An Honest AI Image Policy for Client Work (What I Generate, What I Never Will)

As of September 2026, WBX Creative uses AI image generation for exactly one job: editorial illustration of atmosphere and objects, credited as illustration, on pages I control. No people, no faces, no hands, no letters, no logos, no real products, and no real place posing as itself. Nothing generated goes into a client's identity system, a product listing, a testimonial, a team page, or a photograph's job.

The policy is short because the terms behind it are not generous. The tool I run, gpt-image-2 through fal.ai, comes with no indemnity and a terms page that puts the risk on me. The U.S. Copyright Office says a prompt alone earns no copyright. And the metadata that is supposed to announce "this was generated" does not reliably survive an upload. The policy has to do the work the technology and the contracts do not.

How to think about it

Three questions decide every image, generated or not:

1. Who owns it, and what does "own" mean? Every vendor now says some version of "you own the output." Ownership under a contract and a copyright you can register are different things.
2. Who pays if it infringes? A vendor that indemnifies you pays to defend a claim. Most tools a small studio uses do not, and the ones that do carve out exactly the cases a brand cares about.
3. Can the reader tell, and would they mind? Google's guidance says disclosures are useful "for content where someone might think 'How was this created?'" A generated storefront on a business's own website is where someone would think that.

If any answer is uncomfortable, it is a photograph, an illustration by a person, or nothing.

What do the vendor licenses say as of September 2026?

Everything below comes from the vendors' own terms pages as fetched on 2026-09-22. Terms move; check before you rely on one.

Tool	Output ownership	Indemnity	Provenance signal
OpenAI, consumer ChatGPT (terms of January 1, 2026)	You "own the Output"; OpenAI assigns its rights; "output may not be unique"	None for consumers	C2PA metadata and SynthID watermark
OpenAI, API and Business/Enterprise	Customer "owns all Output"	The Service Terms cover third-party IP claims over Output; void if you "disabled, ignored, or did not use" safety features or if Output is "modified, transformed, or used in combination" with non-OpenAI products; excludes trademark claims	Same
fal.ai, gpt-image-2 endpoint (terms of September 8, 2026)	No express "you own Output" clause that I could find; licensed for "Commercial use via fal Partner agreement"	None. fal "does not represent, warrant, or covenant that any Output Content will be original, will not infringe rights of any third party"	Not stated
Adobe Firefly	Firefly models called "safe for commercial use," trained on "licensed content, such as Adobe Stock, along with public domain content"	Enterprise entitlement only, for "select Firefly outputs"; none on Standard or Pro	Content Credentials included
Midjourney (terms of May 27, 2026)	"You own all Assets You create with the Services to the fullest extent possible under applicable law"; companies over $1,000,000 a year in revenue must be on Pro or Mega to own assets	None; you grant Midjourney a perpetual, sublicensable license	Not stated
Google Gemini API, Nano Banana 2	"Google won't claim ownership over that content"; the unpaid tier is used "to provide, improve, and develop Google products"	None on the Gemini API; paid Vertex AI covers "unmodified Generated Output" only	SynthID watermark on all images

Every indemnity on the table is conditional. OpenAI's disappears once the image is modified or combined with another product, which describes every image that passes through Photoshop, and it never covered trademarks. Adobe's is an enterprise purchase. Midjourney, fal, and the Gemini API give you nothing, and Midjourney is being sued by Disney, Universal, and Warner Bros. in federal court in California; as of 2026-09-22 the docket showed a discovery cutoff of September 21, 2026 and no trial date. None of these terms mention a copyright registration, because that is not the vendor's to give.

Why are faces, logos, and real places off-limits?

Faces and people. OpenAI's usage policy, updated October 29, 2025, prohibits using "someone's likeness, including their photorealistic image or voice, without their consent in ways that could confuse authenticity." A generated face is nobody in particular, which is the problem: on a team page it is a fake employee, and beside a real quote it looks exactly like what the FTC's fake-reviews rule, 16 CFR Part 465, bans: a review "by someone who does not exist, such as AI-generated fake reviews." I do not want to explain the difference to a regulator, and there is no problem if there are no people.

Logos and marks. A logo is the one image a client must own completely, register, and defend. It is also the case every indemnity carves out: OpenAI's excludes trademark claims outright. Add the Copyright Office's position on prompts and you have an asset with no registrable copyright and nobody behind it if it collides with an existing mark. Identity work at WBX is drawn, and the logo-system explainer covers why one file was never the deliverable anyway.

Real places and products. The Outpost, a programmatic SEO publication WBX Creative built for Power Lodge Florida, publishes riding guides, model reviews, and Florida law explainers for a powersports dealer. Its photo manifest is public domain and Creative Commons only, each license read off the Wikimedia Commons file page before the image went in, and the rule at the top of the manifest reads: "Attribution is not a license." The generated illustrations there are scenic atmosphere, captioned as illustrations, never posing as a real place. A reader who has ridden the trail will know a generated one. Google Merchant Center goes further for products: "All images created using generative AI must contain meta data indicating that the image was AI-generated" and "Don't remove embedded metadata tags." My pipeline does not preserve embedded metadata, so the compliant policy is no generated product images at all.

How I write an illustration brief

The brief is where the policy becomes enforceable. Every blog hero on this site comes from a queue file; each entry carries the post's slug, an image id, a role, a one-sentence prompt, alt text, and a credit. The generation script appends the same style block to every prompt:

Editorial illustration for a design and technology studio. Photoreal 3D-render
feel, dark charcoal environment, one teal (#008da2) accent light, cinematic
shallow depth of field, clean composition with negative space on the left third.
Strictly no text, no letters, no logos, no brand marks, no people, no faces,
no hands, no real product UI or screens with content.

The entry for this article's own hero asks for "an empty ornate picture frame on a dark gallery wall under a single teal spotlight, a faint pixel grid visible in the shadow," with alt text that describes the frame and the credit "Illustration for WBX Creative." The rules inside that brief:

1. One object, one light, one idea. A junction box, a drafting table, a rack of printing plates. Objects do not have likenesses.
2. Negative space where the title goes. The composition is a layout problem before it is a picture.
3. The negative list rides in every prompt. Letters, hands, logos, and screens with content are the four things that leak in. If one appears, the image is regenerated, not retouched.
4. Alt text describes what is there, not what it means.
5. The credit says "Illustration," printed under the image.
6. The account is WBX Creative's, never a client's. The terms and the risk sit with the studio that made the choice.

Output is 1536 by 1024 pixels, converted to WebP at quality 80. The Outpost brief is the same policy with one more line: no vehicles, because a dealer's readers know exactly what a real one looks like.

Do Content Credentials survive an upload in 2026?

Content Credentials, built on the C2PA standard, are the industry's answer to "how was this made." As of September 2026 the specification is at version 2.4, with Adobe, Amazon, Google, Meta, Microsoft, OpenAI, Sony, and TikTok on the steering committee. OpenAI says images from ChatGPT, Codex, and the API carry both C2PA metadata and a SynthID watermark; Google says all Gemini-generated images carry SynthID; Adobe says Firefly output includes Content Credentials.

Then the file leaves the tool. OpenAI's own help page says the metadata "can sometimes be removed by platforms, editing tools, or file conversions." A platform test published July 13, 2026 and updated August 13, 2026 found LinkedIn displays the credential icon, YouTube reads the manifest at upload, and TikTok auto-labels, while files pulled back from Meta carried neither EXIF nor a C2PA manifest and X strips EXIF.

My own pipeline is one of those conversions. The script re-encodes the generator's output to WebP with sharp and does not carry embedded metadata across. I will not promise a client that a credential survives a process I know discards it, so the disclosure is a credit line under the image with the word "Illustration." A caption survives every platform.

Use of such Output Content is at Customer's sole risk

Where real photography still wins

Anything a customer will compare with reality: the inside of the restaurant, the plate, the crew, the truck after the wrap goes on. The menus and wraps in the WBX portfolio were produced for businesses that exist, and the 300-plus identities on the bio page are vehicles, cards, menus, and apparel in the world, not renders. A generated "Florida beach" behind a Cape Coral business a few blocks from an actual beach is a small lie a reader can check in a minute. The photograph costs an afternoon and is true.

Mistakes I see

• Reading "you own the output" as "you own a copyright." The vendor assigns you whatever it has; for a bare prompt the Copyright Office says that is nothing.
• Assuming the indemnity applies. Reread the carve-outs in the table above before you count on one.
• Faces on the team page, faces beside testimonials, and a generated version of the real place. A reader who knows the place will know.

What it costs and what to expect

The image itself is close to free. As of September 2026, fal lists gpt-image-2 at $0.006, $0.053, and $0.211 per 1024-pixel-square image at low, medium, and high quality; Google's Nano Banana 2 is $0.067 an image; Firefly plans run $9.99 to $199.99 a month, and Midjourney $10 to $120. The cost that matters is the brief and the check: one sentence naming an object, then a look for letters, hands, logos, and screens before it ships. Minutes per image, and the part a client pays for.

The policy in one paragraph, which you are welcome to paste into your own brand guidelines:

Where to go next

This is the policy that shipped with the editorial gate behind The Outpost; the AI content checklist is its writing-side companion. If a client's marketing needs a generated face or a generated storefront, the answer is a photographer, and I will find one. If it needs illustration, the creative services page covers how that is scoped, and the contact page is where to start.