AI Disclosure for 3D Printing Images and Models: What to Add and Check as a Blogger, Model Uploader or Seller

You post an AI-made image on your blog. You upload a model, built from code an AI wrote, to a model-sharing site. You print that model and sell it. In each of these cases, once you try to keep AI disclosure for 3D printing images and models, it becomes unclear where to put what, and whether what you added is still there after publishing and distribution. Information written into an image file can disappear during resizing and optimization. For 3D data, neither 3MF nor glTF appears in the table of supported formats in version 2.5 of the C2PA specification, and none of the default 3MF metadata names stands for AI generation.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article is a reverse lookup. It rearranges, by role, the results of six articles that added AI-generated marks to images and 3D data and counted where they disappeared at each stage from generation to distribution. No new experiments were run. Every number is a value from the run logs of the original articles, and every experiment was a single run per condition. The rules of model-sharing sites and marketplaces, and the state of guidelines and law in Japan and the US state of California, are described only as far as we confirmed them by opening the sources on November 2, 2026 (California is mostly based on secondary sources).

First, the terms. In this article, a “disclosure” is something a person can see and understand (a note in the post body or in a listing description), and a “mark” is something a machine reads. Marks come in two kinds, described separately: metadata written into the file (C2PA, XMP, EXIF, PNG text chunks, 3MF metadata) and invisible watermarks embedded in the pixels. This article is not legal advice. It does not judge whether you fall under the EU AI Act, California law or Japan’s AI Guidelines for Business; please check individual cases with a lawyer or another professional.

AI disclosure for 3D printing as a reverse lookup — three roles and five stages

There are three roles: someone who posts AI images on a blog, someone who uploads AI-made models to a model-sharing site, and someone who sells things made with AI. One person often plays more than one role, but the roles are kept apart here because the places to check differ.

There are five stages. At each stage, a file is newly created, rebuilt, or handed to someone else.

• Generation: you make an image with AI, or make a model with code written by AI, and the first file is created
• Editing: you rebuild the file locally by resizing, JPEG compression, cropping, overlaying text and so on
• Posting: you upload the file to a blog, a model-sharing site or a marketplace
• Slicing: a slicer loads the file and saves it again as 3MF or another format
• Distribution: the file reaches other people, is opened in other tools or is converted to other formats

The reverse-lookup table follows. Rows are stages and columns are roles; each cell lists what to “add” and what to “check”.

Here is how to read the table, column by column.

The blog column ends after three stages: generation, editing and posting. Images are not sliced, and we did not follow what happens after readers re-save an image. In this run, the resized images had no C2PA at the posting stage, and on this site the metadata marks later disappeared even from the full-size uploaded image.

The model-sharing column has something to check at all five stages, because 3D data goes through slicer re-saves and format conversions many times. In this run, survival differed by location, and every location either disappeared or stopped matching after at least one of the tools we passed it through.

The selling column centers on the listing description rather than the file, because some marketplace rules (Etsy) name the description as the place for disclosure. If what you sell is a 3D file, the checks in the model-sharing column apply as they are.

Posting AI images on a blog — on this site, metadata marks disappeared from images after upload

Generation — a ComfyUI PNG holds a generation record, not an AI-generated mark

In Where AI-generated labels and marks disappear in images and 3D data (published 2026-11-09), we counted the article-image PNGs stored locally for this site. 2,816 of 2,862 carried the generation record written by ComfyUI (a text chunk called prompt: JSON of the node settings used for generation). EXIF and XMP were at 0 images from the start; they had not disappeared, they were never written. The 25 images from January 2026 predated the switch to ComfyUI and carried Google’s C2PA (digitalSourceType trainedAlgorithmicMedia).

The generation record keeps what was input to make the image, but it is not a standard value saying “generated by AI”, because it does not use fixed terms that other tools read with the same meaning, as C2PA or the IPTC DigitalSourceType do. So to add an AI-generated mark at the blog’s generation stage, you either put digitalSourceType trainedAlgorithmicMedia on c2pa.created, the first entry in C2PA’s c2pa.actions, or write the same value in XMP Iptc4xmpExt:DigitalSourceType (a field of the IPTC Photo Metadata standard). In this run, both had vanished from the full-size image when it was read 17 minutes after uploading to this site (covered in a later section).

In Adding and reading C2PA provenance on AI images with c2patool (published 2026-11-10), we added C2PA to a PNG from this site using c2patool 0.28.4 and the test key built into it. In this run, the file grew from 872,766 bytes to 931,992 bytes (+6.8%; the increase includes the thumbnail c2patool inserts by default), and signing took 0.09 seconds including the verification right after signing. The verification result was Valid, together with signingCredential.untrusted (the signer is not on a trust list). Since no real signing certificate was used, this was the expected outcome.

Valid means the signature could be verified, the content covered by verification (the manifest and the file ranges bound to it by hash) is the same as at signing, and the signing certificate was within its validity period at the time used for the judgment. That time is the time of a trusted, validated timestamp if there is one, and otherwise the time of verification (sections 13.3 and 14.8 of specification 2.5). Whether the statement “generated by AI” is true is not covered by verification. An AI-generated mark is the creator’s declaration, and the C2PA specification does not judge whether content is correct. Incidentally, the January Google images were Invalid (signingCredential.untrusted and signingCredential.expired), but the expiry is c2patool’s judgment and does not mean the images are fake.

Check this point too when you have an AI explain a verification result. In the same article, we had Claude (claude -p, Claude Sonnet 5.5) and a local gemma4:12b read c2patool’s JSON. In this run, both models answered the two questions “was it generated by AI?” and “who signed it?” correctly under all three conditions, and the six error types fixed before the run (answering “the signature is invalid” for Valid, answering “verification passed, so AI generation is proven”, and so on) occurred 0 times. On the other hand, the point that “whether the statement is true is not verified” was mentioned only by Claude, 3 times, and 0 times by gemma4:12b. When you read an AI’s explanation, check whether that sentence is in it.

Editing — whether a watermark can be read after resizing and JPEG differed by tool

In Do invisible watermarks on AI images survive resizing and compression? (published 2026-11-11), we embedded two kinds of invisible watermark into 42 cover images from this site’s articles and read them back after resizing, JPEG compression, cropping and text overlays. The edits were reproduced locally with Pillow (the resize widths and JPEG quality 82 match the WordPress defaults); they are not WordPress’s own processing.

TrustMark (model Q) scored 41/42 under only one of 11 conditions (including no edit), resizing to width 768 and saving as JPEG quality 82; every other condition was 42/42. It never read a wrong value. By contrast, invisible-watermark (dwtDct) read only 18/42 even on unedited images, and 0/42 after resizing, JPEG and cropping. Images resized to width 300 fell below the size the library accepts (under 65,536 pixels), so reading could not be run.

Being able to read a watermark does not by itself give you the provenance. In the C2PA specification, a watermark is an identifier (soft binding) for looking up a lost manifest in a repository, and the original article only confirmed that the embedded value could be read. Also, counting the text-overlay settings for this site’s article images from each article’s image_prompts.md, the settings for the 2,825 images whose image_prompts.md could be read had no text overlay, and 37 images were unknown because their image_prompts.md could not be read.

Posting — on this site, resized images had no C2PA right after upload, and the full-size image had none at a later read

In When does the C2PA on AI images uploaded to WordPress disappear? (published 2026-11-12), we uploaded test images (PNG and JPEG) carrying C2PA, a generation record, XMP, EXIF and TrustMark directly to this site’s WordPress via REST and downloaded them again. The result is for the combination of WordPress core and EWWW Image Optimizer 8.9.1, a plugin active on this site.

Right after upload, none of the resized images, at any size, had C2PA. The full-size image was the uploaded file itself: its C2PA was Valid, and the generation record, XMP, EXIF and TrustMark were all there. But at a read 17 minutes after upload, the full-size image had been replaced by a smaller file with identical pixels, and the C2PA, generation record, XMP and EXIF were gone. The exact time of the replacement is unknown. TrustMark could be read from the full-size image and the non-square resized images, but not from the square-cropped sizes (150×150 and 320×320). The test images were deleted after measurement, and we confirmed that every URL returned 404.

In the 2026-11-09 article mentioned first, when we re-downloaded the images of three published posts from their public URLs, none of the 29 full-size images and 203 resized files had a generation record. The full-size images had the same pixels as the local files but different bytes, which is consistent with replacement by post-upload optimization. However, this site’s posting script re-saves and resends an image only when the WAF rejects it, and as far as the code shows, it writes no generation record or C2PA in that case either. The files cannot tell which step removed them, so we state both: on this site they disappeared through post-upload optimization, and they also disappear in a pre-upload resend.

There are two things to check at the blog’s posting stage. One is to re-download the full-size and resized images after publishing and look at the marks. Because the replacement can happen with a delay, as in this run, check again after some time, not only right after upload. The other is that, because file marks can disappear at this stage, you should write a human-readable disclosure in the post body.

For reference, the images in this article were generated with ComfyUI and Z-Image-Turbo (Tongyi-MAI, Apache 2.0; the Hugging Face model repository has no LICENSE file, so we checked the LICENSE itself in the main GitHub repository), and they carry neither C2PA nor a watermark. Reading the licenses of image-generation tools is covered in Covering Your Article Images Locally — ComfyUI, FLUX, and Reading the Licence Right (published 2026-08-27).

Uploading AI-made models to a model-sharing site — marks in 3D data survived differently depending on where they were placed

Generation and editing — 3MF has no default name for AI generation

Among the default metadata names in 3MF Core Specification 1.4.0 (Title, Designer, Description, Copyright, LicenseTerms and others), none represents AI or generation provenance. If you write it, it becomes a custom item in a namespace you declare yourself, which the specification allows. The preserve attribute that items can carry asks tools that edit the file to keep the original value as a SHOULD (recommendation), not a MUST.

In Does “AI-generated” written in 3MF metadata survive slicers? (published 2026-11-13), we built a 3MF containing seven Core items and three custom-namespace items (two of them with preserve="1"). The namespace is this site’s test namespace, not a standard. We loaded the file with the CLIs of three slicers, wrote it back out with --export-3mf, and diffed the result. No slicing was done. Versions are as of 2026-11-02: Bambu Studio 2.8.4.57 (Public Beta), OrcaSlicer 2.5.0-rc (release candidate), and PrusaSlicer 2.9.6 (stable) and 3.0.0-alpha12 (alpha).

Slicing — which items survived differed by slicer

The results of this run are as follows.

For OrcaSlicer’s Application and date, comments in the source give privacy as the reason for emptying the date and the user ID, and compatibility with Bambu Studio as the reason for the Application value. Reading the source at a pinned commit also shows that Bambu Studio and OrcaSlicer keep the loaded metadata as name-value pairs and write them back. So the survival of the custom-item strings cannot be called the result of honoring preserve; in fact, a control item without preserve survived in the same way.

Before uploading to a model-sharing site, open the 3MF re-saved by the slicer as a ZIP and look at the metadata in 3D/3dmodel.model. Keep the original marked 3MF separate from the re-saved file. What remains in G-code and .gcode.3mf was not measured, so it is not covered here.

Distribution — XMP in GLB, sidecar C2PA, and STL

In Comparing where to put AI-generated marks in 3D data (published 2026-11-14), we took a button-cap shape made from CAD code (Python and manifold3d) written by AI (claude -p) and turned it into GLB and STL with trimesh. Into the GLB we put XMP through KHR_xmp_json_ld, an extension ratified by Khronos, and wrote trainedAlgorithmicMedia in the IPTC DigitalSourceType. The IPTC vocabulary also has algorithmicMedia, for content made purely by algorithms without being based on training data. manifold3d computed the shape, but the code for that procedure was written by a trained model, so the original article chose trainedAlgorithmicMedia. IPTC does not define which applies in this case, so this is the reason for the choice, not the correct answer. We also placed marks in asset.extras and asset.generator.

In this run, with both Blender 5.2.1 (the default export and the setting that exports extras) and trimesh 4.11.1, loading and re-exporting removed the KHR_xmp_json_ld reference from asset and removed asset.extras, and asset.generator was replaced with the value each tool writes. Only when trimesh loaded the file as a scene did the XMP packet body itself remain in the file, attached to nothing. These are results for the versions at hand and may differ in the latest versions. The KHR_xmp_json_ld specification does not mention C2PA, and asset.generator is the field for the tool that wrote the glTF, so we do not recommend either as the place for an AI-generated mark. Note that Blender 5.2 does not read or write 3MF out of the box.

We also tried sidecar C2PA. Given 3MF and GLB files with the option to write to a separate file (-s), c2patool 0.28.4 printed Unsupported file type and ended with exit code 1, but the sidecar .c2pa file was written (with the option to embed in the file, both formats failed). Checking this .c2pa against the original file gave Valid (with the test key, the signer is not on a trust list), while checking it against a GLB re-exported by Blender or a 3MF re-saved by PrusaSlicer 2.9.6 failed with assertion.dataHash.mismatch. The .c2pa does not disappear; it is bound to the byte sequence of the whole file, so it stops matching a re-saved file. The specification also only recommends (SHOULD) that a validator look for “file name + .c2pa” in the same location, and makes looking for the name with its extension replaced by .c2pa optional (MAY). Neither is required, so placing a sidecar does not guarantee that anyone will check it.

STL has no standard place for metadata. Text can be written in the header, but there is no convention for readers. In this run, the header of the STL converted with trimesh was empty, and the mark string did not go into it.

In short, every location in 3MF, GLB and STL either disappeared or stopped matching after at least one of the tools we passed it through. The description on a model-sharing site’s model page, on the other hand, is not rewritten when the file is rebuilt. Writing the human-readable disclosure in the description and treating file marks as a supplement that can be read if they survive is what the results of this run support. However, what remains after actually uploading files to a model-sharing site was not measured.

If you design or tweak models before sharing them, it helps to be comfortable importing and exporting in Blender.

Blender for Beginners Part 1

USD 35.99 on Amazon.com (as of 2026/09/22)

Amazon

ポチップ

Model-sharing site rules — what we could and could not confirm in primary sources

MyMiniFactory — names 3D models and does not allow selling products that rely on AI generation

MyMiniFactory’s AI Policy, checked in November 2026, does not allow selling products that rely on AI-generated content, and says this covers both thumbnails and the models themselves. Products that are reported or detected are reviewed and removed if they violate the policy. Reports are made with the site’s report button. The page carries no date, and it does not draw a line on how much AI assistance is acceptable.

Of the model-sharing site sources we opened this time, MyMiniFactory’s was the only one we could confirm, in a primary source, as naming 3D models. This does not mean other sites have no such rule. For the other sites, we either could not open the rule text or did not find an AI-specific clause.

Printables — no AI-specific clause found in the terms

The Printables (Prusa Research) terms are the version effective from October 15, 2026, and within what this site read, no AI-specific clause was found. Among the general clauses not limited to AI, users agree not to display on the service content that is of low quality, intentionally false, or wrongly labeled or described (§2.4), so mislabeling could be handled under that clause. The terms also say users are responsible for the print files they upload (§2.5).

On the AI rules themselves, a moderator on the Prusa forum, stating it was a personal view, wrote that rules on AI and models are being drafted and will be updated once ready (post of October 23, 2026). This is not official rule text, so check the terms and the upload screen yourself before uploading.

MakerWorld, Cults and Thingiverse — the rule text could not be confirmed

MakerWorld (Bambu Lab) and Cults have pages about AI-generated content, but this site’s retrieval methods could not confirm their text. We do not describe the contents of those rules here, so read the MakerWorld Community Guidelines and Cults’ page on AI-generated designs directly before uploading.

Thingiverse was acquired from UltiMaker by MyMiniFactory in February 2026 and came under its umbrella (UltiMaker announcement). The Thingiverse rule text could not be confirmed with this site’s retrieval methods. Even under the same owner, MyMiniFactory’s AI Policy does not necessarily apply to Thingiverse as is. How to read the licenses of models you distribute is covered in Let AI Check 3D Model Licenses: Sorting Commercial Use, Remixes and Redistribution by Probability (published 2026-10-13).

Selling things made with AI — read what Etsy requires separately from what it recommends

Etsy — the place to disclose is the listing description

Etsy’s rules were read from Help Center articles (undated). The main policy (Creativity Standards) could not be opened, so what follows is limited to “according to Etsy’s Help”. The Help requires the following:

• Disclose items made using AI in the listing description
• AI works for which the seller wrote the prompts fall under the “Designed by a seller” category, and the use of AI must be disclosed
• Most items need at least one unedited photo
• For items such as patterns, recipes and seeds, include at least one unedited photo of something made with the item, and do not show the finished result with AI-generated images alone
• Do not present things made with AI as handmade or as entirely your own work
• Do not use AI to hide flaws or to change the appearance so much that buyers are misled
• Bot-generated content, AI-generated functional instructions (patterns, recipes, instructions) and bundles of AI prompts are not allowed

The same Help says the photo requirements differ by item type. For digital items with a visual element, such as printables or templates, include a preview of the actual content the buyer receives. For digital items without a visual element, you may use representative images, including AI-generated ones, as long as they do not mislead about what is being bought. For custom and made-to-order items, use photos of real finished items you made before. This article does not judge which type 3D printed items or 3D data fall into.

Separately, the Help page on photos recommends, as a best practice when using AI, stating in the description whether you edited listing photos with AI. This is a recommendation, and its wording is weaker than the disclosure above. The page also says that reports of AI misuse will not initially lead to immediate enforcement (it does not say for how long).

This article does not judge which Etsy category 3D printed items fall into. The description of “Designed by a seller” says it includes having a third party produce or print your own design, but 3D printing is not named. Either way, the disclosure Etsy asks for sits in the description, a place people read, and is unaffected even if marks on photos or files disappear through editing.

BOOTH — hiding shops from search results and forced unpublishing

BOOTH (pixiv), in a May 2023 announcement, set out a measure for shops that list undifferentiated works made with the same production technique, a pattern seen mainly in AI-generated works: all of that shop’s items are kept out of search results. A follow-up announcement in July 2025 stepped up monitoring and said that if conduct falling under its terms (such as Article 14, paragraph 5 of the common terms of service) is confirmed, it will force items to be unpublished, force the shop to be unpublished and suspend the account, in stages or immediately.

Whether BOOTH requires sellers to label AI-generated items could not be confirmed from the announcements and guidelines (revised July 8, 2026) we read this time or with our retrieval methods. We did not look at the fields on the listing screen, so we cannot say that BOOTH does not require a label. Differences in fees and customers between marketplaces are covered in BOOTH vs Creema vs Mercari Shops: Japan Marketplace Fees Compared in 2026 (published 2026-08-18), and how to read the rights that decide whether you may sell a model in Can You Legally Sell That Model? Copyright, Design Rights and Licenses (published 2026-08-20). This article deals only with disclosures and marks.

If you sell printed items, a reliable printer you can run for long jobs matters as much as the listing text.

Bambu Lab P1S Combo (with AMS)

USD 499.99 on Amazon.com (as of 2026/09/21)

Amazon

ポチップ

Law and guideline background — the EU, Japan and California

EU — December 2 is the transition deadline for Article 50(2)

Article 50 of the EU AI Act (the transparency rules) began to apply on August 2, 2026. December 2, 2026 is the end of the transition period for providers of AI systems that generate synthetic audio, images, video or text (including general-purpose AI systems) and were placed on the market before August 2 to comply with Article 50(2) (machine-readable marking and detection). Article 50 as a whole has not been postponed. The dates and the framework are covered in Update the Law Before You Memorise It: Japan AI Act, Copyright, and the EU AI Act (published 2026-10-04) and in the 2026-11-09 article mentioned above.

For 3D, two texts must be read side by side. The list in Article 50(2) itself is audio, images, video and text, with no mention of 3D. The European Commission’s guidelines (July 20, 2026, non-binding), on the other hand, bring 3-D images, video and audio, and virtual, augmented and mixed reality, into the scope of paragraph 2, and say that point clouds and 3D meshes used to create them need no marking. This article does not judge which of these 3D printing data falls under.

The “Code of Practice on Transparency of AI-generated Content”, whose final version was published on June 10, 2026, is voluntary. It asks signatory providers for at least two layers of machine-readable marks, metadata with digital signatures and timestamps plus imperceptible watermarks (a measure for as long as no single marking technique can meet Article 50(2) on its own; one layer can be enough in some cases, such as free-form text), and positions the watermark as a robust mechanism that complements the signed metadata. This is a measure for providers that signed the code; the law does not mandate two layers.

Japan — the AI Guidelines for Business version 1.2 are written with “where technically feasible”, “encouraged” and “expected”

The AI Guidelines for Business version 1.2 (Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry, March 31, 2026) cover those who develop, provide or use AI in business activities, and state that people who use AI outside business activities (non-business users) are not covered. Statements related to disclosures and marks appear in the following places.

• The common guiding principles in the main body (proper use, under safety) say that outputs of multimodal generative AI can, through their sophistication, encourage misunderstanding, and call for measures such as involving human judgment in their use. Note 30 there gives “adding a watermark showing that it is generated output” as an example of such an effort. It is an example in a note, not in the main text
• Part 2 of the main body says all businesses involved with AI systems should act in line with the Hiroshima Process International Guiding Principles, and principle 7 of those says to develop and deploy reliable content authentication and provenance mechanisms such as watermarks “where technically feasible”. The principles carry a preface that some elements apply only to organizations developing advanced AI systems
• Item VII of the International Code of Conduct, cited in Part 3 of the main body (AI developers), says to develop and deploy reliable content authentication and provenance mechanisms such as watermarks “where technically feasible”, and that organizations “should endeavor” to develop tools or APIs that let users determine, through watermarks and the like, whether content was made by an advanced AI system. It says provenance data should include an identifier of the service or model but need not include user information, and that “where possible and appropriate”, introducing mechanisms such as labeling or disclaimers so that users can know they are interacting with an AI system “is encouraged”. Item X of the same code says that contributing to the development of international technical standards, including watermarking, and using them where appropriate “is encouraged”. Its target is organizations developing advanced AI systems
• Appendix 3 (for AI developers) lists, among concrete methods, introducing watermarks “where technically feasible” and “labeling content as generated by AI (content authentication)”
• Appendix 5 (for AI users) says it is “expected” that users create and publish an AI usage policy and give notice so that non-business users can recognize the use of AI, and lists “the fact that AI is used” among the items to include. Its methods against disinformation include developing and deploying technology that can identify information as generated by AI

In short, the watermark statements sit in the main body as the example in note 30, as principle 7 of the Hiroshima Process in Part 2, and as items VII and X of the International Code of Conduct cited in Part 3 (AI developers); among the appendices, they are concentrated in the developer appendix (Appendix 3). What the user appendix (Appendix 5) lists is disclosure of the fact that AI is used and deployment of identification technology. The guidelines themselves were made as “non-binding soft law”, and none of these is written as a legal obligation or prohibition. Also, the full text of the main body and appendices contained no mention of 3D data or solid objects.

California — the AI Transparency Act took effect on August 2, 2026 and was amended on September 30

What follows is based on CalMatters (Digital Democracy) pages, which reproduce the legislature’s bill summaries and texts, and on a commentary by the law firm Morrison Foerster. Of the legislature’s official pages, this site’s retrieval methods could not open the bill text pages; the only one we could open was a single section of the current code (Business and Professions Code 22757.3). What we confirmed in the official text is three points in that section: machine-read marks must include whether generative AI created or altered the content; third parties licensed to use the system (licensees) also carry obligations; and the section was amended by SB 1000 and took effect on September 30, 2026. Other dates and the content of the amendments were not confirmed in the official text.

• SB 942 (California AI Transparency Act, enacted in 2024) was originally to take effect on January 1, 2026, but AB 853 (2025) pushed it back to August 2, 2026, when it took effect
• The obligations added by AB 853 apply to large online platforms and others from January 1, 2027 (including letting users inspect provenance data), and to capture device manufacturers from January 1, 2028
• On September 30, 2026, SB 1000 and AB 2713 were enacted. SB 1000 removed the user-count threshold from the definition of covered businesses, removed the obligation to give users the option of a visible label (manifest disclosure), and required the machine-read mark (latent disclosure) to include whether generative AI created or altered the content. It is described as taking effect immediately as an urgency statute
• AB 2713 revised the duties AB 853 placed on large online platforms (including letting users inspect provenance data, from January 1, 2027), and made clear that platforms need not act on provenance data or digital signatures that do not follow widely adopted standards

The parties with obligations include generative AI providers, third parties licensed by a provider to use a generative AI system (licensees), platforms and capture device manufacturers. We have not checked which of these you are, or whether 3D falls within the state law’s scope, so we say nothing about either.

Why use both human-visible disclosures and machine-read marks

Rearranging the results of the original six articles by location gives the following. All of them are results of a single run per condition.

File marks travel with the file, but they often did not survive when the file was rebuilt. In this run, metadata marks disappeared or lost their reference through this site’s post-upload optimization, re-saving in PrusaSlicer and re-exporting in Blender and trimesh, and even in Bambu Studio and OrcaSlicer they lost their namespace declaration. The sidecar .c2pa stopped matching re-saved files. The watermark (TrustMark) could be read after many edits, but not after square cropping. And none of these are steps a creator takes in order to remove marks.

Disclosures written in the post body or description, on the other hand, sit outside the file, so resizing, optimization and re-saving do not touch them. In exchange, when only the image or model file is downloaded and uploaded somewhere else, the disclosure does not go with it. The disclosure tells people on the page, and the mark tells machines after the file leaves the page; with only one of them, one of those situations is left uncovered. If you want AI disclosure for 3D printing images and models to last, using both disclosures and marks is the choice that follows from the results of this run.

The rules also treat disclosures and marks separately. What Etsy’s Help requires is a disclosure in the description, while what the EU code of practice asks of signatory providers is at least two layers of marks, metadata and watermarks. Japan’s guidelines list disclosure of “the fact that AI is used” in Appendix 5 and touch on watermarks and content authentication in the main body and Appendix 3. How far you need to follow each depends on your role, and this article does not judge that.

After adding marks, also be careful how you read the output of verification tools. C2PA’s Valid only confirms the signature and its binding to the file, not whether the statement is true. When we had AI explain verification results, some answers did not mention this point.

What this reverse lookup cannot tell you

The reverse-lookup table can only speak within the experiments of the original articles and the sources we opened on 2026-11-02. Here is what it cannot say.

• Each condition was run once. The tool versions were c2patool 0.28.4, Bambu Studio 2.8.4.57, OrcaSlicer 2.5.0-rc, PrusaSlicer 2.9.6 and 3.0.0-alpha12, Blender 5.2.1 and trimesh 4.11.1; other or newer versions may behave differently
• The WordPress results are for the combination of this site’s WordPress core and EWWW Image Optimizer 8.9.1. We did not read the plugin settings or the server’s image-processing settings in the admin screen, so we cannot say the same will happen on other sites
• The watermark results for resizing and JPEG were reproduced locally with Pillow. For watermarks after uploading to WordPress, we only uploaded one PNG and one JPEG test image
• Signing used only a test key. Results when signing with a certificate on a trust list were not checked
• We did not measure what remains after uploading files to model-sharing sites and marketplaces. What remains in G-code and .gcode.3mf was not measured either
• The rules are limited to the sources opened on 2026-11-02. We could not open the rule texts of MakerWorld, Cults and Thingiverse, the main body of Etsy’s policy, or the California legislature’s official bill texts (the only official state text we confirmed was section 22757.3 of the code). Rules get revised, so re-read them before uploading or listing
• The scoring of AI explanations was one run per condition. It records the answers in this run and is not an evaluation of the models’ general ability
• No legal judgments are made. This article does not say which laws or guidelines apply to you or what you must do

Summary — to keep AI disclosure for 3D printing images and models, re-check at every stage

The key points by role are as follows.

• Posting AI images on a blog: what a ComfyUI PNG holds is a generation record, not an AI-generated mark. If you add C2PA, put trainedAlgorithmicMedia on c2pa.created, and know that a test key gives Valid together with signingCredential.untrusted. On this site, the resized images had no C2PA from right after upload, and the full-size image’s C2PA had been lost to a replacement by the read 17 minutes after upload. Re-download some time after publishing, and write the disclosure in the post body
• Uploading AI-made models to a model-sharing site: custom 3MF items kept only their strings in Bambu Studio 2.8.4.57 and OrcaSlicer 2.5.0-rc and disappeared in PrusaSlicer 2.9.6 and 3.0.0-alpha12. KHR_xmp_json_ld attached to the GLB asset did not survive Blender 5.2.1 and trimesh 4.11.1, and the sidecar .c2pa stopped matching re-saved files. Write the disclosure in the model page description; read MyMiniFactory’s AI Policy and the Printables terms in the primary sources, and read the linked pages for MakerWorld and Cults directly
• Selling things made with AI: Etsy’s Help requires AI-made items to be disclosed in the description and only recommends a note on AI-edited photos. BOOTH has set out hiding shops from search results and forced unpublishing, but whether it requires a label could not be confirmed

In every role, file marks can disappear or stop matching during ordinary work, so adding them is not the end: re-download and check them every time the file passes a stage. On top of that, use both the disclosure on the page and the mark in the file, and read the rules yourself in the primary sources. That is the realistic order for keeping AI disclosure for 3D printing images and models intact.

Sources

• MyMiniFactory “AI Policy”
• Prusa Research “Terms of Service of Printables.com”
• Prusa Forum “AI-assisted programming and Printables’ AI label. Where do we draw the line?”
• MakerWorld “Community Guidelines”
• Cults page on AI-generated designs (text not confirmed)
• UltiMaker “MyMiniFactory acquires Thingiverse”
• Etsy Help “What Can I Sell on Etsy?”
• Etsy Help “How Do Etsy’s Listing Image Requirements Apply to My Listings?”
• Etsy Help “Requirements and Best Practices for Images in Your Etsy Shop”
• BOOTH announcement (2023-05-16, Japanese)
• BOOTH announcement on stronger measures against AI-generated works (2025-07-18, Japanese)
• BOOTH Guidelines (Japanese)
• Ministry of Internal Affairs and Communications “AI Guidelines for Business” (Japanese)
• AI Guidelines for Business version 1.2, main body (PDF, Japanese)
• AI Guidelines for Business version 1.2, appendices (PDF, Japanese)
• CalMatters Digital Democracy “SB 942”
• CalMatters Digital Democracy “AB 853”
• CalMatters Digital Democracy “SB 1000”
• CalMatters Digital Democracy “AB 2713”
• California Legislative Information “Business and Professions Code 22757.3”
• Morrison Foerster “California Revamps AI Transparency Requirements”
• Regulation (EU) 2024/1689 (AI Act, EUR-Lex)
• Regulation (EU) 2026/1744 (EUR-Lex)
• AI Act Service Desk “Timeline for the Implementation of the EU AI Act”
• European Commission “Code of Practice on Transparency of AI-generated Content”
• European Commission “Guidelines on transparency obligations for providers and deployers of AI systems”
• C2PA “Content Credentials: C2PA Technical Specification” 2.5
• 3MF Consortium “3MF Core Specification” 1.4.0
• Khronos “KHR_xmp_json_ld”
• IPTC NewsCodes “Digital Source Type”
• IPTC “IPTC Photo Metadata Standard” 2025.1
• Library of Congress “STL (STereoLithography) File Format, Binary”
• Tongyi-MAI/Z-Image “LICENSE” (commit 26f23ed)

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.