How to Cite AI Generated Images?

Why AI Image Attribution Is Now a Publishing Problem

If you've ever picked an AI-generated image for a blog post or ad campaign, you've probably felt the quiet tug-of-war that comes with it. These images often perform brilliantly — and that's exactly the problem. A peer-reviewed study in the International Journal of Research in Marketing found that AI-generated banner ads achieved up to 50% higher click-through rates than human-made images, after generating 10,320 synthetic marketing images and running more than 254,400 human evaluations.

So the images that work best are the ones that, when disclosed, take a hit. In a separate peer-reviewed study, AI and human-made images performed comparably when their origin wasn't disclosed — but once AI use was revealed, consumers systematically penalized the AI versions, with a notable effect size favoring human-made images. Disclosure activates what researchers call persuasion knowledge and algorithm aversion: readers start asking what the image is trying to sell them.

This creates a genuine strategic dilemma rather than a paperwork formality:

• Performance pulls toward silence. Undisclosed AI images match or beat human-made ones, so there's a built-in incentive not to mention them.
• Trust pulls toward disclosure. Hiding AI use risks audience backlash if discovered, and the research shows transparency norms are hardening.
• The law pulls toward documentation. The U.S. Copyright Office maintains a human authorship requirement, and purely AI-generated images are not copyrightable — with prompt engineering explicitly not conferring authorship.
• Provenance tools help, but don't solve it. Embedded signals like C2PA Content Credentials can be stripped by screenshots or file conversion, and their absence doesn't prove an image is human-made.

The copyright angle makes this more than a branding question. In the Zarya of the Dawn case, the Copyright Office rescinded protection for Midjourney-generated images in a registered comic book, protecting only the human author's arrangement of the work. Meanwhile, at least 16 lawsuits have been filed against major AI companies over copyright in the past two years — a sign that provenance disputes are moving from theory to courtroom.

The practical takeaway from the research: written attribution in the publication itself is the reliable layer, because metadata and watermarks are fragile. And how you frame disclosure matters — the consumer study found that explaining a protective motivation for using AI (such as privacy) largely erased the penalty, while cost-efficiency framing worsened trust and purchase intention.

The same principle applies beyond images. Verification beats assumption everywhere in AI workflows — which is why SkillzBench is building free, security-scanned benchmarking for agent skills, so you can see what a skill actually does before you rely on it. You can join the waitlist to get access when it launches. Attribution, like skill selection, works best when it's deliberate rather than assumed.

What the Research Says: Disclosure, Framing, and Consumer Trust

If you disclose that an image is AI-generated, expect your audience to like it less — unless you tell them why you used it. That counterintuitive finding comes from a peer-reviewed series of experiments on consumer reactions to AI imagery, and it reshapes how publishers should think about attribution.

The baseline result is sobering. In one experiment, undisclosed AI-generated images performed just as well as human-made ones with consumers. But once AI use was disclosed, attitudes shifted significantly toward human-made images — a consumer study measured the effect at d = 0.52 favoring human work. Disclosure alone, in other words, activates what researchers call persuasion knowledge and algorithm aversion, and consumers penalize accordingly.

The penalty isn't fixed, though. The same research found that framing the motivation behind AI use changes everything:

• Framing AI use as protecting privacy (e.g., avoiding photos of real people) led to evaluations comparable to human-made images — the penalty disappeared.
• Framing it as cost efficiency caused significant declines in both trust and purchase intention — the penalty got worse.
• Framing AI as augmenting rather than replacing human work also softened negative reactions tied to diminished perceived authenticity.

The practical takeaway: disclose proactively and explain why. "Image generated with AI" is the bare-minimum framing the research suggests will cost you trust. "Image generated with AI to avoid using identifiable photos of real people" is the kind of motivation framing that eliminates the penalty. Hiding AI use entirely is not a strategy either — provenance signals like C2PA Content Credentials and SynthID watermarks are increasingly embedded in generated images, and while OpenAI's documentation notes these signals can be lost through screenshots or file conversion, the direction of the ecosystem is toward verifiable disclosure, not concealment.

This matters more as AI imagery becomes the norm. A large-scale marketing study found AI-generated banner ads achieved up to 50% higher click-through rates than human-made images — so the quality and performance incentive to use AI images is real, and the trust question won't go away.

The same verification-first mindset applies beyond images. Just as provenance standards verify where an image came from rather than trusting a file at face value, agent skills — the instruction packages powering AI coding agents — need verification before you rely on them. That's the problem SkillzBench is building for: a free directory where agent skills are security-scanned, version-pinned, and benchmarked, with a waitlist now open for anyone who wants to equip their agents with skills they can actually trust.

The Attribution Toolkit: Provenance Metadata, Watermarks, and Written Citations

A visible credit line takes seconds to write; invisible provenance metadata can vanish the moment someone screenshots your image. That asymmetry is why serious publishers treat attribution as a layered system rather than a single label.

The technical layer starts with C2PA Content Credentials, an open standard the coalition behind it describes as "like a nutrition label for digital content" — a machine-readable record of where an image came from and how it was edited, available for anyone to inspect at any time (c2pa.org). OpenAI embeds a dual signal in its generated images: C2PA metadata plus an invisible SynthID watermark, precisely because metadata can be stripped by platforms during upload or conversion, while watermarks may survive some transformations (OpenAI's provenance documentation).

Free verification tools make this layer practical. OpenAI offers a public checker at openai.com/verify and a Content Provenance API for developers, currently free with usage limits (per its help center). Detection is far from perfect, though — text watermark detection rates at a 1% false-positive threshold range from 69.0% in Spanish down to 42.2% in Romanian.

Here is the critical limitation, and it is research-backed: OpenAI itself states that provenance signals do not guarantee content is accurate, unedited, legally owned, or correctly contextualized, and that the absence of a signal proves nothing. A screenshot or file conversion can strip Content Credentials entirely; heavy editing degrades watermarks. Embedded signals are supporting evidence, not the attribution itself.

That leaves written attribution in the publication as the reliable layer:

• State AI generation in a caption, credit line, or disclosure note that travels with the image wherever it is republished.
• Name the generating tool (e.g., DALL-E 3, Midjourney v6) rather than a generic "AI-generated" label.
• Frame the motivation where relevant — research shows privacy-protective framing eliminates consumer penalties, while cost-efficiency framing damages trust (a peer-reviewed consumer study).
• Keep provenance checks as corroboration: verify credentials before publication, and note the result alongside your written credit.

The same verify-don't-trust logic applies beyond images. SkillzBench is building a free directory that security-scans and benchmarks agent skills before you rely on them — worth a look if AI now fills your content workflow.

The Legal Reality: Copyright, Human Authorship, and Disclosure Obligations

If you publish an AI-generated image under your own byline, you may be publishing something nobody owns. That is the uncomfortable starting point for attribution in the United States, where the Copyright Office reaffirmed its long-standing human authorship requirement: outputs generated wholly by AI are not copyrightable at all.

The Office has been equally clear that prompt engineering does not confer authorship. Because the GenAI model controls or determines the output rather than the human prompter, typing detailed prompts does not satisfy the human authorship requirement. In the Zarya of the Dawn precedent (February 2023), the Office rescinded protection for Midjourney-generated images in a comic book, protecting only the human author's selection, coordination, and arrangement of those images.

Publishers face a disclosure obligation, not just a stylistic choice. The Copyright Office's official Registration Guidance for Works Containing AI-Generated Materials (88 Fed. Reg. 16,190, March 16, 2023) requires applicants to disclose AI-generated content in registrations, and the Office relies on voluntary disclosure from rights seekers to identify which elements are excluded from protection. Over 10,000 comments poured into the Office's August 2023 Notice of Inquiry by December 2023, signaling how contested this territory remains.

The stakes are visible in the litigation landscape: at least 16 lawsuits have been filed against almost every major AI company over copyright infringement in the past two years. For publishers, that means documenting rights differently depending on how an image was made:

• Purely AI-generated images carry no U.S. copyright under current guidance, so your documentation should record the tool, prompts, and date rather than assume ownership.
• AI-assisted composite works protect only the human contributions — selection, arrangement, and substantive edits — which must be identified and disclosed in any registration.
• Licensing records should state which elements are AI-generated, since those parts are unprotected and cannot be exclusively licensed.
• Written attribution in the publication itself matters more than embedded metadata, which can be stripped by screenshots or file conversion.

The law diverges sharply outside the U.S., so attribution practices should follow the jurisdiction of publication. The UK grants copyright in computer-generated works to whoever arranged the creation, while China recognizes copyright when prompts are sufficiently detailed. An image that is unowned in New York may be protectable in London or Beijing — which is precisely why consistent disclosure of AI-generated elements is the safest editorial policy across borders.

Verification is the underlying discipline here, and it extends beyond images: the same skepticism that applies to "who made this picture?" applies to "can this agent skill be trusted?" SkillzBench is building exactly that for agent skills — pinning every version to a commit SHA, security-scanning, and benchmarking in a sandbox before anything earns a trust score, all free when it launches. Join the waitlist to be first in line.

Your Attribution Playbook: A Practical Checklist

Attribution for AI-generated images works best as a repeatable process, not a one-off judgment call. The evidence points to a handful of concrete steps that protect both your credibility and your legal position.

First, disclose AI use in a visible caption or credit line — not buried in metadata. Consumer research across three experiments found that when AI use is disclosed, consumers systematically penalize AI-generated images, with a significant effect favoring human-made images. Hiding the disclosure doesn't avoid that penalty; it just adds a trust violation on top of it.

Second, frame the motivation behind the AI use. The same research found that disclosing a protective motivation — such as privacy protection — produced evaluations comparable to human-made images, while cost-efficiency framing caused significant declines in trust and purchase intention. A caption like "generated to avoid using identifiable customer imagery" lands very differently than "generated to save money."

Third, keep provenance metadata intact where possible and verify it with available tools. C2PA Content Credentials function like a nutrition label for digital content, documenting origin and edits. OpenAI images carry both Content Credentials and invisible SynthID watermarks, and the company offers free verification at openai.com/verify. But treat these as supporting evidence, not proof: metadata can be stripped by platforms, lost through screenshots or file conversion, and — critically — an absent signal does not prove content is human-made.

Fourth, disclose AI elements in copyright registrations. The U.S. Copyright Office's 2023 registration guidance requires disclosure of AI-generated material, and for good reason: purely AI-generated images are not copyrightable, and prompt engineering does not confer authorship. In the Zarya of the Dawn case, protection was rescinded for Midjourney-generated images in a comic book, covering only the human author's arrangement. It's tempting to omit AI elements since those parts lose protection, but undisclosed claims put your entire registration at risk.

Finally, never rely on a missing watermark as proof of human origin. OpenAI states plainly that a watermark is only evidence an AI model likely generated content — it does not establish authorship, ownership, or who is legally responsible.

Your practical checklist:

• Add a visible AI disclosure in the caption or credit line, framed around privacy or quality — not cost-cutting
• Preserve C2PA metadata and verify provenance with available tools before publishing
• Disclose AI-generated elements in any copyright registration
• Treat a missing watermark as inconclusive, never as evidence of human authorship

The underlying principle is the same one we're building SkillzBench around: verification beats trust. Just as image provenance demands scanning and checking rather than taking a file at face value, agent skills demand security-scanning and benchmarking — published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is building exactly that: a free directory where every skill is pinned to a commit SHA, security-scanned, and benchmarked for real performance lift, plus a free MCP tool to connect verified skills to your agents. You can browse the free skills directory and join the waitlist while the platform finishes its launch build.