How to Check if Your Photos or Artwork Were Used to Train AI

🔎 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.

📁 Upload Image 🔗 Paste URL

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

Load

💡 Right-click any image online → "Copy image address" → paste here.

—

Dimensions: — Size: — Type: —

✕

Search this image on:

🔵 Google

🔍 Lens

🟦 Bing

🔴 Yandex

🟢 TinEye

📌 Pinterest

Select all engines 🔎 Search All Selected Engines

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

If you want to check if your work was used to train AI, start with an honest fact: no public tool can prove that a named AI model trained on your exact photo. What you can do is build a clear, dated record of the evidence that does exist, and this guide shows you how, step by step.

Short answer: There is no public tool that can prove a named AI model trained on your exact photo. The closest check is to search an indexed dataset such as LAION-5B, which holds 5.85 billion image-text pairs, for a matching record, then combine that with a reverse image search and the provider’s own policy. A match is a lead, not proof.

Training data is the collection of images, text and other material that an AI model studies so it can learn patterns and generate new work. When photographers and illustrators ask whether their work trained an AI, they are usually asking whether a copy of their specific image ended up in one of those collections. This guide explains how to check the places where that question can actually be answered, what each result does and does not prove, and what to do with what you find.

What Does “Used to Train AI” Actually Mean?

The phrase covers several different events, and they are not interchangeable. Your image might appear in a public dataset, it might have been collected from a public web page by a crawler, or it might have been submitted directly to a service you used. Each of those is a different event, with different evidence and different people to contact.

A public dataset is a large, organised collection of image and text pairs that researchers and companies use to train models. The best-known open example is LAION-5B. A web crawl is different: a company downloads pages from the open web, including the images on them, and uses that material in training. A direct submission is different again: you uploaded the image to a product, and that product’s terms say what happens to it next.

Confusing these events is the most common mistake creators make. Finding your image in a dataset does not prove that a particular commercial model trained on it. Finding a generated image that looks like your style does not prove that your file was in any dataset. And a provider’s policy about one product does not describe what happened to a photo you posted on a different platform three years ago.

Copyright protects your specific work, not a general style, subject or genre, so a model producing something “in the style of” your art is a different question from a model having trained on your actual file. Keep the two questions separate throughout your investigation, because they need different evidence.

Can You Prove an AI Model Trained on Your Photo?

The short answer is usually no, at least not with public tools alone. What you can prove is narrower but still useful: that a matching record exists in a specific indexed dataset on a specific date, that your image appears on particular web pages, and what a provider’s published policy said for a particular product and period. Those three facts, properly recorded, are the foundation of any later complaint or legal question.

The table below compares the five kinds of evidence you can gather, what each one can show, and where each one stops.

Use this claim-strength checklist before you describe your findings to anyone:

• Strong and supportable: “The searched LAION-related index displayed a record matching my image on 7 November 2026.”
• Supportable with policy evidence: “This account setting was enabled for new submissions from the date shown, according to the provider’s policy.”
• Too strong without direct documentation: “Model X trained on my image.”
• Also too strong: “The model copied my style, so it must have used my work.”
• Overbroad: “My opt-out removes the work from every AI dataset.”

The useful goal is not a certainty that public tools cannot supply. It is a dated, reproducible record showing what an index contained, what a provider’s policy covered, what setting you applied, and where the uncertainty remains.

How Do I Check If Your Work Was Used to Train AI?

For photographs and artwork, the most practical public check is to search an indexed dataset. LAION-5B is the dataset behind much of this discussion: according to LAION’s own announcement, it contains 5.85 billion CLIP-filtered image-text pairs, including 2.3 billion in English, 2.2 billion in more than 100 other languages, and about 1 billion with no clear language. The images in it were gathered from the public web, and the dataset is distributed as metadata under a Creative Commons CC-BY 4.0 licence, while the images themselves remain under their own copyright.

The creator-facing tool built for this search is Have I Been Trained?, run by Spawning AI, which searches indexed material including records associated with LAION. Treat it as a lead-generation tool, not as a model-training detector. One honest warning for November 2026: when this guide was checked, both Have I Been Trained? and Spawning’s own site displayed a maintenance notice, and LAION’s own search demo did not load. Tools in this space change often, so if one is unavailable, try again later and use the reverse image search method in the next-but-one section in the meantime.

Follow these steps when the tool is available:

1. Choose the best source image. Start with the original file, then test important crops and variants. A thumbnail, a crop, a transparent-background export or an altered colour profile may match differently from the original.
2. Read the search service’s terms first. Before uploading sensitive or unpublished work, check how the service handles submitted images, retention and deletion. A search provider has its own data practices, separate from the dataset it searches.
3. Submit the image and note the coverage. Record which dataset or index is being searched, its displayed version, and whether the service searches image URLs, captions or both. A tool cannot report records outside its coverage.
4. Inspect every potential match. Open the associated record and capture its identifier, source URL, caption, displayed image and any publication information. A visually similar image is not enough if the record is actually a different work.
5. Compare provenance. Check whether the source URL is your original page, a mirror, a scraped copy, a licensed stock submission or an unrelated page using a similar image. Compare dimensions, visible edits, watermark placement, filename clues and publication dates.
6. Save the result. Export or screenshot the match, record the search date, and preserve the exact dataset name and version shown by the tool. Dataset contents and availability can change.
7. Describe the finding precisely. The defensible statement is: “A matching record appears in the searched dataset index.” Do not state that a named model trained on your work unless you have direct documentation linking that model’s training data to the record.

A negative result is not clearance. No match may mean the image was never in that index, but it may also reflect a crop, a transformation, missing metadata, a private collection, a different dataset, or a version of the work that the search tool cannot recognise. A work may have been downloaded before an index was rebuilt, or removed afterwards. Never present “not found” as proof of non-use.

Worked Example: Tracing One Illustration from Portfolio to Dataset Record

This worked example follows a concrete case through every step, so you can repeat the same method with your own work. The illustrator is fictional, but every step is the real procedure described above, and the outcomes shown are the kinds of results each check actually produces.

Maya is a freelance illustrator in Leeds. In March 2024 she published a watercolour illustration of a fox in autumn leaves, titled “Autumn Fox”, on her portfolio website, and posted the same image on Instagram the same week. In October 2026 a follower tells her that an image generator keeps producing foxes in a strikingly similar style. She wants to know whether her illustration was used to train AI, so she works through the checks in order.

Step 1: preserve the evidence. Maya saves the original high-resolution file, notes the portfolio URL and the 12 March 2024 publication date, and screenshots the Instagram post with its date. She calculates a file hash and stores it with the original, so she can later prove that a downloaded copy is the same file.

Step 2: search the dataset index. Maya uploads the original illustration to Have I Been Trained? and runs the search. Suppose the tool returns a record whose image matches her fox, with a source URL pointing not to her portfolio but to a wallpaper site that reposted the image without credit. That is a lead, not a verdict: it shows the index contains a record of her image, sourced from the wallpaper copy, and it does not by itself prove that any named model trained on it.

Step 3: check the source. Maya opens the wallpaper page and finds her illustration with its watermark cropped out, dated two months after her portfolio post. She screenshots the page, records the URL and date, and notes that the copy, not her original, is the version the index points to. If the search had returned no match, she would record that too, with the date and the exact image variant used, and remember that a no-match does not clear the work.

Step 4: run a reverse image search. Maya searches the same image with Google Lens and TinEye, as described in the next section, and finds the illustration on three more sites: her own portfolio, the wallpaper site, and a stock-photo page that is selling it without a licence. Each copy is a separate fact to record, because each one may have been the version a crawler collected.

Step 5: check the policies that apply. Maya reads the policies of the services she actually used: her portfolio host, Instagram, and the image generator she is concerned about. She notes what each one said about training, and for which dates, and saves copies of the relevant pages.

Step 6: write the finding down honestly. Maya’s evidence log now says: a matching record exists in the searched index, sourced from an unlicensed copy; the image appears on four public pages; and the provider policies she read do not confirm per-work training. That is a solid, defensible record. What she cannot honestly say is that a named model trained on her fox, because no public source establishes that link.

How Can Reverse Image Search Help You Find Copies of Your Work?

If you want to check if your work was used to train AI, do not stop at one dataset search. A reverse image search shows you where your image actually appears on the public web, and that is often how you discover the copy that a dataset or a crawler picked up. If your photo has been reposted, sold, or scraped, the copies are usually the trail that leads back to the source.

Start with a reverse image search of your original file, then repeat it with your most important crops and variants. A cropped or recompressed copy can match differently from the original, and the copy a dataset holds is often not the version you published.

Follow these steps:

1. Search with Google Lens. Upload your image at lens.google.com or through the Google app, and open the results that show the same image rather than merely similar ones. Our Google reverse image search guide walks through the desktop and mobile steps.
2. Search with TinEye. TinEye is useful because it sorts matches by date, so you can see the earliest appearance of an image and work out which copy came first.
3. Try Bing Visual Search as a third engine. Different engines index different pages, so a copy that Google misses may appear on Bing, and the reverse is also true.
4. Open every distinct source. For each match, open the page and record the URL, the publication date, and whether the image is your original, a crop, a recompressed copy, or a different work that only looks similar.
5. Compare dates against your own publication date. A copy dated after your original post is evidence of copying; a copy dated before it needs a closer look, because it may be an earlier version of your own work or a coincidence.
6. Save screenshots and URLs in your evidence log. Pages change and disappear, so a dated screenshot with the URL visible is worth far more than a memory of what you saw.

Reverse image search is also the fastest way to spot an AI-generated copy of your work. If a generator has produced something very close to your image, searching that generated image can lead you back to the pages and datasets it was associated with. Our guide on how to tell whether an image is AI-generated explains the visual tells to check first.

What Do the Major Providers Say About Training on Your Work?

Provider policies are general statements, not per-work answers, but they are still essential evidence. They tell you which products collect what, what the defaults are, and where the opt-out controls live. Read the policy that applied at the relevant time, for the exact product and account type you used, because consumer, business and API accounts often have different rules.

OpenAI says that inputs and outputs from its business products, including ChatGPT Team, Enterprise and the API, are not used for training by default unless the organisation opts into data sharing. Consumer ChatGPT is different: eligible content from individual services may be used to improve models, and the data controls apply prospectively, so check the exact product, the feedback rules and any separate settings for tools such as Codex. OpenAI announced a Media Manager opt-out tool in May 2024, but at the time of writing it still had not launched.

Adobe distinguishes product-improvement analysis from generative-AI training. Its Content Credentials guide describes a generative-AI training and usage preference that can be embedded in each file, currently supported by Firefly workflows. Selecting it can also prevent the work from being uploaded to Adobe Stock, and Adobe says the preference is recognised by Firefly and Spawning.

Google’s Gemini Apps privacy notice says that with Keep Activity switched off, future chats are not used to train models unless you submit feedback, although Google retains those chats for 72 hours for responding, feedback processing and safety. Meta uses public posts from Facebook and Instagram to train its AI models on an opt-out basis, so if you post your work there you need to submit an objection through settings, and it does not necessarily cover everything you have already posted.

DeviantArt excludes all deviations from third-party AI training datasets by default, following its November 2022 reversal after the DreamUp backlash, and it sends a noai directive as an HTML tag and an HTTP header. ArtStation’s NoAI tag is the opposite: it is not applied automatically and has to be enabled manually in your account settings.

Two legal developments matter for 2026. Under Article 53 of the EU AI Act, providers of general-purpose AI models have had to publish a sufficiently detailed public summary of the content used to train their models since 2 August 2025, using a template the European Commission published on 24 July 2025. Those summaries describe sources at a general level, not individual works, and models already on the market before August 2025 have until 2 August 2027 to comply. In the UK, the High Court handed down judgment in Getty Images v Stability AI on 4 November 2025: Getty had withdrawn its primary copyright claims, the court dismissed the secondary infringement claim because model weights are not an infringing copy, and Getty succeeded only on limited trade mark grounds over iStock watermarks in early model versions. An appeal is due to be heard on 10 November 2026, so the position may still change.

What Should I Do If I Find My Work, and How Do I Protect Future Work?

If your checks turn up a dataset record or an unlicensed copy, work through these steps in order, and keep a written log of everything you do.

1. Preserve the evidence first. Save the dataset record, source URL, metadata, screenshots and your original file, and record the time zone and date of each step.
2. Determine the actual source. A match may point to your own page, a licensed stock submission, a mirror or an unrelated image. Establish which copy is involved before you contact anyone.
3. Identify the responsible operator. The dataset maintainer, the website hosting the source and the model provider may be different organisations. Contact the one that controls the relevant record.
4. Ask a narrowly framed question. Request confirmation of whether the identified record is retained, how removal requests work, what future datasets are affected, and whether completed training or downloaded copies can be reversed. Avoid claiming model use that your evidence does not establish.
5. Keep the response with your log. Save ticket numbers, replies, policy versions and effective dates. A removal request may change a maintained dataset without erasing copies already downloaded or undoing completed training.
6. Escalate rights questions appropriately. If the work has significant commercial or legal value, obtain advice from a qualified copyright or technology professional in the relevant jurisdiction. Our guide to reverse image search for lawyers explains how to document image evidence properly before you speak to one.

For future work, these protective measures are worth setting up now, with their limits understood:

• Register an opt-out where one exists. Spawning AI’s Do Not Train registry lets you flag work that participating trainers commit to skip, and you can also check a photo’s Content Credentials to see whether a training preference is already embedded in a file.
• Confirm platform settings. Check the NoAI setting on DeviantArt and ArtStation, and the training objection on Instagram and Facebook, rather than assuming the default protects you.
• Use protective tools on work you post publicly. Glaze and Nightshade, free tools from the Glaze Project at the University of Chicago, add imperceptible changes to an image: Glaze disrupts style mimicry, while Nightshade is designed to corrupt what a model learns from the image. They add friction, not certainty, and researchers continue to develop ways to strip such protections.
• Watermark and keep originals. A visible watermark does not stop scraping, but it makes unlicensed copies easier to identify, and a high-resolution original with its metadata intact is your best evidence of authorship.
• Control what you upload. Do not upload unpublished or high-value work to a service until you have read its training policy, and remember that an opt-out set today cannot undo training or collection that already happened.

No single measure guarantees that your work will never be used to train AI. The steps in this guide show how to check if your work was used to train AI with the tools that exist today, and the realistic aim is to make your work harder to take, easier to trace, and backed by a dated record that shows exactly what you found and when.

FAQs

Can I prove an AI model trained on my photo?

Usually not with a public search alone. A matching record in an indexed dataset shows that the searched index contains related material, but it does not prove that a named model used that record or that it entered a particular training run. Direct provider documentation about the model’s training data would be needed for that stronger claim.

What does a match in LAION-5B actually mean?

It means the searched LAION-related index displayed a matching or related image-text record, with a source URL and caption. LAION’s own announcement reports 5.85 billion image-text pairs, but that figure is the dataset’s scale, not a count of your work or proof that every model used the dataset.

Does a “no match” result mean my work was not used?

No. The work may be in another dataset, a private collection, a different version, or a transformed form that the tool cannot recognise. A record may also have been removed or become unavailable after it was downloaded, so never present a “not found” result as proof of non-use.

Can I stop my photos being used to train AI in future?

Partly. You can register opt-outs, set Content Credentials preferences, confirm platform settings such as DeviantArt’s NoAI label, and use tools like Glaze before posting. These measures affect future collection by services that respect them, but they cannot remove copies already downloaded or undo training that already happened.