How AI Unmakes Images: What Photographers Must Know About Copyright, Training Data & Creative Control
The talk at The Photographers' Gallery
A recent talk — How AI Unmakes Images: The Legal Aesthetics of Copyright — unpacks a deceptively simple idea: generative AI doesn’t “copy” images the way we think. Instead it digests millions of pictures, learns statistical relationships (colour, texture, composition), and synthesises new images from those learned patterns. That process dissolves the idea of a single, discrete photographic object that law and markets have traditionally protected. The talk lays out why that technical reality is now a legal and economic headache for image-makers. (The Photographers Gallery)
Watch the full talk on YouTube
If you want the full presentation and the examples the speakers use, the video is available online — it’s a clear primer for anyone working with images today. The speakers walk through not only the tech, but the political and property questions that follow when an image becomes “training data.” (YouTube)
What “unmakes images” actually means
When an AI model trains, it doesn’t tuck away whole photo files for later replay. It internalises statistical patterns across datasets: the way light hits skin, the grain of film, compositional rules. When you prompt a model, it generates something new from those internal patterns. From a photographer’s perspective that’s both fascinating and worrying — “new” images can feel eerily similar to your style or even accidentally reproduce unique elements (watermarks, compositional signatures). Because there’s no neat copy to compare, the usual tests for infringement — “is there a substantial similarity to a specific work?” — become harder to apply.
The legal landscape — still unsettled
Courts and regulators are moving fast, but they haven’t settled everything.
In the U.S., an appeals court recently affirmed that works created solely by machines without meaningful human creative input are not eligible for copyright — a reminder that human authorship remains a legal baseline. That decision signals that photographers who use AI as a tool (with clear human creative choices) are in a different bucket from fully autonomous outputs. (Reuters)
In the U.K., a headline-making case between Getty Images and Stability AI reached the High Court. The ruling was complicated: the judge found instances of trademark/watermark misuse but did not accept a broad claim that the model reproduced copyrighted images as literal copies stored in the system — leaving open the larger question of whether large-scale training on copyrighted archives is itself an infringement. The case is widely read as important but inconclusive. (Courts and Tribunals Judiciary)
That mix — some rules reinforced, other questions left open — means creators should plan for uncertainty. Major platforms and courts are still figuring out where liability, licensing, and transparency obligations land.
Practical steps photographers can take today
You can’t change the legal fog overnight, but you can take steps that both protect your work and strengthen your position if a dispute arises.
Document everything. Keep original RAW files, layered edits, metadata and timestamps. If your work is registered, keep those records handy — administrative proof matters.
Record your human contribution. If you use AI tools, save prompts, iterations, and notes about curatorial choices. Clear evidence of direction, selection and editing strengthens claims of authorship.
Watermark and track — strategically. Watermarks aren’t perfect, but they make wholesale scraping and reuse riskier and provide visible provenance. Consider forensic watermarking for high-value archives.
Prefer platforms with opt-outs or licensing offers. When uploading work, read terms carefully. Some platforms offer explicit opt-outs for dataset training; others claim broad rights. Negotiate or choose services that respect creator control.
Join collectives and push for transparency. Collective bargaining and public campaigns are already pushing platforms to disclose dataset sources and licensing models. There’s power in numbers.
Treat style as a brand, not a secret. If a generated image mimics your distinct style, you may have rights under trademark-like theories or unfair competition rules in some jurisdictions. Document what makes your style distinct.
What publishers and platforms should do (and why it matters to you)
Platforms and marketplaces are the first line of either protecting or exploiting photographic work. Good platforms will:
Offer clear toggles for creators to opt in/opt out of dataset use.
Publish dataset provenance statements (what was used and whether consent was obtained).
Support metadata standards so “human authorship” flags travel with an image.
When platforms adopt those practices, photographers regain control — and the market can better distinguish licensed, creator-approved uses from extractive scraping.
Final note — adapt and document
The phrase “How AI Unmakes Images” is provocative because it captures a practical truth: the technology changes what an image is in legal and economic terms. But technology also creates new opportunities — new tools for expression, new licensing models, and new marketplaces. The smart move for photographers is not to bury their heads, but to double down on the fundamentals that matter in court and commerce: provenance, evidence of human creativity, sensible contracts, and collaboration with other creators.
If you shoot, edit, curate or sell images, start today by auditing where your work lives online, saving origin files and asking platforms where your images might end up. This is a fast-moving conversation, but documented creators will be the ones best placed to shape the rules to come.
.png)
