to wonderful articles and resources, most of which
appeared in her original post texts and are included here in this compilation.
OCTOBER 2022
10.7.22
FP Post (Kim Jung Gi RIP 10.3.22)
"As
we mend our broken hearts from the recent passing of one of our brightest, this
person was busy taking the beautiful work of our friend and trained an “AI”
model. Not once wondering if the family would be ok with it, or if it was a
good idea. By training he means taking the work of our passed, claiming it
"draws" when he means generated, and going against the beautiful
things Master Jung Gi stood for. But please don’t forget *HIS* credit.
If there is any shred decency in this person, they’ll delete the whole thing out of respect to the Master, his friends, his family, and the artistic community."
(This
Muddy Colors link has many Kim Jung Gi videos.
https://www.muddycolors.com/2022/10/remembering-jung-gi/?fbclid=
July 2022
7.30.22
Facebook pg
Karla’s “call to action” post:
“Sooo
I tried one of those fancier AI apps (DALL.E)
My
thoughts so far: I think folks job's are safe, for a while at least.
Artists
are still EXTREMELY important, because even with a lot of iteration all you get
from the AI is a huge giant mess of shapes. Completely unusable for any kind of
serious production.
For
now, it's great for generating inaccurate/messy references!
I'll
check in with it in a couple years and see if that's changed, but honestly,
giving it a try calmed me the heck down 🙂
Ps.
Still super uncomfortable about training AI to imitate living artist’s work
without their consent. That should be a no no. Could imagine some solid lawsuit
coming out from this.
Ps
#2 I want this tech to be done ethically. Honestly I think artists NEED to get
together and start preparing for class action lawsuits. Make it really clear
real early that this tech cannot just simply use our work to create prompts,
even if those prompts are messy. Sooo…who wants to do this?"
7.31.22
Facebook pg
Midjourney
Karla’s post: “Ok did a deep dive of Midjourney. I got some concerns.
They have a basic subscription model, between $10-$54 dollars per month. So if your name is being used as a prompt on that site (and if you’re an even remotely publicly known artist, actor, public figure, etc, your name is on there.) they already profit from it.
They also offer a corporate subscription model for $600 per year. Take a look at their commercial license. They claim they can give their corporate clients full ownership to the assets Midjourney produces, which is absolutely wild considering the overreach of their prompts.
Their terms of service is extremely vague overall, but if they train an AI to use copyrights, or anything of the sort, they basically shift the blame to the users. But what about the profits they made from said user?
Outside of extremely odd wording, there’s also no follow up on what they actually do when copyright infringement occurs. They say you can file a DMCA takedown, but how does that work? Do they erase all content that had copyrighted prompts? Do they guarantee that the AI can no longer utilize that learned imagery? What about if you do not want your name utilized there as Midjourney profits from it the second someone uses it? This all seems extremely vague and honestly shady to me. (next 5 images were shared by Karla on her post)
Midjourney
looks like an absolute legal mess. I can 100% see folks using a celebrity’s
name as a prompt, get a fairly decent likeness, and utilize that for an ad if
you decide to buy a corporate license. For example an illustration of Scarlett
Johansson for a sci fi book as pictured below. Note, this is just 1 example out
of many. I saw hundreds of realistic looking celebrities, flagrant use of ip in their discords
everywhere, etc. it blew my mind. That they have not yet been sued is a
miracle, probably made possible because unless you have access to the discord,
you cannot see the images created.
As
mentioned in previous threads, i think AI is an interesting tool that can lead
to exciting use cases, but it needs to be approached in an extremely ethical
and measured way. And from what I’ve witnessed today, Midjourney is anything
but…"
AUGUST 2022
8.13.22
Facebook pg
Karla’s post: “Ethical and technological concerns aside, when I look at AI, it makes me think how starved folks are for figurative work. Most of the prompts I see are people trying to recreate Bouguereau, Mucha, Sargent, etc.
Makes
me wonder if collectively we truly need aesthetically pleasing and tangible
art, far more than what the art markets provide. Or far more than what we
imagine ourselves.
Also,
real talk, it looks like a singular new incredible artist came into the scene,
as it kind of looks like work from the same person to me. But yeah.
Interesting, enchanting, concerning, frightening. Really feels like someone
opened Pandora’s box, huh? Wonder what it’ll be like in 5 years.”
8.14.22
Facebook pg
Muddy
Colors 8.11.22 link. https://www.muddycolors.com/2022/08/robots-vs-lawyers/?fbclid=
Karla’s
post: “Excellent article on the state of affairs of AI and IP laws by the
wonderful and insightful Lauren Panepinto via Muddy Colors”
8.30.22
Facebook pg
Karla’s
post: “Here’s a very interesting tool that checks which images were used and
from where they were taken, by the first Stability Diffusion dataset.
I
searched my name and saw plenty of my work. Note from the authors “ this is
only a small subset of the total training data: only 0.5% of the 2.3 billion
images that it was first trained on.
Funny
enough all of my images are mostly scraped from Pinterest, of work that I’ve
never even uploaded to that site. Including my own traditional work 😭. (Posted in comments below)
So
much trademark content in there too☠️
Here’s
the tool to check out if your work is in a tiny subset of one the three massive
databases Stability Diffusion was initially trained in:
https://laion-aesthetic.datasette.io/laion.../images
Fun
note this tool only looks at .05% of 2.5 billion images. The current data base
for Stability Diffusion is approximately 5.8 billion images!!!
The
article explaining the tool is fascinating full of all kinds of interesting
information and notes:
Great
work by Andy Baio and Simon Willinson!
SEPTEMBER 2022
9.1.22
Facebook pg
Karla’s
post about 8.31.22 Town Hall meeting w/ link for You Tube download.
“Hello
everyone!
Here's
the recording for yesterday's AI & Industry: Impact, Issues and Solutions
Townhall.
Hosted
by the Concept Art Association & myself, with guest Christian Alzmann,
Lauren Panepinto , Craig Mullins, Frank Gonnello and Abhishek Gupta (further
guest information in video!) .
I
believe this was a very much needed conversation for this pivotal moment, so
hope you an all give it a watch here:”
https://www.youtube.com/watch?v=LYO9sii1eKA
9.3.22
Facebook pg
Karla’s
Post:
“Oh
boy. This is just the Bloodborne marketing art pasted on some of these ai
generations. ☠️
I
know fans dont like to hear this, but stuff like this is an issue, especially
when commercial rights are granted. It’s
just so messy at the moment. 😔
Ps.
these images were found by using Lexica! A really handy tool that let’s you
check some of Stable Diffusion imagery by prompts alone.
You
Tube clip shared on comments thread
https://www.youtube.com/watch?v=e_mq6ANKAm4
9.10.22
Facebook pg
Karla
post: “On local hellsite, twitter dot com, the wonderful artist and ai prompts
fan favorite, Greg Rutkowski, expressed the following: "Well I guess soon
I won't be able to find my own work on the internet cause it will be flooded
with ai stuff."
An
unforeseen consequence of all this is that if you get popular enough as a
prompt, your own work can be drowned out by intentional imitations of your work
in search results and social media. Could this harm our ability to advertise?
Shame
to all the companies that took the names and work of artist without consent,
permission or compensation, and are making insane profits by doing so.
I
mean, even early AI software, Disco Diffusion, used Greg’s name as an EXAMPLE
prompt (pictured below. Source at end of post) He was never once asked.
AI
visual companies should’ve asked Greg for permission and worked partnerships
with him to *opt in* to this. They could have given him a small percentage of
profit each time his name or work was used. They could’ve materially changed
his life, while users enjoyed the results of his hard work. But instead they
engaged in an egregious overreach of data, claiming all of our works and our
names as theirs to use as they see fit.
These
companies say this will “democratize art”. What they don’t say is that they do
so on the backs of artist’s work and their names, with no way out. Hell, not
even the families of our dead can have a say in how their loved ones art gets
used (Syd Mead, H.R Giger, Thomas Kinkade, Frank Frazetta, Moebius, etc).
The technology is indeed exciting, and it shows so much promise. But this current iteration is deeply irresponsible.
Worst
in all of this is that even if opt out measures are hastily implemented, in the
form of blocked prompt lists, because of how the software works, it wouldn’t be
able to fully forget the works of artists who do not wish to be a part of this.
This is because once the Ai associates, let’s say Greg’s work, with other
generic terms such as “fantasy” or “oil painting” his work will continue to be
used to generate imagery even if using other prompts, as the software has made
that association.
So
the AI software will continue to use his work, even if these companies managed
to block his name from being a prompt, if he so wished it.
Its
frustrating because it didn’t have to be this way. Here’s how it could be:
Smart
companies could future proof and build models based only on public domain
works. That alone could create wonders (hello, Sargent is in there! ). For more
modern works, seek partnerships with artists, photographers, actors, animators
etc.
Companies
could pay to use artist’s works and names, each time the work was utilized.
This also celebrates the work of the artists they utilize, while ensuring that
artists who do not wish to be a part of this don’t end up in their software.
Respectful and beneficial.
Companies
like this would also be able to offer true commercial licenses to users,
without any concerns of copyrighted data being utilized at any point of the
process. This would make for a technology that respects creatives, doesn’t
steal from them and shares profit.
We
need responsible leaders in the Ai world who will not only call out rogue
companies that operate on nefarious practices, but will also lead the way to
more equitable technologies. Especially as these technologies could potentially
harm our creative workforce ecosystem.
In
conclusion, if you say you care about art, about creativity, don’t then turn
around and take the work of photographers, fine artists, illustrators, concept
artists, or artists in general, so your software generates better images,
without paying them and not even giving them a true way out.
Notes:
1. Disco Diffusion: To find Greg's name (alongside Thomas Kinkade's) utilized as an example, scroll to the "Prompt" section. As of today, its clearly there.
With
this link: https://colab.research.google.com/github/alembics/disco-diffusion/blob/main/Disco_Diffusion.ipynb?fbclid=
9.14.22 Facebook page –
Content
Warning: more AI stuff, and more long writings/musings ahead. If exhausted by
the subject please, skip this. I swear I'll flood your Facebook feed with nicer
things in the future! Ok with the warning done, let's begin:
Today,
a fantastic article from the Guardian was released that allows us to see a
glimpse of the views of former hedge fund manager and venture capitalist, now
founder of Stability AI/ Stable
Diffusion's founder, Emad Mostaque ( fun fact, he also provided an initial
grant for Midjourney!).
This
section is particularly concerning:
""
But Mr Mostaque says he's not worried about putting artists out of work - the
project is a tool like Microsoft's spreadsheet software Excel, which - he notes
- "didn't put the accountants out of work, I still pay my
accountants".
So
what is his message to young artists worried about their future career, perhaps
in illustration or design? "My message to them would be, 'illustration
design jobs are very tedious'. It's not about being artistic, you are a tool"."
He
suggests they find opportunities using the new technology: "This is a
sector that's going to grow massively. Make money from this sector if you want
to make money, it'll be far more fun"." You can read the full article
here: https://www.bbc.com/news/technology-62788725
This
is dehumanizing language, intent on reducing human artists to mere
"tools". It is a deeply offensive tactic, but there is purpose behind
these words. So let's discuss it.
Your
labor, your craft, your dedication, and your human experience has to be reduced
to that of a tool. He needs your artistic existence and your contributions
erased in order to make *his* tool more palatable, to ease the potential
existential threat this is to human artists.
He
also pushes the perception of professional creative endeavors as tedious, or
another implied word, "menial". Because that is the promise of
technology. To rid us of menial labor, like digging holes with shovels, and
he's here to sell you the latest hole digging software!
But
art, even in professional settings, is not menial. It is skilled creative work.
It is humans communicating ideas to other humans, whether they be visual,
audio, language, etc. It is an incredible process at the heart of humanity, one
human communicating with another human.
Let's
take a look at a professional artist who Mr Mostaque, claims is a tool. HR
Giger is an admired and loved artist for
our community and the world. His work, contribution and life experiences
brought us the haunting imagery of the Xenomorph from Alien & so much more!
But
it makes sense for Mr. Mostaque to wish to see HR Giger's contributions to
humanity as mere a tool. In his own software, Stable Diffusion, HR Giger's name
and work has been utilized as a prompt
24,130 times (as of 8/13/22). All of this without the consent of HR
Giger's estate.
But
that's the purpose here. When you see humans as tools, then it's easier to
erase their contributions and easier to disregard their human rights.
Ironically,
Mr Mostaque, is dead set on telling the world how creative and artistic his
software is! The software that was built upon the work and contributions of
those artists he now trying to diminish. The software that is
"democratizing" art, but
"do please dismiss artist's vote and voice, in fact dismiss them
entirely".
I'm
not sure if he says this out of deep belief,
or if he NEEDS to believe this, as it's easy to dismiss the harm caused,
when you dehumanize those you harm.
I'd
honestly love to talk to him face to face to see which is it!
But
one thing is clear to me, the disdain, lack of empathy and misunderstanding of
human artists cannot be the defining feature of our communal AI future. In my
opinion, a person with such harmful beliefs towards those he benefits from, is
just not fit to lead this moment.
The
harm this will create to humans, to artist and our collective human creativity
would be immeasurable if the defining philosophy to this monumental development
was "make money, take credit, use real people's craft and their humanity,
as tools for you to use and discard".
To
me this is is a wake up call for all writers, musicians, photographers, fine
artists, performers, animators, game developers, poets, anyone in the creative
fields (heck, even coders). You must know there are plans in motion by Stability
AI to expand to every reach of our industries (seriously read through his AMA
from a couple days ago, link below).
If
there was ever a time for us to unite it is now.
If
we do not advocate for our creative humanity, if we do not organize and mobilize,
the ones at the table dictating our future will be those who see us as tools to
be profited from and then discarded.
It's
already happening in UK. AI companies are seeking to push laws to be made where
all of our copyrighted works can be utilized by their companies, for commercial
purposes without needing to seek permission.
Full consultation papers in links at end of post.
Anyhow
I know it's dreary, but this doesn't have to be our future. We can work
together to create models of AI that compensate our work, if we so desire. We
can work together so no artist's work is a part of these data bases, so our
names are not used without our consent. We can work together to ensure limits
of AI are implemented in industries, so whole teams aren't at risk of
replacement! We can do so much! We WILL do so much!
I wish we weren't living through this. The pandemic, the wars, the state of our world, is hard enough as it is. But this is the time for us to step up, reach out to each other, and help define what this technology will mean to us, to our children and to all future artists.
Let's
end it with a better author than I:
“I
wish it need not have happened in my time," said Frodo.
"So
do I," said Gandalf, "and so do all who live to see such times. But
that is not for them to decide. All we have to decide is what to do with the
time that is given us.”- J.R.R Tolkien
Sources:
Reddit
AMA thread: https://www.reddit.com/r/StableDiffusion/comments/x9xqap/ama_emad_here_hello/?utm_source=share&utm_medium=ios_app&utm_name=iossmf
UK
consultation papers:
9.15.22
Facebook page
Karla
on AirTalk
https://omny.fm/shows/airtalk/rise-of-ai-art-9-14-22
Here's
the excerpt of yesterday's chat on KPCC NPR Air Talk with Larry Mantle.
Apologies
for my own audio quality. If there's a next time I'll be sure to upgrade my
sound! There's a couple points I wish I could've made but honestly as it was my
first radio interview like this, I was a tad nervous lol
Anyway
loved being a part of this, alongside Claire Leibowicz! Such a wonderful chat!
9.16.22
Facebook pg
Karla’s
post: “I think everyone should read this fantastic piece by Melissa Heikkilä.
Full
transparency, I’m quoted briefly, but more so important read Greg’s story!”
With
this link from Technology Review:
9.18.22
Facebook pg
Karla’s
post: “Seeing some folks dismiss serious grievances towards emerging AI
technologies. I get it, it’s exciting and could potentially be a great tool!
But
all I’ll say is, if AI researchers, AI engineers, AI ethic advocates, and even
AI task force members to the White House are horrified to find out how this
technology actually works, how it uses data scraped from illegal sources (Like
Pinterest, Shopify, Deviant Art, or Fine Art America) and how it’s a huge tech
overreach of copyright and privacy data,
then maybe it’s not such good idea to diminish legitimate concerns?
Just
a thought!
Anyway
here’s a wonderful read. Check out in particular Daniela Braga’s thoughts here.”
With
this link from Forbes:
9.18.22
Facebook pg
Karla’s
text:
I
wrote this as a response to a friends comment (im so sorry its so long 😭) but I figured the research and info was worth
sharing. note: I changed the original text a bit to better fit as a post:
The
question overall is, do AI companies utilize copyrighted material?
So
I found out via their Discord MidJourney utilizes a data set called Laion. If
I’m not mistaken they’re upgraded from Laion 400 to Laion 5b.
Both
of these data sets are currently being studied in depth, but before getting
into it, you can actually search these data sets on your own!
In
there you’ll find what images the ai is being trained on, and lemme say its
riddled full of copyrighted material from a lot of different sources not just
artists but also businesses of all kinds. You can search the data base here
(search result set to HR Giger so you can see how many of his works are in the
data base).
You
can also use tools such as lexica.art, Libraire.ai or haveibeentrained.com to
find your own works.
Anyways,
good news is these data sets are being studied upon. Specifically where these
images are scraped from. Here’s a study on the initial data set:
https://waxy.org/.../exploring-12-million-of-the-images.../
And
here’s another study on a larger data set:
What
research has shown is that the data sets scraped data from places like Shopify
(36,469,372 times!) Pinterest (30,908,667), Deviant Art (500,000+) times, and
heck even business like Nordstrom or Disney.
This is a huge legal mess because the ToS of so many of these companies
disallow for data scraping, especially for commercial purposes.
Do
keep in mind this isn’t a study on the full data set. This is only research of
one of the 6 data sets that Laion 5b is trained on. These studies have only
looked at about 600 million images out of 5.8 Billion images. Not to mention
this is not only scraping data from artists without permission but also from
all kinds of businesses, both big and small!
So
if the questions is “is there copyrighted material used to train MidJourney and
Stable Diffusion” the answer is 100% yes.
Now
what about the question, does ai generation create something new? Even ai
experts are unsure of how deep learning actually works (tho most of them will
say it’s definitely not like how humans learn). But we have seen AI generated
imagery straight up copy paste material like for example the Bloodborne
marketing material.
Compare
Stable Diffusion generations here:
https://lexica.art/?q=Bloodborne
To
Bloodborne’s actual marketing material (added as an image on this post) . If
you also scroll through MidJourney’s discord you’ll find a bit more variety but
dig long enough and you’ll still find copy paste like material.
I
have a suspicion that a lot more copy paste occurs but because the data sets
are so huge it’s difficult to verify. An example is the case of the painting
“Scorched Earth” done by artist Arcipello in 2012
(https://www.deviantart.com/arcipello/gallery ). A user generated an Ai image
that has an uncanny resemblance to Arcipello’s original painting. Tho important
to note the AI could have also used this rip off of Arcipello’s painting by
this amazon business as a base:
(https://www.amazon.com/Diamond-Painting.../dp/B09Q8PW5VY )
Now
the interesting thing is that when the user provided the prompt used, nowhere
was the name Arcipello used.
This
could potentially be an example of how ingrained artists work is in the data
sets, as even when generating imagery without the use of an artist name, we can
get similar images, and potentially copies of an artists work. This could also
potentially exacerbate theft issues, as one could argue it sampled amazon’s rip
off and not the original painting (examples in the post). “Theftception”?. Too
many questions. Too many legal grey areas . Its wild.
The
tech is indeed exciting, and could be a wonderful tool, but its current
implementation of it is a mess imo 😞
TL;DR:
Yes, there is without a doubt a LOT of copyrighted data being used. Also we’re
not sure how much of the generation process is remixed or just copy pasted.”
With
photos and this text:
“Prompt:
a volcanic fire waterfall from the mountains of cherry blossom trees on to
ocean.”
9.20.22
Facebook post w/ this comment from Karla:
“Regardless
how anyone feels about AI, this kind of exploit is just not ok.”
With
this Artnet link:
9.21.22
Artist’s Medical record photos posted Link posted on her FB page 9.21.22
9.21.22
Getty Images Link posted on her FB page 9.21.22 w/ this excerpt:
“There are real concerns with respect to the copyright of outputs from these models and unaddressed rights issues with respect to the imagery, the image metadata and those individuals contained within the imagery,” said Peters. Given these concerns, he said, selling AI artwork or illustrations could potentially put Getty Images users at legal risk. “We are being proactive to the benefit of our customers,” he added.”
9.29.22 Facebook pg
This
is exciting stuff but we're gonna need regulations on this stuff a.s.a.p!
I
also hope all creatives realize this isn't an issue for just concept artists or
illustrators, it'll soon be an issue for all of us. These technologies will
only get better.
Btw,
if that proposed law in the UK comes to pass, note that anyone's copyrighted
data can be used freely, without permission and for profit, by companies like
Stability Ai, but also behemoths like Meta/Facebook too...
Edit:
holy shit the entire database that trained this was taken from Shutterstock,
potentially without permission AND from a Microsoft Asia Research set that
scraped 3.3 million Youtube Videos. HOLY SHIT.
All these data sets were created by research licenses but FB is likely
to use this for commercial purposes. Wow. Read all about it here:
Are
you an Illustrator in the UK?
Do
you know there is a proposed law change in the UK that could allow corporations
(specifically AI companies) to use your work and data without any permission or
compensation, and use it even for commercial purposes?
Tell
the Association of Illustrators how you feel about it! They can’t fight for you
if they don’t hear from you!! 🔥 🔥
And
to read their full explanation on this
check out their twitter thread here:
9.29.22
FB pg
"Is
AI art “art”" is the most boring discussion.
In
my opinion the better discussions are:
How
can we ensure this tech isn’t exploitative, that creative labor is protected
and compensated, and that the public has the tools to know if something is AI.
What policies will we need to do all this?
Speaking
of which, out of curiosity, what policy and/or regulations would YOU like to see
implemented? Both for companies providing AI tools AND companies utilizing it.
Her
reply to Miles Vesh Miller (Design Engineer – Spectacles at Snapchat) from
Comments thread to this post…
From
Miles:
Miles
Vesh Miller
(replying
to Feifei Wang) -- there are no “source
images” being composited. This comment along with many others are deeply
misunderstanding the very core of how deep learning works.
It’s
hard watching so many artists I respect going through an existential crisis
with advances in ML, but also disheartening to see the calls for “what policies
should we implement” from a community that quite literally has no clue how this
works.
I’ve
been hesitant to even waste time posting a response to any of these threads, as
I’m literally walking into a lion’s den.
Karla
Ortiz, if you’re ever interested, I’d love to talk in depth with you and share
the perspectives I have. I’m a big fan of yours (and came to one of your
gallery showings in SF a few years back!). I’ve been in this field for ~5 years
now starting from before the first StyleGAN paper was released. I’ve worked at
both small startups, Riot Games, and most recently Snapchat (hired as an ML
engineer although I lead an R&D team for our AR glasses these days). I
think it will be helpful to understand from first principles exactly what is
possible and how information is represented in ML.
I’m not a professional (or good) artist, but I did spend a non trivial amount of time grinding in art and I think I could offer unique perspective that most other ML engineers don’t have.
Karla
Ortiz
Miles
Vesh Miller i think what the person is referring to is imagery trained upon the
data base. And having a potential tool that could showcase everything that was
used in the process. Very difficult to implement at the moment but a proper
goal nonetheless.
Also
hi Miles! I hope all is well.
Personally
I gotta say I’m personally not having an “existential” crisis, but rather
seeing exploitative practices and wishing for these tools to not have them. In
other words, this doesn’t come from a space that hates advances, but rather is
very concerned about poor intentional practices from certain AI companies.
Also
please note, while I’m certainly no expert, I have spent almost this entire
month speaking to Ai experts, engineers and ethics advocates such as Abhishek
Gugpta , and folks from Partnership AI and they’ve agreed that there are
specific practices that are indeed exploitative and deeply problematic in the
industry.
But
how about this rather than see this as a lion’s den (and if you’ve been to my
show you know how nice we really are 🙂 ) why not instead
see it as an opportunity to understand the root of our concerns (potential
exploitation of our data, both creative and private) and perhaps state what
things you think could address our concerns?
Policy
isn’t a scary thing btw, its just good rules of engagement for companies to
follow.
So
what in your industry do you see could be done better? What do you think could
be a policy that could ease artists concerns? Public domain? Licensing?
Anything more specific?
Let us know! Also hope all is well!
Miles
Vesh Miller
Karla
Ortiz the imagery in the training set for stable diffusion is known and public.
It’s a subset of LAION. All images in that dataset could be argued to have
“gone into” any output, because the model is learning extremely high
dimensional abstract features from an enormous corpus of raw signal.
Didn’t
mean to project anything about a crisis for you personally btw - sorry if it
came off that way. I lead a small team at Snap working in a different space and
my 3D artist has been having a difficult time with the rapid pace of
advancement and the feeling that all of her grinding is being made irrelevant.
Some of my best friends are industry artists I met when I was a game designer
on League at Riot, and I’ve had a lot of heartfelt convos with them.
I’m
not going to pretend I truly understand what artists are going through right
now.
Training
data licensing is a noble cause baseline, but as I mentioned in a different
comment it’s ultimately not going to stop what I imagine is the core problem
for working artists - having their jobs be replaced.
If
there are sides to be chosen in this whole generative AI arms race, I tend to
fall pretty hard in line with Stability AI’s philosophy. My concern is that big
companies funded by billions of dollars will ultimately be the ones capturing
the value. Stability AI publishing the actual weights of Stable Diffusion was a
*big deal* to me. Now more than ever, I think we need to keep progress in the
hands of the people and not in a few big corporations.
On the artist side though, I think the battle being fought might be the wrong one. It’s not my place to say, but trying to ban prompts isn’t going to get anywhere. I’m happy to delve into why, but that deserves its own space. The battle I think we need to be fighting is for Universal Basic Income, or some equivalent that guarantees as jobs are made automated that people can still live and pursue their passions.
Miles
Vesh Miller
Karla Ortiz also I’ve always appreciated your approach in conversing around difficult topics. Your posts are definitely a much safer space to engage in this type of discourse than anywhere else right now.
Karla
Ortiz
Hi
Miles.
Thank
you for your post.
I
do have to address some misconceptions here.
It’s
important to know Stability Ai funded LAION to be created. Its also important
to know LAION 5B was originally for research which is why it could get away
with scraping data from places like Pinterest, Shopify and so on, some of them
, like Shopify 11 million times. (You can see the data on ONE of the data sets
of Laion here https://twitter.com/ummjackson/status/1565136733250809857... ) .
There
are two major issues i see, the raw data it learns from is littered with not
only copyrighted data , but also private data (like medical record photos). It
was taken without consent, permission, or compensation.
The
process the machine utilizes to “learn from” is already bunk because the source
it “learns from” allows for copyrighted and private data. I’ve seen too many
examples now where it also showcases the machine can copy and paste as well, so
how different is it really?
We
really cannot ignore or excuse that companies took this kind of data like this,
in the future forcing everyone to opt out, and taking our choice to opt in. I
find that egregious.
Secondly,
I never said I wanted to ban AI! I said I want to regulate it and it should be
regulated. I also want companies to ditch their current data sets, and transition
to public domain only, retrain their models based on public domain. That seems
the correct choice if they so wish to profit from this. And further laws will
be needed to better regulate research to commercial pipelines, what kind of
data can AI be trained on, pausing the use of people’s full names (unless an
individual want to)and provide the public tools necessary to know is it AI or
not. These are very practical stances and not at all the “i want to ban” stance
that was claimed.
And
lastly, I actually find Stability Ai very irresponsible. Emad Mostaque released
to the public and immensely powerful tool with no safety mechanisms and a
copyright/privacy data ridden data set it uses and was trained on.
He
did not think of all the potential harm like blackmail, violent imagery,
potential revenge porn, identity theft, and general use of data that isnt
theirs (copyrighted works, medical records, etc)
The
media did ask him about potential consequences and he didnt care.
https://techcrunch.com/.../a-startup-wants-to.../
Heck
I asked him and all he said was that a paper would be published soon…he
released it a couple days after and that paper never came…
And
this isnt just me.
I’ve
spoken to various and prominent Ai ethics advocates about this and they found
his actions to be single-handedly the worst thing to have happened to Ai tech
in terms of ethics. The tech was just not ready yet.
And
here’s where it pisses me off even more. Emad funded this, released it to the
world, and now when pressed about the issues the models contain, he’ll say “oh
there’s not much we can do, it’s already out there”…. My dude, HE did that! He
put the tech out there.
So
I find them to be the most irresponsible in the space . Which is why they must
be regulated…they’ve shown they do not care about consequences yet will sing
songs about creativity and how great they are, all while pocketing massive
cash, and knowing they released a problematic technology to the world..
It
didnt have to be that way. Basic safety measures could have been added that was
mindful of people’s data and potential harms .its why others in the space have
moved slowly, because this kind of tech is so impactful it cant be treated so
irresponsibly.
And
yes, we’ll need universal income regardless, but we’re not there yet.
Anyhow…Miles,
do you have any ideas on how to best regulate your industry? Anything that you
think would be important to protect people? let me know!
(End
of this exchange from the comments thread)
9.29.22
FB Post:
Amazing
article by Andy Baio that so succinctly describes one of the many big issues AI
companies have, Data Laundering! The process where academic and non profit
research, sometimes funded by corporations (like Stability Ai did with LAION),
is then utilized for commercial purposes by major corporations.
This
is a way to shield companies from responsibilities and accountability when
gathering data they normally would not have access to!
So get familiar with the term Data Laundering, because it is *precisely* what these companies are doing.
With
this link: https://waxy.org/2022/09/ai-data-laundering-how-academic-and-nonprofit-researchers-shield-tech-companies-from-accountability/?fbclid=I
9.30.33
FB Post
Last
AI post before I leave social media for the weekend (seriously we all need
breaks lol).
You
know how I know AI companies like Stability AI, knowingly and purposefully took
and used visual artists and general public copyrighted data and private data
without a care or concern (aside from seeing data on where they scraped their
data from.)?
Let
me present to you the way Stability AI has handled their music model, Dance
Diffusion.
They
built their data sets entirely out of copyright free music and allowed
musicians to opt in. Listen to this admission
“Because
diffusion models are prone to memorization and overfitting, releasing a model
trained on copyrighted data could potentially result in legal issues. In
honoring the intellectual property of artists while also complying to the best
of their ability with the often strict copyright standards of the music
industry, keeping any kind of copyrighted material out of training data was a
must.”
Where
was this care and consideration for visual artists, small businesses and the
general public? What are companies like Stable Diffusion/Dream Studio and
Midjourney doing to rectify this situation? Will they compensate affected
artists? will they shut down operations in order to comply with the standard
Stability Ai’s own music program complies with?
I
need artists to understand, these folks are exploiting your work and your
rights. And the only reason they do it to us is because we assume “its bound to
happen” or we play into the game of “oh its just a tool” when its more than
that. Its companies deliberately taking, using and profiting off your data.
That they can abide by more ethical practices elsewhere shows us that they
could have done that with us…but they chose not to.
Further
resources:
Link
to Dance Diffusion where screenshot was grabbed from:
More
info on Dance Diffusion data sets:
This ends my summary of Karla's posts from July - Sept on AI.
Thank you Karla for putting this and other important artist life topics on the radar.
Karla's thoughts on NFTs are featured in the 12.21 update on this blog post from 4.21
https://stuartngbooks.blogspot.com/2021/04/nfts-in-news-case-studies.html
No comments:
Post a Comment