Building a "no AI allowed" art auction site, mostly because I got annoyed
So I've been building this thing called Atelier for a few weeks now. It's a live auction marketplace, but for one specific kind of art: 100% human-made, physical pieces only. No AI art, no digital-only assets, no prints-of-prints. Every listing has to be studio-verified. The artist has to show their actual process, not just upload a jpeg.
The whole thing started from being annoyed, honestly. I kept seeing "art marketplaces" where you genuinely couldn't tell if a piece was hand-painted or generated in an afternoon, and it felt like nobody was building specifically for the opposite of that. If you're a collector, you're not just buying a picture, you're buying the fact that a real person made a real thing with their hands. That story is the actual product. Most platforms treat it as an afterthought.
Where it's at right now:
Live auctions with real bidding, anti-sniping protection, reserve prices
Artist studio pages: bio, process video, the actual philosophy behind their work
A verification system where "studio verified" actually means something (a human reviews it, it's not just a checkbox the artist ticks themselves)
Where it's NOT at yet, to be upfront: payment is still faked for testing, and I haven't onboarded a single real artist yet. It's a live demo, not a live business. I've been finding and fixing real bugs almost daily. Today alone I caught a filter that silently broke for half the categories, and an ownership bug that locked one test artist out of their own listings. Building this in the open, warts and all.
What I'm actually hoping for from this post:
If you're an artist (or know one) who's tired of AI slop crowding out real work, would something like this even matter to you? What would actually make you trust a platform enough to list a real piece?
If you're a collector, what's stopped you from buying art online before? Trust? Provenance? Just never occurred to you?
Know anyone in art communities (Discord servers, local collectives, whatever) who might want to poke at an early demo and tell me what's wrong with it? I'd rather hear "this is dumb because X" now than after I've built more on top of it.
Not asking anyone to bid real money or list anything yet, genuinely just want to know if this resonates with anyone outside my own head before I go further.
Happy to drop the link if anyone wants to look, or just reply here. Thanks!
I actually think the provenance angle is the strongest part here. “Human-made” alone is a filter, but showing the artist’s process and story is what makes a collector care. The hardest part will probably be building trust and getting the first group of artists onboard, but the idea definitely resonates. Would love to check out the demo.
Since the ambiguous middle is the part you said you do not have a clean answer for, here is the line that I think actually holds up under pressure: ask whether the tool originated any part of what the buyer is looking at.
A reference image did not originate the painting, the painter did, and the reference never reaches the buyer. Removing dust from a photograph of a real oil painting did not originate anything either, it just stopped the photo from misrepresenting an object that already exists. But generative fill that invents a corner of the canvas has authored part of the thing being sold, even though it is a small part, and even though the artist did everything else by hand. That test also handles the bio case cleanly, because the bio is not the artwork. It is marketing copy about the artwork, and nobody is bidding on the bio.
The useful property of framing it that way is that it scales to tools that do not exist yet. Any rule written as a list of banned software goes stale the moment Adobe ships something new. A rule about authorship of the final artefact keeps working, because you apply it to the output rather than to the toolchain.
One practical note on publishing it: state the principle first and give examples second, explicitly labelled as illustrations rather than as the exhaustive list. Otherwise the first artist with a case not on your list will reasonably argue it is therefore permitted.
This is the cleanest version of the answer I've seen, and I think it actually resolves the thing I said I didn't have a clean answer for. "Did the tool originate any part of what the buyer is looking at" is a real test, not just a vibe. The reference image example makes it click: the painter originated the painting, the reference never reaches the buyer, so it's a non-issue. Dust removal is the same logic, the object already existed, the photo was just being corrected to represent it accurately. Generative fill is different in kind, not degree, because now something in the actual thing being sold was authored by a tool, even if it's one corner of the canvas and a human did everything else. And you're right that the bio question dissolves the moment you separate "the artwork" from "marketing copy about the artwork." Nobody's bidding on the bio.
I also think you're right about the publishing structure, and it's not a small point. A banned-tools list ages badly by definition, someone ships a new feature and the list is instantly out of date or has to be argued about. A principle about authorship of the final object doesn't have that problem. I'll write it up as: state the rule, then a handful of examples explicitly labeled as illustrations, not the full list, so the first real edge case doesn't get treated as automatically permitted just because it wasn't named. Genuinely appreciate you working through this with me, this is going straight into how I actually write the policy.
Thanks, genuinely glad it resonates. Being upfront though, it's still early enough that I'd rather set expectations than oversell it, payment is faked for testing right now and I haven't onboarded a real artist yet. Happy to send the link if you want to poke around and tell me what feels off. Are you coming at this more as a collector, or just curious how the thing works?
Love the "verification isn't a checkbox the artist ticks themselves" bit — that's the part most marketplaces fake. Building something in a similar spirit myself (anti-AI-flattening, just for writing not art). Good luck with the bugs, would poke at the demo if you drop the link!
Thanks, and yeah that line matters a lot to me, most "verification" I've seen elsewhere is really just an honor system with a badge on it. Curious what your writing project looks like, is it tackling the same "prove a human actually made this" problem, or more about something else in that space? Happy to send you the Atelier link either way, would genuinely value someone poking at it from outside my own head.
I think the interesting part isn't even the "no AI" angle; it's trust. If collectors are paying for original physical artwork, they need confidence they're buying something authentic, and your verification process seems to be addressing that. I'd be curious to see how you'll keep that trust as the platform grows. Human review works well early on, but maintaining consistent standards at scale is a tough challenge. Looking forward to seeing how Atelier evolves.
Honestly, I don't have a great answer to this yet, and I'd rather say that than pretend I do. Right now it's just me reviewing, which works at 5 artists and obviously doesn't at 500. My rough thinking is some combination of a written rubric so it's not just my gut each time, spot-audits on already-approved listings rather than only front-loading everything at submission, and eventually more than one reviewer so it's not a single person's judgment call. But I haven't built any of that, it's still just me looking at a process video and deciding. Consistency at scale is the real unsolved problem here, not a detail I've worked out and forgot to mention.
very much needed as we are all drowning in a sea of AI art
Thank you, that's exactly the frustration this came out of.
“No AI allowed” is a clear positioning promise, but the operational definition seems like the hard part.
I’d decide early what the platform is actually certifying: an artist’s declaration, evidence of the creation process, or verification performed by the platform. Trying to guarantee perfect detection could create a lot of friction for honest artists while still missing determined bad actors.
A lighter starting point might be artist attestation, clear evidence requirements for disputes, and random audits.
How are you planning to handle a buyer challenging an artwork’s authenticity after an auction has already closed?
This one I don't have a full answer for either, but there's a piece of the architecture that already helps: payment sits in escrow until the buyer confirms delivery, so there's a real window after the auction closes but before money actually changes hands where a dispute could pause that release rather than trying to claw something back after the fact. What I don't have yet is an actual defined process for what happens inside that window, who reviews the claim, what evidence would settle it, how long a buyer has to raise it. Your lighter-model suggestion, attestation plus clear evidence requirements plus random audits rather than trying to guarantee perfect certainty up front, feels closer to the right shape than anything I'd come up with, mostly because trying to make verification airtight would just push all the friction onto honest artists while a determined bad actor still finds a way through. I think this needs to be written down before it's needed for real, not improvised the first time someone actually disputes a sale.
That escrow window changes the problem quite a bit. Instead of needing perfect verification before the auction, you mainly need a predictable dispute process before the funds are released.
For a first version, I’d keep it deliberately narrow: the buyer has a fixed period to raise a claim, the claim has to point to something specific, the artist responds with predefined evidence such as sketches, source files, progress photos, or dated work-in-progress material, and the reviewer only decides whether the evidence meets the platform’s policy — not whether human authorship can be proven with absolute certainty.
You could also phrase the promise as “human-made under our verification policy” rather than implying the platform can guarantee that no AI was involved. That feels enforceable without overpromising.
And yes, writing the process before the first real dispute matters, because that first case will probably become the precedent for everything after it.
Disclosure first, since it is relevant here: I work at Ojin, an AI company. We build real-time conversational avatars, which are explicitly synthetic and disclosed as such, so I am not quite the thing you are annoyed by. But sitting on the generation side is why I think the most useful thing I can offer is where this gets hard rather than encouragement.
The architectural call you have already made correctly, and should probably never revisit: you verify process, not artifact. Do not bolt an AI detector onto the uploaded image, ever. Those tools are unreliable, and the failure mode that actually matters is not a generated piece slipping through, it is falsely flagging a real artist’s real painting. One public incident of telling a working painter their own work looks machine-made will cost you more trust than a hundred quiet false negatives. Human studio verification sidesteps that completely, and it is the reason the model holds up.
The constraint doing most of your heavy lifting is one you almost mention in passing: physical pieces only. That is a far stronger guarantee than any verification layer, because the item being auctioned is an object that demonstrably exists. It reduces provenance to "does this thing exist, and did this person make it," which is a photography and logistics problem rather than an AI problem. I would foreground that harder than the no-AI framing, since no-AI describes an absence and physical-object describes a thing.
The part worth deciding before you onboard artist number one, because it will not stay theoretical: AI-assisted is not AI-generated, and the binary will break on a real case. An artist who used a generated image as a compositional reference. Someone who ran generative fill to clean dust off the photograph of their oil painting. An artist whose bio was drafted with ChatGPT. None of those are slop, and all of them technically involve AI. Whichever way you rule, the first ambiguous case is a real person you either publicly reject or publicly accept, and that becomes your policy whether you wrote it down or not. Most no-AI policies do not die on obvious slop, they die right there.
On your actual question about what would make an artist trust it: likely that the verification is appealable and the criteria are published before they submit. Artists have been algorithmically flagged and delisted elsewhere with no recourse, so a human reviewer only reassures if there is a named process standing behind it.
Really appreciate you laying this out, especially the disclosure up front, that actually makes the rest land better, not worse.
The detector point hits home. I hadn't fully articulated it that way but you're right, the asymmetry is the whole thing. A generated piece slipping through once is a bad day. Telling a real painter their own work looks machine-made is the kind of thing that gets screenshotted and never forgotten. I wasn't planning to bolt one on, but "never" is a better rule than "probably not," so thank you for putting it that plainly.
The physical-object reframe is genuinely useful and I think you're right that I've been underselling it. "No AI" is a negative claim, it's defined by what it excludes. "This object exists and this person made it with their hands" is a positive one you can actually point a camera at. I think I led with the AI framing because it's the more emotionally charged hook, but you're right that it's not the part doing the real work.
The ambiguous case is the one I don't have a clean answer for yet, and I'd rather say that than pretend I do. My gut says reference images and cleanup tools (getting dust off a photo of a real oil painting) are obviously fine, and a fully generated composition obviously isn't, but I know the actual hard cases live in the middle of that, not at the edges. I think you're right that I need to write the policy down before artist number one shows up with a real edge case, not after, because you're also right that whatever I decide in that first real moment becomes the policy whether I meant it to or not.
On appealable and published before submission, that's fair and I don't have it built yet. Noted as something that needs to exist before I'd feel good onboarding a real artist, not just as a nice-to-have.
The interesting part isn't only excluding AI — it's turning human authorship and process into something collectors can trust.
What would convince you that buyers are choosing Atelier because of verified human creation, rather than because they simply want better provenance and confidence when buying art online?
That's a sharper question than the one I've been asking myself, honestly. I don't think I have a real answer yet. My honest guess is I can't actually tell the difference from the outside until I see it in a specific moment, like a collector choosing a verified piece over a similar unverified one at a similar price, or being willing to pay more specifically for the ones with a strong process story. Right now I don't have either of those data points. I think the test is something like: show the same piece with and without the studio verification and story, and see if behavior actually changes. If it doesn't, then you're right that people just want confidence in general and the human-creation angle is more of a marketing frame than a real driver. I'd rather find that out now than after building more on top of an assumption.
This comment was deleted a month ago
The insight that "the story is the actual product" is really sharp here. Most art platforms treat artist bio and process like checkbox requirements, but you're flipping that - making verification and process the differentiator. Have you talked to any collectors about whether they'd actually pay more for pieces where they can see the studio process and artist philosophy? That pricing signal would tell you if the provenance premium is real enough to sustain the business, since your cost per listing (human review) is going to be higher than just spam-checking checkboxes.
No, I haven't, and reading your comment I think that's a real gap rather than something I can keep deferring. I've been treating the story/process angle as the differentiator without actually pricing it, which is backwards given you're right that human review is a real cost, not a free checkbox. I think the honest next step is literally asking a handful of real collectors "would you pay X more for this piece because you can see the process," not asking if they like the idea in the abstract. Abstract "yes I like transparency" answers are cheap, an actual price point is the real signal. Adding that to what I ask people directly, thank you for pushing on it.