Where Every Major Art Platform Stands on AI Art in 2026
Which art sites ban AI-generated work? Which require labels? And which platforms train AI on your uploads? A plain-English guide to AI art policies in 2026 — from DeviantArt and Pixiv to Cara, ArtStation and artbase.
Few topics move artists between platforms faster than AI policy — and few are as confusingly reported. Most arguments actually mix up two separate questions:
- Can AI-generated images be posted here? (a content policy question)
- Will my artwork be used to train AI models? (a data policy question)
A platform can ban AI art and still leave your uploads scrapeable; another can allow AI art while defending your files from training datasets. This guide answers both questions for every major platform, in plain English.
Last checked: 16 July 2026. Policies in this space change quickly — we’ll keep this post updated, and the date here will tell you how fresh it is.
The short version
| Platform | AI art allowed? | Labelling rules | Your art vs. AI training |
|---|---|---|---|
| DeviantArt | Yes | Label required (“created using AI tools”) | Opted out of third-party datasets by default |
| Pixiv | Yes | Mandatory “AI-generated” flag; unflagged work can be auto-flagged | FANBOX bans AI content entirely |
| Cara | No | — | Auto “NoAI” tags + built-in Glaze |
| ArtStation | Yes | ”CreatedWithAI” tag required for marketplace items | Opt-in “NoAI” tag bars training use |
| Newgrounds | No (Art Portal) | — | — |
| artbase | Yes | Label required on all AI work | Never used to train AI, by policy |
| Instagram / Meta | Yes | Platform applies “AI info” labels | Trains on public posts; limited opt-outs |
| X (Twitter) | Yes | None | Grok trains on public posts by default |
| Bluesky | Yes | None | Platform won’t train; open network is still scrapeable |
| Tumblr | Yes | None | Per-blog “Prevent third-party sharing” toggle |
Now the detail, because the detail is where the surprises are.
Dedicated art platforms
DeviantArt — allowed, labels required, opted out of training by default
DeviantArt permits AI-generated work but requires it to be labelled as created using AI tools, backed by an automated detection model for work that isn’t declared, and browse settings that let members dial down how much AI art they see. On the data side, DeviantArt made every deviation opted out of third-party AI training datasets by default and introduced the noai / noimageai directives that tell scrapers to stay away — though, like all such directives, they rely on scrapers choosing to honour them.
Pixiv — allowed, mandatory flags, and a 2026 crackdown
Pixiv requires any work whose creative process was entirely or mostly AI to carry an AI-generated flag, and reserves the right to apply the flag itself when creators don’t. Guideline updates that took effect in March 2026 went further: tag abuse (including unflagged AI work) is explicitly prohibited, and works judged likely to violate the guidelines can be hidden from search. Pixiv’s creator-support service FANBOX bans AI-generated content outright.
Cara — banned outright
Cara is the hard-line option: no AI-generated portfolios, period, until it considers the ethical and data-privacy issues resolved by regulation. It pairs the ban with detection tooling, automatic “NoAI” tagging on uploads, and an integration with Glaze to protect work against style mimicry. This stance is the platform’s founding identity, and its community holds it firmly.
ArtStation — allowed, with an opt-out tag for artists
ArtStation allows AI-generated work (a stance that sparked large artist protests when announced) and requires marketplace products made with generative AI to carry a CreatedWithAI tag. For artists on the other side, the NoAI tag marks your projects as off-limits to AI training, with the ban written into ArtStation’s terms of service. Note that NoAI is opt-in — untagged work doesn’t get that protection.
Newgrounds — banned in the Art Portal
Newgrounds bans AI-generated images from its Art Portal. Minor, incidental AI elements inside substantially human-made work are tolerated, but AI pieces as the artwork itself are not welcome.
artbase — allowed, but it must be labelled, and never used for training
artbase takes the transparency position: AI-generated and AI-assisted work is permitted, but it must be clearly labelled when uploaded. Nobody browsing should ever have to guess whether they’re looking at human-made work — labels keep discovery honest and let the community engage with each kind of work on its own terms.
On the data side, artbase’s policy is unambiguous: member artwork is never used to train AI models, full stop. No opt-out toggle to hunt for, no default to check — it isn’t on the table in the first place. artbase is open to everyone, and as the platform grows this policy will be kept current here.
Social platforms
Social media flips the question: posting AI art is broadly allowed everywhere, and the real issue is what happens to your uploads.
Instagram / Meta — trains on public posts, opt-outs are limited
Meta trains its AI models on public posts, and 2026 has sharpened artists’ concerns: July’s launch of its Muse Image model let anyone use photos from public accounts as references for AI generations — opted in by default, with no notification. After backlash led by SAG-AFTRA, Meta walked back part of the feature within days, but the underlying training practices remain. If you stay, review the “Sharing and reuse” controls in your account settings, and read our fuller assessment of Instagram for artists.
X (Twitter) — Grok trains on your posts unless you opt out
Public posts on X are used to train Grok by default. You can opt out — Settings → Privacy and safety → Grok — but the setting is only reachable on the web version, and opting out doesn’t unwind training that has already happened.
Bluesky — won’t train on your work, but can’t stop others
Bluesky has said it won’t train generative AI on user content, which makes it unusual among social platforms. The honest caveat: it’s an open network, so public posts remain technically scrapeable by third parties, and proposed user-consent signals will depend on scrapers respecting them.
Tumblr — opt-out toggle, per blog
Following reports of content-licensing deals with AI companies, Tumblr added a Prevent third-party sharing toggle. It’s off by default and set per blog, so if you run multiple blogs, each needs switching individually.
Protecting your work, whatever platform you use
A few defences travel with you everywhere:
- Glaze and Nightshade — free tools that alter images imperceptibly to disrupt style mimicry and training. Glaze is built into Cara and usable before uploading anywhere.
- Opt-out tags and toggles — set them wherever they exist (they take minutes), but treat them as polite requests rather than locks.
- Watermarks and resolution limits — old-fashioned, still useful: post at display resolution, keep print files private.
- Platform choice — ultimately the strongest lever, which is why this post exists.
FAQ
Which art sites ban AI art completely?
Cara and Newgrounds ban AI-generated work outright, and Pixiv’s FANBOX bans it for creator monetisation. DeviantArt, ArtStation, Pixiv and artbase allow it with mandatory labels or tags.
Does posting my art on social media let companies train AI on it?
On X, yes by default (with an opt-out); on Meta platforms, public posts are used for training with limited opt-outs; on Tumblr, sharing with third parties can be disabled per blog; Bluesky says it won’t train on your content itself, but its open network can be scraped by others.
Do “NoAI” tags actually stop scraping?
Not technically — they’re signals, honoured by cooperative scrapers and written into some platforms’ terms, but there’s no enforcement mechanism in the file itself. Pair them with Glaze or Nightshade if this is a priority for you.
What exactly is artbase’s AI policy?
AI-generated and AI-assisted work is allowed but must be clearly labelled at upload, so viewers and the community always know what they’re looking at. Unlabelled AI work isn’t permitted, and on the data side, artbase never uses member artwork to train AI models — there’s no opt-out needed because it’s simply not done.
Final thoughts
The platform landscape has settled into three camps: ban (Cara, Newgrounds), allow with labels (DeviantArt, Pixiv, ArtStation, artbase) and allow with your data on the table (most social media). Which camp suits you is a personal call — but it should be an informed one, made on current policies rather than year-old headlines.
We’ll keep this page updated as policies move. If a labelled, transparency-first community sounds like your camp, artbase is open to everyone now.