Public Domain In, Not Necessarily Free Out
"We only used Creative Commons and public-domain material" is the most confident wrong answer in AI content review. It is wrong in four separate ways, and three of them are cheap to fix at ingestion and expensive to fix after the model is trained.
"Creative Commons" is not a permission level
It is a naming scheme for six licences built from four elements. A dataset described as "Creative Commons" can be anything from CC-BY, which asks only for credit, to CC-BY-NC-ND, which forbids both commercial use and adaptation — a narrower grant than most paid stock licences. An ingestion pipeline that records "CC" and nothing more has recorded that a licence exists, not which one.
This matters more for AI than for a blog image, because the harm is retroactive. Discovering a bad image in a blog post is one takedown. Discovering NC material in a training corpus after the model is in production raises a question about the model, and the only clean remedy is the one nobody wants to pay for.
The four elements, read as pipeline requirements
Attribution
NEEDS A FIELD YOU PROBABLY DO NOT HAVECredit the creator, provide the licence and a link, indicate whether you modified the work, and keep any notices supplied with it. The obligation is to do this in a manner reasonable to the medium.
Almost no generation pipeline carries provenance from input to output. If you intend to rely on BY material, the attribution has to be designed in at ingestion — a credits page, a dataset manifest you can publish — because it cannot be reconstructed later.
NonCommercial
FILTER AT INGESTIONNo use primarily intended for or directed toward commercial advantage or monetary compensation. A model that supports a paid product is difficult to describe as anything else.
There is no cure for this one after training. Treat NC as a hard exclusion filter at the point of collection, and audit the filter — NC material is heavily represented in exactly the free image and audio repositories teams reach for first.
ShareAlike
UNSETTLED — TREAT AS RISKAdaptations must be licensed under the same terms. Whether a model, a fine-tune or a particular output is an adaptation of a training input is a question no court has answered.
If proprietary output is the point of your product, this is not a risk to price — it is a filter to add. If you are building something you intend to release openly anyway, SA material may be perfectly usable.
NoDerivatives
INCOMPATIBLE WITH GENERATIONYou may share the work, but not distribute a modified version of it. Fine-tuning on it, generating variations from it, or compositing it are all difficult to square with the condition.
ND material is usable as reference a human looks at. It is not usable as pipeline input. The distinction is worth writing into the policy explicitly, because it is the one people talk themselves out of.
The licence terminates automatically, and 4.0 gives you a window
A Creative Commons licence ends by its own terms if you fail to comply with it. Under the 4.0 suite it reinstates automatically if you cure the failure within 30 days of discovering it — a genuine improvement over earlier versions, where termination was permanent absent express reinstatement. The practical consequence is that noticing matters. A missing attribution you find and fix inside a month is a non-event; the same omission surfaced by the creator two years later is a use that has been unlicensed since the day it started, which is how statutory damages arguments begin.
Four labels that all read as "free" and are not the same
CC0
An affirmative dedication by someone who held rights and gave them up, with a permissive fallback licence where waiver is not possible. The strongest of the four, and still no warranty that the dedicator owned what they dedicated.
Public Domain Mark
A third party's assertion that a work is already free of copyright. It is a label, not a grant, and the person applying it usually had no relationship to the work. Verify the underlying facts — the author's death date, the publication date — before relying on it.
A reproduction of a public-domain work
The underlying painting is free; the photograph of it is a separate question that turns on jurisdiction and on how much creative judgement the photographer exercised. Many holding institutions also impose contractual download terms that bind you independently of copyright.
“Free to use” on an aggregator
A site's own terms, which can be withdrawn, can be wrong about the uploader's rights, and frequently exclude exactly the use you want — AI training is now an explicit carve-out on several large libraries. Record the terms as they read on the day you downloaded, not the day you are asked.
The record you need is small, and it has to be made at ingestion
Everything above collapses into a manifest. For each asset entering the pipeline, capture the source URL, the exact licence identifier and version, the creator string the licence expects you to reproduce, the date of collection, and a copy of the terms as they read that day. Five fields.
The reason to do it at ingestion is that it is the only moment the information exists. After the corpus is assembled, the answer to "where did this come from" is an afternoon of guessing per asset, and the honest result of that exercise is usually a decision to rebuild the corpus. A team that can produce the manifest on request has a defensible position even where the law is unsettled; a team that cannot is negotiating without knowing its own facts.
Where the attribution actually has to appear
"Reasonable to the medium" is the standard, and for a website that means a credits page or a per-asset caption a reader can reach. It is worth checking that the page you chose is one a human can use: credits rendered as an image, buried behind a control with no accessible name, or in a collapsed panel that a keyboard cannot open are attribution you have technically published and functionally withheld. The licence asks for credit that reaches the reader, not credit that exists in your HTML.
Can a reader actually reach your credits page?
Attribution only counts if it is reachable. Scan your credits or licence page free and see whether the links have accessible names, whether a keyboard can open the panel the credits live in, and whether any of it is locked inside an image. It takes about thirty seconds.
Scan Your Credits Page for Free →Frequently Asked Questions
We only used CC-licensed material. Is that not the safe option?
It is safer than scraping, and it is not the same as unrestricted. A Creative Commons licence is a copyright licence with conditions; using the work outside those conditions is ordinary infringement with the added problem that you have a signed-in-advance record of the terms you broke. 'Creative Commons' names a family of licences ranging from nearly-anything-goes to narrower than most stock agreements.
Does the attribution requirement really follow content through a generator?
The condition attaches to your use of the licensed work, so it follows the use rather than the pixel. If a CC-BY photograph is in a training set, in a retrieval index, or dropped into a composite, you have exercised rights under the licence and the attribution condition applies to that exercise in the manner reasonable to the medium. Most pipelines have no field to carry it — which is a design gap, not an exemption.
Does NonCommercial block our internal use? Nobody is paying for it.
NonCommercial excludes use primarily intended for or directed toward commercial advantage or monetary compensation. Training a model that supports a commercial product is fairly read as directed toward commercial advantage even if the model itself is never sold and the output is internal. Treat NC material as unusable in any pipeline that touches revenue, and filter it at ingestion rather than at review.
If we train on ShareAlike material, do our outputs become ShareAlike?
This is genuinely unsettled and should be treated as a risk rather than a resolved question. SA obliges you to license adaptations under the same terms, so everything turns on whether the model, the fine-tune or a given output is an adaptation of the input. Reasonable lawyers disagree. If your business depends on proprietary outputs, the cheap answer is to exclude SA material rather than to bet the product on a position no court has confirmed.
Is CC0 the same as public domain?
Close enough for most purposes, and not identical. CC0 is an affirmative dedication by a rights holder who had rights to give up, with a fallback licence for jurisdictions that do not permit waiver. The Public Domain Mark is a label applied by a third party asserting that a work is already free of copyright — it is an assertion about someone else's work, it carries no warranty, and it is frequently wrong.
The painting is from 1650. Surely the photograph of it is fine?
The painting is out of copyright; the photograph is a separate work and the answer depends on jurisdiction and on how much creative judgement went into it. US practice has long treated a faithful two-dimensional reproduction as lacking the originality for its own copyright, while several European jurisdictions and many holding institutions take a different view and enforce contractual terms on the download regardless. Record where the file came from and what its terms of use said, because that is a contract question as much as a copyright one.