A Beader’s Guide to Spotting AI-Generated Beadwork
(Don't forget that you can click on images to enlarge them!)
Let me also mention, that while environmental impact from AI is a major concern, it is not what this particular blog post is about. 😉
Websites we thought we could trust, such as Amazon, Etsy, eBay are riddled with mass produced AI beadwork kits, books, and tutorials.
And there are entire groups on Facebook dedicated to 100% AI generated "beadwork", touted as inspiration.
Not sure about how you obtain your inspiration, but for me I'd prefer it be from something actually doable!
And not just for newbies who may not yet have the experience to know what beads can and cannot physically do. It's becoming harder for experienced beaders, too, as AI image generators become more and more advanced...like, by the second!
It's most certainly going to be hard for us to keep up.
And my biggest fear is that as these generators learn how to create more realistic-looking images, there is going to come a time when we simply can't tell what they've created from what an actual, living, breathing human beadwork artisan has created.
That is going to be a very sad day! And creates a huge potential for financial loss, both for the buyer being duped by something unobtainable, but also to designers who will eventually suffer the consequences of beaders giving up hope! 😢
But for now, I thought I'd compile a little (okay! a LOT) information that will hopefully help us determine the real deal from the AI-generated.
First, there are the blatantly obvious images
where beads are doing things that beads just can't do.
Where's the thread???
Those are the easy ones. Unfortunately, they're not all going to be that easy.
1. Follow the Thread Path
This may be one of the best tests we have as beaders.
Pick one small section of the piece and ask yourself:
“Could I actually stitch this?”
Hint...no, you can't! 😉
In real beadwork, there has to be a physically possible thread path. Beads don't simply attach themselves to one another because that's where they'd look pretty.
Look for:
- Beads that appear connected without any possible thread path.
- A row that begins as peyote and then inexplicably changes structure.
- Beads sitting in positions where they couldn't possibly be supported by the surrounding beads.
- Connections that disappear behind another bead and never logically emerge.
- Components that seem to merge into one another.
- Impossible joins between motifs.
AI can be very good at creating the appearance of intricate beadwork without necessarily understanding how that beadwork would actually have to be constructed.
So don't just look at the design.
Try to mentally bead it.
If you can't figure out how a needle and thread could possibly travel through what you're looking at, take a closer look.
2. Look for Beads That Morph Into Other Beads
This is one of those things that can be really easy to miss
Instead, zoom in and follow a single row of beads.
You may find something that starts out looking like
a seed bead and gradually becomes:
- a crystal
- a pearl
- a drop
- part of the background
- half of another bead
- some strange little blob that isn't any bead you've ever seen in your life
You may also see two neighboring beads partially fused together.
Real beads can overlap visually in a photograph, of course.
But they're still separate physical objects.
A Delica doesn't decide halfway through the necklace that today it identifies as a bicone.
3. LOOK AT THE HOLES!
AI-generated beadwork may contain:
- Holes in impossible locations.
- Some seed beads with visible holes and supposedly identical beads without them.
- Holes that don't correspond with the orientation of the bead.
- Thread entering or exiting somewhere other than a hole.
- Beads that appear to be drilled sideways when that bead is actually drilled lengthwise.
- Two-hole beads with inconsistent hole placement.
Sometimes what looks like a bead hole isn't actually a hole at all. It's basically a dark spot that the generator has learned belongs on something that looks like a bead.
Those of us familiar with beads such as SuperDuos, GemDuos, Tilas and other multi-hole beads have an advantage here.
We know where those holes actually are.
If you see what appears to be a thread or connection coming out of an area of the bead where there is no hole...
Well...
Nope.
Beads haven't learned how to violate the laws of physics yet.
4. Does That Bead Actually Exist?
This is another very bead-specific giveaway.
AI can create a gorgeous little object that looks sort of like:
- a SuperDuo
- a Ginkgo
- a GemDuo
- a PaisleyDuo
- a Czech glass flower
...except it isn't actually any of them.
AI can essentially invent manufactured beads.
And sometimes the giveaway isn't the bead itself, but the fact that supposedly identical beads throughout the design aren't actually identical.
One is pointed. The next has mysteriously become rounded.
Manufactured beads have variations, we certainly all know THAT! but they aren't individually reinvented every time they're placed into a design.
If you don't recognize a bead, that doesn't automatically mean its AI. There are thousands upon thousands of beads out there that we haven't all seen.
But it might be worth asking the designer:
“What bead is that?”
5. Examine the Fringe Carefully
Fringe can apparently be an AI playground.
A beautiful sweep of fringe may look completely
convincing until you stop admiring the whole thing and
trace one individual strand from beginning to end.
Look for strands that:
- disappear
- merge together
- divide into two
- terminate for no apparent reason
- emerge from nowhere
- pass through another strand
- contain impossible bead sequences
- attach to the base in impossible places
The overall fringe may look absolutely spectacular while no individual strand could actually be made the way it's shown.
So, pick a strand. Start at the top. And follow that puppy all the way down. How's is not all falling onto the floor?
6. Look at Bezels Around Cabochons and Crystals
This is another excellent place to investigate. We don't have an endless selection of bead shapes and sizes, and certainly not an array of them to make a curve.
If it's Peyote, for example, the bead relationships
and decreases have to actuall work.
Something has to secure that cabochon or crystal.
AI can produce a beautiful-looking ring of beads that:
- randomly changes bead count
- contains impossible decreases
- changes from one stitch structure into another
- has beads partially embedded inside the stone
- has no plausible means of holding the cabochon in place
To a non-beader:
Gorgeous!
To a beader:
Wait a minute. How the heck is that stone actually being held in there?
That's exactly the kind of question we should be asking.
Other Telltale Signs
There are plenty of other things that should at least make you stop and take a second look.
Missing body parts!
Egads! They've chopped off their arms!
If there's a model wearing the jewelry, count the fingers. Look at the ears. Look at where the necklace actually meets the neck. AI has gotten considerably better at people, so this isn't the giveaway it once was—but bizarre anatomy is still worth noticing.
Not to mention...could those AI models be anymore flawless. I mean, c'mon! Freckles aren't that bad, are they?
It basically looks too good to be true.
Now, be careful with this one!
There are some phenomenal bead artists out there creating work that absolutely DOES look too good to be true.
So “Wow, that's amazing!” is certainly not evidence of AI.
But combine that with impossible construction, mysterious beads, nonsensical thread paths and other inconsistencies?
AI loves color and sparkle and can throw an extraordinary number of subtly different colors into something that supposedly uses manufactured beads.
The stitches mentioned don't match the image.
If the listing says you're buying a peyote-stitch tutorial and you're staring at something that clearly couldn't have been made using peyote stitch...ask yourself why!
So...Who Do We Buy From?
Now, what I would like to say—because it would make this whole thing super simple—is:
Just buy from established designers you recognize.
Buy from people you've purchased from before. Designers whose names you've seen frequently on Planet Bead. People you've watched teach. People whose work you've seen other beaders make.
But...
That's really not fair to new and up-and-coming designers.
Every established designer was an unknown designer at some point.
And I think new designers deserve our love and support, too. 💜
So rather than automatically deciding that an unfamiliar designer must be selling AI-generated work, I'd suggest stopping for a moment before clicking that:
BUY ME NOW BECAUSE YOU LOVE ME BUTTON
...and send the seller an email or message first.
Questions to Ask Before Buying a Kit or Tutorial
“Can I see another photo of the finished piece?”
Ideally, ask for the back, side, clasp or another angle...not simply another beauty shot.
Someone who possesses the finished piece should normally be able to photograph it.
“Can you show me a photo of the back?”
This can be especially useful with bead embroidery.
AI can create a spectacular front without ever having to figure out how the piece is actually constructed and finished on the back.
“Do you have a photo of the piece being worn?”
For jewelry, this can help establish that it's a physical object while also showing its actual size and drape.
And we all know our average beady designer does NOT look like those flawless AI models!
“What stitch is used to make this section?”
A designer should be able to explain whether it's peyote, RAW, CRAW, herringbone, bead embroidery, netting, etc.....including areas that aren't immediately obvious.
“What beads are used here?”
This can be particularly revealing when an unusual bead shape appears in the image.
Ask for the manufacturer, bead name and size.
“What size is the finished piece?”
Then compare those measurements with the known sizes of the beads shown.
If the proportions make absolutely no sense, that's another reason to investigate.
“Are the photographs in the listing photographs of the actual finished project?”
I really like this question because it doesn't accuse anyone of anything....but the answer could be VERY telling!
“Did you personally make the sample shown in the listing?”
Again, straightforward and reasonable.
“Has this tutorial been test-beaded?”
I think this is an especially good question.
Has another beader successfully completed the project by following the instructions?
“Are there step-by-step photographs or diagrams showing the construction?”
A tutorial consisting primarily of written instructions accompanied by one spectacular rendered image would certainly make me ask more questions.
“If I have trouble with the instructions, do you provide support?”
Someone selling a design they actually developed should understand its construction well enough to help when a customer gets stuck.
If I Could Ask Only ONE Question...
It might be this:
“Before I purchase, could you please send me a quick photo of the finished piece from the back or from a different angle than the photos in the listing?”
That doesn't require the seller to defend themselves against an accusation of using AI.
You're not saying:
“PROVE YOU'RE NOT A SCAMMER!”
You're simply asking them to demonstrate that the object they're selling instructions to make actually exists.
Now, granted...
Someone who's deliberately trying to deceive buyers can lie in response to some of these questions.
So none of this is foolproof.
And THAT is why I think asking for an additional photograph is so valuable.
One Important Distinction
I want to make this really clear:
Using AI isn't necessarily the problem (because we're not discussing environmental impact in this blog post, remember!) . Selling instructions for an object that has never actually been constructed IS.
A designer might legitimately use AI tools to remove a background, create advertising graphics, improve the presentation of a photograph, help edit written instructions, or create beautiful lighting.
Personally, I don't think that's the same thing as generating a spectacular image of imaginary beadwork and then selling a “tutorial” that supposedly teaches another human being how to make something that the designer has never actually made themselves.
There's also a gray area that we're going to have to become more aware of:
A photograph may begin with real beadwork but then be altered with generative AI.
If the AI only changes the background, that's one thing.
But if it starts “improving” the actual beadwork—adding beads, changing shapes, fixing symmetry, inventing fringe or reconstructing portions of the piece—then the photograph may no longer accurately represent the project being sold.
That's something buyers and designers alike need to think about.
This is a perfect example of that using one of my projects. The image on the left is the actual bracelet, laying flat on a plain background. AI stood that baby up, balanced precariously on the beads and added some beautiful lighting! Jealous of that lighting? Yeah, me too! But...if I didn't let me viewers know, I'd be deceptive...because the bracelet doesn't actually look like that!
What About AI Detectors?
I wouldn't put too much faith in a website that tells you:
“This image is 97% AI-generated!”
Current AI-image detection is not foolproof, and as image generators change, detection systems don't necessarily generalize reliably to new generators.
There are newer methods of establishing an image's provenance—basically information about where the image came from and how it was created.
One of those is called Content Credentials, based on the C2PA standard. Some AI generators, cameras and editing tools can attach verifiable provenance information to an image.
OpenAI, for example, currently puts C2PA information and an invisible SynthID watermark into images generated with its tools and provides a verifier that can check for supported OpenAI provenance signals.
BUT—and this is important—
Finding no AI information does NOT prove an image is real.
Metadata can disappear when an image is resized, converted, uploaded, downloaded or screenshotted, and watermarks can sometimes be degraded. Even OpenAI specifically warns against treating the absence of those signals as proof that an image wasn't AI-generated.
So we're right back to using our most valuable tool:
Our knowledge as beaders.
Below are some examples of real beadwork vs. AI-generated imagery.
Take your time looking at them.
Don't just ask yourself:
“Does this look real?”
Instead ask:
“Could I actually bead this?”
Because, at least for now, that may be one of the best AI detectors we've got.
Here is an amazing example of one of Sheila Root's beaded birds. Many of us recognize her work and while it is outstanding and a bit intimidating for some, it can, fact, be beaded!
But those little guys from that mass produced kit seen splattered all over the internet...yeah, good luck! Not only does the kit not contain the beads you need, but the instructions are dismal, incomplete, and not even beading stitches.
Hopefully you're starting to get an idea of how unrealistic AI looks?
Plus when you look closely enough, there are made up beads, and all the other issues mentioned above.
Happy Real Beading!
Love, Beki
No comments:
Post a Comment