Several approaches to giving creators control over training use have emerged, with limited coverage.
Robots and crawler directives
Signals asking crawlers not to collect.
Which are voluntary and respected inconsistently.
Opt-out registries
Databases of works declining training use.
Which depend on developers checking them.
Licensing agreements
Deals between developers and rights holders.
Which cover large catalogues and not individual creators.
Legislative proposals
Requirements for consent or remuneration.
Which several jurisdictions are considering.
Why none of them works comprehensively
Crawler directives are voluntary, registries require developers to consult them, and licensing deals cover large catalogues while individual creators have no practical route.
Which leaves most working artists with no effective mechanism.
Proposals for statutory consent or remuneration would change that, and they face the practical difficulty that training datasets were assembled before any of these mechanisms existed.
Retroactive problems
Models already trained on material collected earlier.
Which no forward-looking mechanism addresses.
Machine unlearning
Research into removing specific data from trained models.
Which is technically difficult and not currently practical at scale.
Collective licensing
Models from music and reprography being proposed.
Where this is heading
Legislative rather than technical resolution.
Why these questions are unusually hard to settle
Creative work has always involved learning from other people’s work, and the law reflects that: ideas and styles have never been protected, only specific expression. That principle exists for good reasons and nobody wants it reversed.
What has changed is that a capability which previously required years of practice now requires a sentence, at essentially unlimited volume. Whether that is a difference of degree or of kind is the actual argument, and it is not resolvable by pointing at existing law, because existing law was not written with this in view.
Where the practical action is
Not, so far, in courts. Collective bargaining has produced enforceable terms faster than litigation has produced rulings, provenance standards have advanced faster than detection, and platform and publisher policies have moved faster than legislation.
That means the rules governing this in practice are currently a patchwork of contracts, platform terms and professional norms, all of which vary and none of which is comprehensive. Anyone commissioning, producing or publishing creative work is operating inside that patchwork whether or not they have noticed.
A general note
Copyright, contract and likeness law differ substantially between countries and are developing quickly. Nothing here is legal advice, and anything with money or rights attached warrants a qualified adviser familiar with the relevant jurisdiction.
What working creatives actually seem to want
Across surveys, union negotiations and professional body statements, the demands are fairly consistent: consent before their work is used in training, some form of compensation when it is, credit where their contribution persists, and disclosure so that clients and audiences know what they are looking at.
None of those is a demand that the technology stop existing, which is worth noting given how the debate is frequently characterised. They are requests for the same terms that apply when any other use is made of someone’s work.
The argument on the other side
Requiring consent for training would be practically impossible for datasets already assembled, would advantage large companies able to license at scale over smaller developers and researchers, and would sit awkwardly with a legal tradition that has never required permission to learn from published work.
Both positions are held seriously by people who have thought about them, and the resolution is likely to be legislative and messy rather than clean in either direction.
What to do in the meantime
If you commission creative work, specify what tools may be used and what disclosure you expect. If you produce it, be clear with clients about what you did. If you publish it, have a policy and state it. If you consume it, check provenance rather than trying to spot artefacts.
All four are available now, none depends on the law being settled, and between them they cover most of the situations that actually arise.
Where to follow this properly
Professional bodies for the relevant discipline publish position statements, survey data and practical guidance, and they are considerably better informed than general coverage. Court filings and judgments in the major cases are public. Standards bodies publish their specifications openly.
All of that is more useful than commentary, and most of it is free. The field is moving quickly enough that anything written more than a year ago should be checked against the current position before it is relied on.
One closing observation
Most of the heat in this argument comes from treating a set of separate questions as one. Whether training was lawful, whether outputs can be owned, whether a style can be protected, whether disclosure should be required and whether the market effects are acceptable are five different questions with five different answers.
Separating them does not resolve any of them, and it makes the disagreement considerably more tractable than it appears when they are argued together.