Opt Your Photos Out of AI Training: Every Control That Actually Exists

🔎 Reverse Image Search

Image search techniques made simple: upload a photo or paste a URL and search 6 engines at once. 100% free, nothing stored.

🖼️ Drag & drop an image, or click to browse JPG · PNG · WEBP · GIF · BMP — Max 8MB

🔒 Your image is never stored on our servers longer than 24 hours and is never shared.

To opt out of AI training photos, use each platform’s own control: object through Meta’s form for Instagram and Facebook, keep DeviantArt’s default NoAI label switched on, set Flickr and 500px photos to private or All Rights Reserved, and block training crawlers in your own site’s robots.txt. No switch covers everywhere, and opt-outs only affect future training, not models trained.

That is the honest starting point for this guide. There is no single button that removes your photographs from every AI training dataset, and anyone who promises one is overselling. What does exist is a set of real, working controls — one per platform, plus file-level signals and crawler rules for your own website — and used together they cut your future exposure sharply. This guide walks through every control that actually exists in November 2026, platform by platform, with the limits stated plainly. One boundary to note up front: this article covers prospective controls, meaning steps that stop future training use. If you want to find out whether your photos were already used to train an AI model, that is a different job, and our guide on how to check if your photos were used to train AI covers it.

What Does It Mean to Opt Your Photos Out of AI Training?

An AI training opt-out is a setting, form or file-level signal that asks a company not to use your photos to train its AI models.

That one sentence matters, because the word “opt-out” hides three different things. A platform setting tells the service hosting your photos what it may do with them — Meta’s objection form is one example, and DeviantArt’s NoAI label is another. A file-level signal travels with the image itself, such as a noai meta tag or a Content Credentials preference embedded in the file. A crawler rule in your robots.txt file tells automated visitors to your own website what they may collect. Each layer reaches a different audience, and none of them reaches everybody.

Four questions will keep you from being misled by any single control:

Whose systems are covered? The host itself, a named partner, or any crawler that visits your site?
What use is covered? Training a model is not the same as product improvement, safety checks, search indexing or recommendations, and a setting that covers one may not cover the others.
Who is eligible? Your country, your account type and whether the photo is public can all change the answer.
When does it apply? Almost every control is prospective: it affects future collection, not material already copied or already used to train a model.

Keep those four questions in mind as you work through the platforms below, and check the current wording on each service before you rely on it, because settings and policy pages move without warning.

Can You Really Stop AI Companies Training on Your Photos?

No, not completely — and any guide that says otherwise is misleading you. A photo that is public on the internet can be downloaded, screenshotted, reposted or scraped by a bot that ignores every rule you publish. What you can do is close the documented, lawful routes one by one, so that the companies which do follow the rules stop collecting your work, and the only remaining exposure comes from actors who were already breaking the rules.

That is still worth doing, for three reasons. First, the biggest trainers are the ones most likely to honour an opt-out, because they publish their controls and face regulators. Second, the law is moving in your direction: Article 53 of the EU AI Act requires providers of general-purpose AI models to put in place a policy to comply with EU copyright law, and in particular to identify and comply with reservations of rights expressed by rights holders, and those obligations have applied since 2 August 2025, according to Article 53 of the EU AI Act. A clear opt-out on your site is exactly the kind of reservation of rights that provision refers to. Third, an opt-out creates a record: a dated setting, a confirmation email or a published robots.txt rule is evidence of your intent if you ever need to complain or claim.

The limits, stated plainly:

Models already trained cannot be untrained. Once your photo has contributed to a finished model, no opt-out removes it from that model.
Signals are requests, not locks. A robots.txt rule or a noai tag only binds crawlers that choose to read it.
Partners are a separate question. A platform’s promise covers its own systems unless it says otherwise; ask whether the opt-out reaches its partners too.
Public still means copyable. Even with every control set, a public photo can be reposted somewhere you have never heard of, and scraped from there.

With that understood, here is every control worth setting, in the order that matters most for photographers.

How Do You Opt Out of AI Training Photos on Instagram and Facebook?

For Instagram and Facebook, the control is Meta’s objection form, and it is available to adults in the EU and the UK. Meta’s April 2025 announcement described plans to train its AI on public posts and comments shared by adults, and on interactions with Meta AI, and it described notifications containing an objection form for users in the EU. Public data from accounts belonging to people under 18 is excluded, and private messages are not used unless a person shares them with Meta AI.

Follow these steps:

Open the notification Meta sent about AI training, or go to Settings & privacy, then Privacy Centre, and look for the Generative AI section.
Open the objection form and submit it, giving a short reason for objecting. The form is the formal route Meta provides for the EU and UK; do not rely on a general privacy toggle, because there is no simple account switch for this.
Wait for the confirmation email. Meta says it will send an email confirming that your public content will no longer be used for future AI training, so keep that email as your record.
While you are there, reduce what is public. Switch old public posts to friends-only where you can, and consider making the account private if you do not need the reach. A private account’s photos sit outside the training scope Meta describes.
Do not mistake Instagram’s Sharing and Reuse setting for a training opt-out. That setting controls whether other people can make things from your posts using Meta’s AI features; it does not stop Meta training on your public photos.

If you are outside the EU and the UK, Meta’s own pages describe no equivalent form for your own public posts, so your practical levers are the ones in step 4: make less public, delete what you no longer need online, and restrict what logged-out visitors can see. The form has also been reported as available in Australia since July 2025, so check the Privacy Centre shown for your region rather than assuming the EU steps apply to you.

How Do You Opt Out on Flickr, 500px and DeviantArt?

Photo platforms differ sharply. One gives you a genuine default opt-out, one gives you visibility and licence controls but no AI switch, and one gives you no documented AI control at all. Here they are, honestly, one by one.

Flickr: control the licence, because there is no AI switch

Flickr does not offer a one-click AI-training opt-out in its account settings, so the levers are visibility and licence. The licence matters more than most photographers realise: IBM’s Diversity in Faces dataset was built from around a million Creative Commons-licensed photos downloaded from Flickr, and IBM said the images could be used because their licences permitted reuse. That episode is the clearest proof that the licence you choose on Flickr is also an AI-training decision.

Open your Flickr settings and check the default licence for new uploads. If it is a Creative Commons licence that permits reuse, change it to All Rights Reserved for work you do not want reused.
Review your existing public photos and change the licence on the ones that matter, one by one or in a batch through the organiser.
Set sensitive or client work to private or friends-only rather than public, so it is not sitting in the open for any crawler to collect.
Remember that changing a licence does not recall copies already downloaded under the old licence, so do this before the next upload, not after a problem appears.

500px: keep photos out of Licensing

500px publishes no clearly documented AI-training toggle in its help pages, so do not waste time hunting for one. The control that matters is Licensing: submitting a photo to 500px Licensing grants 500px a broad worldwide licence to use, distribute and sub-license that photo, which is exactly the kind of permission that can put it into commercial datasets. Keep photos you care about out of Licensing entirely, use private settings for anything sensitive, and treat your 500px profile as a shop window for finished work rather than an archive of everything you have shot.

DeviantArt: check the default, do not assume it

DeviantArt is the strongest platform in this list. In November 2022 it introduced the noai and noimageai directives, opted every deviation out of third-party AI training datasets by default, and published the directives openly so any site could adopt them. When content is marked as not authorised, a page carries noimageai in its robots meta tag and an X-Robots-Tag header with the noai directive is sent when the image is downloaded directly from DeviantArt’s servers.

Open your DeviantArt account settings and confirm that your deviations are marked as not authorised for AI datasets. This should be the default, but check it rather than assuming it.
Check an individual deviation as well, because per-deviation settings exist alongside the account-wide one.
To verify the signal is really being sent, open one of your deviation pages, view the page source and search for “noai”. If it is there, the label is live.
Do not opt in to dataset use for a quick feature or promotion unless you mean it: the permissive direction requires an explicit opt-in from you, and that is the protection working as designed.

Two honest caveats. DeviantArt also runs its own AI image generator, DreamUp, and permits labelled AI-generated content on the platform, so being opted out of training datasets is not the same as the platform being AI-free. And the noai signal only binds parties that choose to respect it; a scraper that ignores it will still fetch a public image. ArtStation works in a similar way: projects receive a NoAI designation by default, and you must actively remove it to disable the protection, so leave it in place.

Where your photos live	Control that exists	What it covers	Main limit
Instagram and Facebook	Meta objection form (EU and UK adults)	Future training on your public posts and photos	No equivalent form in most other regions
Flickr	Visibility and licence choice	Who can see and reuse the photo	No dedicated AI-training switch
500px	Staying out of Licensing	Commercial sub-licensing of your photos	No documented AI-training toggle
DeviantArt	NoAI label, on by default	Third-party AI training datasets	Only binds parties that respect it
ArtStation	NoAI designation, on by default	Generative-AI datasets and inputs	Does not stop unrelated scrapers
Your own website	robots.txt rules and file signals	Compliant crawlers visiting your domain	Voluntary compliance only

How Do You Protect Photos on Your Own Website?

Your own website is the one place where you set the rules, and the controls here are the most technical — and the most effective — in this guide. Work through them in order.

Block Google-Extended. Add a robots.txt rule naming Google-Extended and disallowing it. Google’s crawler documentation explains that Google-Extended is the control for whether your content may be used to train future generations of Gemini models and for grounding in Gemini Apps, and that blocking it does not affect your site’s inclusion in Google Search and is not a ranking signal, according to Google’s crawler documentation.
Block the other training crawlers too. Add matching rules for GPTBot (OpenAI), ClaudeBot (Anthropic) and CCBot (Common Crawl), since each company only honours its own named crawler.
Add the new shared token — but keep the company rules. On 30 October 2026, Google, Microsoft, OpenAI and Anthropic announced support for one shared robots.txt token, __usage_ai-training, based on an Internet Engineering Task Force draft, so that a website can express its AI-training preference once instead of once per company. Google plans to support it in early 2027 and Microsoft from 1 February 2027, so add the token as its own group now, and keep the Google-Extended, GPTBot and ClaudeBot rules in place until each company confirms support.
Add the noai file signals. Put <meta name="robots" content="noai, noimageai"> on pages carrying your work, and ask your host to send an X-Robots-Tag header with the same values for your image files. These cost nothing and put your intent on record in a form machines can read, although only crawlers that choose to read them will respect them.
Embed a Content Credentials training preference. Adobe’s Content Credentials preference asks supported services, currently named as Firefly and Spawning, not to train on a file or use it to create new content. Our guide to Content Credentials (C2PA) walks through how to check and apply it, but note the trade-off: embedding the preference makes the work ineligible for upload to Adobe Stock.
Keep originals off the public web. Publish only the size you need for display, keep full-resolution files outside the public web root, and put client proofs behind a login rather than relying on a crawler request to protect them.
Write it down. Add a short statement to your site’s terms page that your photographs may not be used to train AI models without a licence, and note the date you published each rule, because a dated record is far more useful than a vague memory.

A practical example of the rules in step 1 to 3, for a photographer’s portfolio site, looks like this in robots.txt: a group for Google-Extended disallowing everything, a group for GPTBot disallowing everything, a group for ClaudeBot disallowing everything, and a group for the shared __usage_ai-training token disallowing everything. Test the file with a robots.txt tester after publishing it, and re-check it whenever you change hosts or redesign the site. That is the fullest way to opt out of AI training photos on a site you control. AI training is only one risk your photos face online, so pair these rules with the wider steps in our guide to protect your photos online, which covers watermarks, reverse image search monitoring and takedowns.

What Still Trains on Your Photos Anyway?

Even with every control above set correctly, some routes stay open, and you should know them before you promise a client that a photo is safe.

Copies already collected. Photos scraped before you opted out are already in datasets, and datasets get copied, so an opt-out cannot recall them.
Non-compliant scrapers. A robots.txt rule is a request, and a bot that ignores it will still download a public image. Blocking at your firewall or CDN is the only step that does not depend on anyone’s goodwill.
Screenshots and reposts. A follower can screenshot your photo and post it somewhere with no opt-out set, and from there it can be scraped as if it were theirs.
Uploads to AI chatbots. If you paste a draft or a photo into an AI assistant, that is a separate submission. For an individual ChatGPT account, turn off “Improve the model for everyone” in Settings, then Data controls, before uploading anything sensitive, and do not send feedback on a conversation containing confidential work, because the associated conversation may still be used to improve models despite the opt-out.
Missing tools. OpenAI promised a Media Manager opt-out tool in May 2024 and committed to shipping it by the end of 2025, but as of early 2026 the tool still does not exist, according to reporting on the company’s own announcements, so there is still no registry where you can flag images and expect them to be excluded from future GPT or DALL-E training.
Your face, not just your photos. If your concern is your own face appearing in face-search results, that is a different system from training datasets, and it needs its own opt-out. Our guide on how to opt out of PimEyes covers one of the largest face-search engines step by step.

None of this makes opting out pointless. It makes it a layered defence: each control closes a real route, and the routes that remain are narrower, harder and less likely to be the ones a mainstream AI company uses.

Opt-outs tell crawlers not to train on your photos, but they cannot change what a model learns from an image that is already public. Glaze and Nightshade add that extra layer, cloaking your photos before you post them so a model learns the wrong style or concept.

Worked Example: A Photographer Opts Out in One Afternoon

Sarah is a landscape photographer in the Lake District whose work sits in four places: a public Instagram account, a Flickr portfolio, a DeviantArt gallery and her own WordPress portfolio site. She has one free afternoon, so she works through the controls in this guide in order, and keeps a dated note of everything she changes.

First she takes an inventory. She runs a reverse image search on three of her best-known photographs and writes down every place they appear, including two reposts by walking blogs she had forgotten about. That list becomes her working map: she cannot protect a copy she does not know exists, and the search shows her which platforms hold her work.

Next she opens Instagram, goes to Settings and privacy, then Privacy Centre, and submits Meta’s objection form for her account, using the same name and email the account is registered under. Meta confirms by email that her public content will no longer be used for future AI training, and she saves that email in a folder with the date. She also switches a dozen old public posts to friends-only and leaves the account public only because her bookings come through it.

On DeviantArt she checks her account settings and finds the dataset opt-out already on, as expected, then opens two individual deviations and confirms the same setting on each. She views the source of one deviation page and finds the noai marker in the page, which confirms the signal is being sent. On Flickr she changes the default licence for new uploads to All Rights Reserved, then works through her public photos and changes the licence on the ones she still sells as prints, leaving only a few older Creative Commons images as they are because they were already widely reused.

Finally she opens her own site. She adds robots.txt groups disallowing Google-Extended, GPTBot and ClaudeBot, adds the shared __usage_ai-training token group, and asks her host to send the noai X-Robots-Tag header on image files. She embeds a Content Credentials training preference in the files she exports from now on, and she adds one line to her terms page stating that her photographs may not be used to train AI models without a licence. She checks the robots.txt file with a tester before finishing, and writes the date in her notebook.

What Sarah has not done matters as much as what she has. She has not removed any photo from a model that already trained on it, because no control can do that. She has not stopped a determined scraper, and she has not checked whether her photos were already used — that is a separate check for another day. What she has done is close every documented route a mainstream AI company would use to collect her work from tomorrow onwards, and she has a dated record to prove it.

FAQs

Can I opt out of AI training after my photos are already online?

Yes, for future training, but not for the past. Every control in this guide is prospective: it stops a company collecting or using your photos from the day you set it. It cannot remove your work from a dataset that already holds it, or un-train a model that already learned from it.

Does deleting a photo remove it from AI training?

No, not from a model that already trained on it. Deleting a public post stops that photo being used for future training, and Meta says a deleted post is not used for future training, but the model itself does not change. If the photo was copied or reposted elsewhere, deleting your copy does not delete those copies either.

Is robots.txt enough to stop AI training on my website?

No, on its own it is not enough, because robots.txt is a voluntary request rather than a lock. It works well against the big companies that publish their crawlers and honour the rules, such as Google, OpenAI and Anthropic. For anything else, combine it with noai signals, Content Credentials and firewall or CDN blocking.

Do I need to opt out on every platform separately?

Yes, you must opt out of AI training photos on every platform separately. There is no universal switch: Meta has its objection form, DeviantArt has its NoAI label, and your own website has robots.txt. The shared __usage_ai-training token announced in October 2026 will eventually cover websites in one step, but it does not reach photos hosted on social platforms.