Generative software that reworks a photograph into something resembling a centuries-old portrait has quietly become common in marketing, publishing, and personal use. The technology itself is straightforward enough to describe in a sentence. The legal questions it raises are not, and they now sit squarely in front of lawyers advising creative agencies, advertising firms, publishers, and individual artists.
Why Authorship Doctrine Is Under Pressure
Copyright law was built around a presumption that a human being made the creative choices behind a work. That presumption is now being tested directly. In the United States, the Copyright Office has repeatedly declined to register works produced by a machine without meaningful human creative input, a position reinforced by the ongoing Thaler v. Perlmutter litigation, which confirmed that a work lacking human authorship falls outside the scope of the Copyright Act entirely. The Office has since issued guidance clarifying that AI-assisted works can still qualify for protection, but only to the extent of the human-authored elements, meaning a generated image may be only partially protectable, and courts have not yet settled how that partial protection should be applied in practice.
Other jurisdictions have taken different paths. The United Kingdom’s Copyright, Designs and Patents Act contains a provision, section 9(3), that assigns authorship of a computer-generated work to the person who made the arrangements necessary for its creation, a solution that predates modern generative AI by decades and is now being revisited for how well it actually fits. The European Union has generally maintained that copyright requires an “author’s own intellectual creation,” a standard that machine output struggles to satisfy on its own. This divergence is not academic. A marketing agency producing AI-assisted imagery for a campaign running in the UK, the US, and continental Europe simultaneously may hold enforceable rights in one market and effectively none in another, with real consequences for licensing, enforcement against copycats, and contractual warranties given to clients.
Drawing the Line Between AI-Generated and AI-Assisted
The degree of human involvement is doing most of the legal work in this area, and lawyers need a working framework for evaluating it rather than a binary label. Selecting a source image, iterating on instructions, choosing among multiple outputs, and manually editing the result afterward are all activities that can push a piece of work from “generated” toward “assisted,” which is the category more likely to attract protection under current US guidance. Consider a practical example: a designer uploads a photograph to a generative editing tool that reimagines it in a historical portrait style. If the designer simply accepts the first result, that output likely lacks the human authorship needed for protection. If the same designer selects a source image deliberately, tests several variations, and substantially edits the final composition before it goes into a client deliverable, the analysis shifts toward the kind of human creative control that supports a copyright claim. Advising clients on which side of that line their workflow falls is quickly becoming routine work for IP counsel, and it is a question that will only get harder as the tools improve and the human input becomes less visible.
Style Versus Substance: What Copyright Actually Protects
A recurring source of confusion, including among clients, is the difference between imitating a style and infringing a specific work. Artistic style is not, on its own, protectable subject matter under copyright law in the US, UK, or EU. Nobody owns the concept of chiaroscuro lighting, painterly brushwork, or a period-appropriate color palette, regardless of how closely that aesthetic is associated with a particular movement or artist. This is precisely why output modeled on a broad historical style tends to sit on firmer legal ground than output that reproduces a specific, identifiable work.
The harder question lies upstream, in how the underlying models were trained. Litigation including Getty Images v. Stability AI in the UK and multiple consolidated cases against AI developers in the United States centers on whether ingesting copyrighted images at scale to train a generative model constitutes infringement, or whether it falls within fair use, text-and-data-mining exceptions, or similar doctrines. These cases remain unresolved as of this writing, and their outcomes will determine not only how AI companies must license their training data but how much downstream legal exposure ordinary corporate users absorb simply by relying on the resulting tools. Firms with clients in advertising, publishing, or design should be monitoring this litigation closely, since indemnification clauses in software licensing agreements are becoming a live negotiating point precisely because of this uncertainty.
Practical Guidance for Advising Clients
A few points are worth building into standard advice for clients experimenting with generative tools. Commissioning agreements and creative service contracts should now specify whether AI tools were used in production and allocate responsibility if a resulting image is later challenged on infringement grounds. Clients producing AI-assisted commercial material should retain records of prompts, source materials, and subsequent edits, both to support a copyright claim where one is available and to demonstrate good faith if a dispute arises. This record-keeping is easy to overlook precisely because the tools are so simple to use; a browser-based photo to renaissance painting generator, for instance, can produce a finished image in seconds, and clients rarely think to preserve the intermediate steps unless counsel asks them to. Where a business operates across multiple jurisdictions, counsel should flag that a single generated image may carry enforceable rights in one market and none in another, which has direct implications for licensing terms and takedown strategy. Finally, indemnification language in vendor contracts for AI software deserves closer scrutiny than it has typically received, given how unsettled the training-data litigation remains.
Looking Ahead
None of this is likely to resolve quickly. Legislatures in the EU, UK, and US are each considering, at different paces, whether existing copyright frameworks need statutory amendment to address generative AI directly, and courts are being asked to apply decades-old doctrine to a technology its drafters never anticipated. For legal practitioners, the task is not to wait for certainty but to help clients make defensible decisions under genuine ambiguity, and to stay current as courts and legislatures continue to redraw these boundaries.
