Tuning Into Reality
Connecting to the world
ExperiencesConnecting to the world
You see AI outputs that resemble your work, or hear that a provider collected material from a site where you publish. Can you find out whether your own work was used to train a model?
Build an evidence trail with a specific provider, model, dataset, and version where possible. Similar output, an AI-origin label, and a record in a dataset are different kinds of evidence. None should be substituted for the others.
This question feels personal, but the evidence is often buried in systems you cannot inspect. Public model descriptions may discuss broad categories of material rather than listing individual works. A plausible concern therefore needs a more careful conclusion than either “the model resembles me, so it trained on my work” or “I found no record, so it did not.”
It helps to treat the investigation as a sequence of connections. First identify the work. Then identify any collection record. Then investigate what is known about the use of that collection. Finally, separate those findings from any complaint about a particular output. You may be able to document some connections and leave others unresolved.
Write down the suspected use. Was your song uploaded by a user as reference audio? Was your artwork included in a published dataset? Is a chatbot retrieving a current page from the web? Or do you suspect the provider used your work during model training? A request becomes more useful when it identifies the operation you are asking about.
The US Copyright Office’s AI report series treats digital replicas, copyrightability, and generative-AI training as separate subjects. Its legal analysis is US-specific and does not settle your rights worldwide.
For investigation purposes, use “collection” for obtaining a copy or record of material, “training” for its use in developing a model, “reference input” for material supplied in a particular generation workflow, and “retrieval” for bringing external material into a response. These are working distinctions to make your inquiry precise, not a legal classification of a provider’s conduct.
A screenshot showing your artwork as an input in a user’s generation project would support a different inquiry from a record showing your image URL in a named dataset. A chatbot quoting a current web page also raises a different question from whether that page was included in a historical training run. Preserve the evidence you have without assigning it to a different operation.
Useful leads include a dataset entry containing your work’s URL, a documented dataset version, a provider’s model documentation, a license, an upload record, or correspondence confirming a particular use. Preserve the full page, identifier, and date, not just a cropped search result.
When a dataset search service is available, record exactly what collection it searches. A match can show that a record appears in that searched collection. Connecting it to a particular trained model requires additional evidence about that model’s data and version. No result only describes the searched coverage; it does not exclude unindexed or private datasets.
As checked on 11 November 2026, Have I Been Trained? displays an “Under Maintenance” page. Do not make an investigation depend on that service being operational. Recheck its current availability and scope before using it.
Consider a fictional illustrator who finds a record pointing to her portfolio image in a searchable collection. She saves the dataset name and version, the entry, the image URL, and the date of her search. She checks that it refers to her actual work rather than a similarly named image or a thumbnail of something else.
That creates a documented lead about the collection. If a model’s documentation names a different dataset version, she cannot simply treat the versions as interchangeable. The next question is whether documentation, provider confirmation, or another reliable record connects the identified entry to the relevant model’s data. The search result may be valuable even when that connection cannot be established publicly.
A search may also reveal only a URL or descriptive record. Preserve what the record actually contains and avoid claiming it proves a particular stored file or use unless the supporting documentation says so. Do not try to access restricted datasets or private accounts to fill gaps.
Write down the service, collection, search terms or image used, date, and stated coverage. This prevents “I searched one index” from becoming “I searched all training data.” It also makes the investigation easier to revisit if the service expands its coverage or a provider releases new documentation.
For music, the Suno credentials checker looks for provider metadata, rather than identifying which songs were used as training data. For images or video, provenance records can supply available creation history, but missing provenance indicators do not establish authenticity. An output resembling your work remains something to investigate, rather than a complete record of model development.
Use the provider’s current privacy, rights-holder, or support channel. Identify the works, original URLs, dates, and any dataset records. Ask: “Do your records show collection or use of these identified works? If so, for which purpose and model or dataset version? What process can I use to challenge that use or request exclusion?” Request clarification when the response only describes general policy.
If your concern involves personal data, such as information identifying you, consider whether a local access, objection, or deletion process applies. This is a separate route from a copyright complaint. The country help directory links to privacy authorities. Eligibility and exceptions depend on the applicable law; these requests do not automatically entitle you to a complete training manifest.
Send a manageable list of clearly identified works. Include the original URL, title, publication period, relevant identifiers, and the specific evidence that prompted the request. If the content moved between domains, explain that history. Ask about the named model or collection rather than demanding an answer about every product the company has ever operated.
Keep copyright and personal-data requests distinct. A request concerning a photograph that identifies you may raise a personal-data question as well as an ownership question. A landscape painting may involve your copyright concern without making every aspect of the image personal data about you. Describe your relationship to the material and follow the applicable request process.
If the provider gives a general answer, ask which part of your specific inquiry it addresses. For example: does a promise to honor future exclusions cover the URLs you listed, a named dataset, or a particular product? Keep the original request and the reply together so an adviser can assess what remains unanswered.
Do not assume a future-use opt-out, deletion of a hosted file, and removal of a work’s influence from an already trained model are the same operation. Ask for a written description of the action and its scope.
The result may be a confirmed record, a documented exclusion request, an explanation of what a provider can remove, or an unresolved question. Record the scope of any confirmation. “The provider acknowledged my request” is different from “the provider confirmed my work was used,” and both differ from a finding about whether the use was lawful.
If you want compensation or a binding legal remedy, an unresolved support exchange may need professional assessment rather than more repeated submissions. Bring your evidence trail, contracts, country information, and desired outcome. Use the country help directory to locate appropriate starting points; no single global complaint process resolves every training-data dispute.
Generating many outputs until one resembles your work can produce a large folder without clarifying the source of the resemblance. If you already have an important output, preserve its available generation context, inputs, settings, and date. A selected example without that context is harder to interpret.
Focus first on records that connect the work to the suspected use. If you commission a technical investigation, ask what evidence the method requires, which alternatives it can distinguish, and how uncertainty will be reported. A claim that a tool can prove training membership needs more scrutiny than a claim that it can locate a visible similarity.
Maintain a dated record of your originals, licenses, dataset leads, provider requests, and replies. If a particular output also copies identifiable material, investigate that output through the music or artwork guide in parallel. You can describe and report the observed use without asserting a training history you have not established.