Executive summary
There are specific tags and directives that creators can attach to published works to signal it should not be used for machine learning training, including robots.txt disallow directives, noai and noimageai meta tags and X-Robots-Tag HTTP headers.
However, these restrictions have varying legal weight in different jurisdictions, are sometimes disregarded (intentionally or unintentionally) and the works may ultimately be incorporated into model training.
Both creators and downstream users of trained models should be mindful of the resulting long-tail risks that are difficult to identify when using the models.
Introduction
Independent creators use online platforms to share and monetize work as model repositories, digital galleries, photography portfolios, and code-sharing communities provide discovery spaces. Visibility can support commissions, licensing, and community engagement.
Some platforms are now offering a “NoAI” metatag option as part of the publication process that can be opt-in or opt-out, where a creator, can select licence options (e.g., similar to publishing under different variants of Creative Commons licences). These restrictions typically rely on voluntary compliance as opposed to technical enforcement.
An example metatag can indicate: "NoAI: This model may not be used in datasets for, in the development of, or as inputs to generative AI programs." Robots.txt similarly lists the types of bots and their permissions (or restrictions).
Automated crawlers often collect public content at scale, creating large datasets for training machine learning models. The content becomes numerical representations, and model parameters absorb features into model weights, and often the original works are not stored as a retrievable file.
However, recent complaints and litigation have been raised in scenarios where automated crawlers have allegedly ignored the NoAI and similar restrictions, and it is important for end users of the machine learning models to be mindful of the long-tail risk introduced by their vendors.
Canadian considerations
Currently, Canada’s Copyright Act contains no text and data mining exception and no equivalent to a European Union style opt-out framework.
A Canadian creator’s tag or directive does not have the legal operation that a valid reservation can have in other jurisdictions with formalized rights. Developers must instead rely on fair dealing under section 29 and following, confined to enumerated purposes including research, private study, criticism, review, and news reporting, and on section 30.71 for temporary reproductions essential to a technological process.
Whether large-scale ingestion to build a training corpus qualifies as “research” has not been decided by a Canadian court; the position is unsettled.
The federal Copyright in the Age of Generative Artificial Intelligence consultation and resulting “what we heard” report confirmed existing legislation does not clearly address text and data mining or large-scale model training.
Submissions were divided between a standalone text and data mining exception or broadened fair dealing and licensing frameworks with remuneration for creators.
For Canadian creators looking to protect their rights in a scenario where their work has been misappropriated, the strongest levers may be contractual rather than statutory: website terms of use, platform terms, and licence conditions. Technological protection measures should also be considered, as they provide a separate ground for recourse.
There is pending litigation in Canada regarding model training that may result in changes in jurisprudence.
US developments
In the United States, litigation trends have been shifting from training in the abstract toward dataset provenance, meaning where a training corpus came from, how it was assembled, and what records establish chain of custody. The recent complaints also consider whether attribution, licence terms, and other rights management information travelled with a work through collection and ingestion.
There are a number of ongoing disputes regarding three-dimensional models against technology companies, raising allegations concerning the Digital Millennium Copyright Act’s copyright management information provisions, rather than as conventional infringement claims.
Creators upload three-dimensional models to public repositories under licences that typically require attribution and may restrict commercial use or impose share-alike conditions. Dataset curators compile the assets into large research datasets. Commercial developers then download, render, and ingest the assets into generative three-dimensional systems while stripping attribution and licence metadata needed to preserve the applicable licence terms.
The affected datasets are model datasets with millions of 3D object assets. The complaints allege that each work from which information was removed is a separate violation, so statutory damages can scale across an entire dataset in a way conventional infringement claims may not. They emphasize that the dataset preserved links back to source repositories and creator accounts, addressing the argument that attribution at scale is impractical.
Practical steps for small creators
For independent creators, these steps may improve practical control:
- Use layered signals: robots.txt rules, noai/noimageai tags, and X-Robots-Tag headers for non-HTML files.
- Date each reservation, including when the signal was deployed and what it covered.
- Consider sharing only at reduced quality or behind access controls. Access restrictions may reduce reach and commercial benefit, so balance those effects against preferences regarding automated use.
- Use licence terms addressing machine-learning use and platform opt-out toggles where available.
- Creators should put machine-learning restrictions in site terms of use and licence terms, not rely on tags alone, and try to preserve attribution and licence metadata in the file itself, not only on the surrounding webpage.
Downstream risk for model users
For downstream trained model users, adopting a model or widely used public research dataset can carry exposure inherited from its assembly, so dataset-level diligence matters as much as model-level diligence.
Material can enter a corpus through third-party datasets, web-scale crawls, mirrored copies without a “NoAI type opt-out” signal flag, or collection before deployment. Opt-out indicators may have applied at site or file level and are sometimes lost when copied. A good-faith crawler may therefore ingest tagged material unintentionally, leaving a risk that noai material was included.
The risk that the output is very similar to a single work is greater for more specific prompts, as there may be a limited number of training examples for a narrower domain.
Widely used public research datasets are a common point of inherited exposure. Many organizations have built on the same small number of corpora, so diligence should identify the specific datasets underlying a model and whether any are the subject of pending claims.
The risk can be somewhat managed through documented complaint, triage, takedown, output-suppression, filtering, escalation, and retraining or substitution procedures.
Another approach for risk management includes using IP insurance, and coverage should be assessed for training, output, defense, and regulatory claims. Contracts with model providers can also be negotiated to include IP warranties, indemnities, defense control, and pass-through protections.
This publication provides general information, not legal advice. Legal positions vary by jurisdiction; readers should seek advice for their circumstances.