If you are approving AI images for published campaigns, buy Adobe Firefly on an enterprise contract that references the Firefly product description, and generate with Adobe's own Firefly models. Use Getty Images' generator for the few hero images where you want an uncapped cover, and Imagen on Vertex AI when the images come out of an API pipeline your engineers run. Shutterstock comes last for this buyer, because you cannot read its AI indemnity terms before you sign. Every one of these indemnities stops at a named product, plan or use, and none of them covers the trademark in the background of your shot.
"Commercially safe" on a vendor's homepage means the vendor will pay to defend certain claims about certain outputs. It does not mean the image is cleared, and the conditions differ enough that the same campaign asset can be covered by one vendor and not another.
| Option | What the indemnity covers | Plan you need | Exclusions that bite on campaign work | Price (checked Oct 18, 2026) |
|---|---|---|---|---|
| Adobe Firefly (Adobe models) | IP claims on "select Firefly outputs" | Eligible enterprise contract that references the Firefly product description | Beta and trial features; any non-Adobe model | Individual plans $9.99 to $199.99/month; enterprise indemnity is contact sales |
| Adobe Firefly (Google/OpenAI partner models) | IP claims (copyright, patent, trade secret) on unmodified output | Creative Cloud Edition 4 with Premium Stock, or a license linking to the partner-model terms | Trademark, publicity and privacy claims; unlisted partner models get nothing | Contact sales |
| Imagen on Vertex AI (Google Cloud) | Training data claims and unmodified output claims | Any paid use of a listed, generally available service | Trademark claims from use in commerce; modified output; preview models; post-notice use | Per image on Google Cloud's pricing page |
| OpenAI image models (API, Enterprise) | Copyright claims on output | API or ChatGPT Enterprise | Trademark claims; modified or combined output; disabled safety features | gpt-image-2: $30 per 1M image output tokens |
| Getty Images generative AI | Legal protection per generated image | Any purchase; uncapped only on custom enterprise | $50,000 per image cap below enterprise | 25 generations $49, 100 for $149; enterprise is contact sales |
| Shutterstock AI generator | Indemnity on images that pass human review | Enterprise customers, on request | Human review per image; terms not published | Contact sales |
How We Compared Them: One Campaign, Six Contracts
The test workload is a quarterly campaign of about 200 finished images: generated, retouched in Photoshop, placed next to a product shot with your logo, and run in paid social and print in the US and the EU. That is what a CMO actually ships, and every step after generation is a place where the fine print can drop the cover.
Three questions decide whether a claim against that campaign is covered:
- Does the cover follow the model or the app you used? This is the partner-model question, and Adobe is where it matters most.
- Does editing the image void it? Google and OpenAI both write conditions about modified output.
- What kind of claim is it? Copyright in training data, copyright in the output, trademark, or someone's likeness. Most vendors cover the first two and exclude the rest.
Prices below were checked on each vendor's live page on October 18, 2026. Where the indemnified tier is "contact sales", the table says so rather than estimating.
Why Adobe Firefly Is the Default for Campaign Work
Firefly is the best pick because its cover attaches inside the tool your designers already use, and Adobe has published where it stops. Adobe's plans page states that "eligible enterprise customers can also receive IP indemnification for select Adobe Firefly outputs", and that enterprise plans add Firefly Custom Models and indemnity through a sales conversation. The individual plans on the same page run from $9.99 a month (2,000 generative credits) to $199.99 a month (50,000 credits) and carry no indemnity. A team that bought Firefly Pro seats on a corporate card has commercial-use rights for Firefly model outputs and no Adobe defence.
Training provenance is the other half of the argument. Adobe's Firefly FAQ says its models are trained on licensed content such as Adobe Stock and public domain content where copyright has expired, and that editorial Adobe Stock content is excluded. That claim took a dent in April 2024, when Bloomberg reported that some Firefly training images were AI-generated pictures, including Midjourney output, uploaded to Adobe Stock by contributors. Adobe put that share at about 5% of the training set and said the images went through Stock's moderation. "Licensed" here means Adobe holds a licence from the uploader, which is a weaker guarantee than knowing the uploader made the picture. The indemnity exists to absorb that risk, and it is one more reason not to buy Firefly without it.
Firefly is the wrong pick for a team that will not sign an enterprise agreement, and for a team whose designers prefer partner models Adobe does not list as eligible. Both end up holding the risk themselves.
The Partner-Model Trap: The Cover Follows the Model
Firefly's app now hosts third-party models, and the Firefly IP indemnity does not travel with them. Adobe's partner-model terms say capabilities powered by non-Adobe-trained models are excluded from the Firefly IP indemnification. A separate "Creative Partner Model Supplemental Coverage" term, effective May 8, 2026, covers only the partner models Adobe lists as eligible. It applies to customers on Creative Cloud Edition 4 with Premium Stock, or whose licence expressly links to that page.
What that supplemental cover contains is narrower. Adobe's enterprise help page says it provides indemnification for certain copyright infringement claims for eligible, generally available Google and OpenAI media models. The contract term itself, the PSLT (2026v1), is a little wider: it covers claims that unmodified output infringes patent, copyright or moral rights or misappropriates a trade secret. The help page says it excludes trademark, right of publicity and privacy claims, and you lose it if you knew or should have known an output infringed, lacked rights to your input, or ignored the safety features. Firefly's own FAQ tells users that for partner model outputs, deciding whether the content fits the project is the user's job.
So the same designer, in the same Firefly window, can produce one covered image and one uncovered image by changing a dropdown. The fix is administrative: keep any team producing published work off partner models, through your admin settings where Adobe exposes them and by written policy where it does not, or limit them to Google and OpenAI models on a contract that references the supplemental terms. Beta features are excluded in both regimes, which puts anything labelled beta, including the Firefly AI Assistant while it carries that label, outside the cover.
Imagen on Vertex AI: The Broadest Written Cover, Built for Pipelines
Google Cloud publishes the most complete indemnity of the group, and it is the right pick when images come from code rather than a designer's canvas. Section 20 of Google Cloud's Service Specific Terms gives two separate indemnities. The training data indemnity covers claims that Google's use of training data to build its models infringes someone's rights, with no listed exclusions. The generated output indemnity covers claims that an unmodified output from a listed service infringes.
The listed services are the condition. Google's indemnified services list (snapshot of April 22, 2026) names the Gemini Enterprise Agent Platform API, formerly the Vertex AI API, used with generally available versions of Codey, Gemini, Imagen, PaLM and Veo. Preview model versions are off the list, and so is any service Google gives you free of charge. A team testing a preview image model on free credits has no Google cover at all. (Our Vertex AI tool page covers the platform itself.)
The output exclusions are where marketing teams get caught. Google's cover does not apply if you used an output you knew or should have known was likely infringing, disabled or worked around Google's filters, kept using it after a rightsholder's notice, or if the allegation is a trademark claim arising from your use of the output "in trade or commerce". A campaign is use in commerce by definition. The word "unmodified" also matters: an image retouched in Photoshop is no longer the generated output Google indemnified. Keep the raw generation, its prompt and its timestamp in your asset system so you can show what Google produced and what your team changed.
On price, Google bills Imagen per image on Vertex AI's pricing page. For a reference point, its Gemini API lists Nano Banana Pro (Gemini 3 Pro Image) at $0.134 per 1K or 2K image as of October 18, 2026, but check that the Gemini model and endpoint you use appear on the indemnified list before you rely on the cover.
Vertex AI is the wrong pick for a brand team without engineers. The cover is strong, but it attaches to an API, and your designers will not work in one.
OpenAI's Image Models: Covered on Paper, Loose for Campaigns
OpenAI extends its output indemnity, which it launched in 2023 as Copyright Shield, to API and ChatGPT Enterprise customers under its Service Terms. A summary of the terms updated September 29, 2026 lists the exclusions: the cover does not apply where you knew the output was infringing, disabled citation or safety features, modified or combined the output, lacked rights to the input, made a trademark claim through use in commerce, or used a third-party offering. Free and Plus ChatGPT accounts get nothing.
For a campaign, "modified or combined" is the clause to read twice. Compositing a generated background behind a product shot is combining. OpenAI's terms are also the only ones in this comparison we could not open directly on October 18 (the page blocked automated access), so have counsel read the version your contract incorporates rather than relying on any summary, including this one.
Price is metered by token: OpenAI lists gpt-image-2 at $8 per 1M image input tokens and $30 per 1M image output tokens, with Batch output at $15. OpenAI's image quality is a real reason to want it. Using it through Firefly adds Adobe's supplemental copyright cover, but that term also covers only unmodified output and excludes combining it with other material.
Buying OpenAI directly is the wrong pick for a brand team whose workflow ends in compositing, which is most of them.
Getty Images: The Uncapped Cover Costs an Enterprise Deal
Getty's generator has the cleanest training story and the most specific dollar figure. Its generative AI page says the model is trained exclusively on licensed visuals from its creative library, explicitly excludes web-scraped and public domain data, and that images you generate are never offered to anyone else to license. Generated visuals carry automatic legal protection of up to $50,000 per image. Uncapped protection appears only as a benefit of a custom enterprise plan. (The same page also says protection "starts at" $50,000, so get the figure written into your order form.) Introductory pricing is 25 generations for $49 or 100 for $149, expiring a year after purchase.
Getty's June 2026 10-Q describes "trusted copyright claim protections, model and property releases" for licensees and confirms Getty terminated its Shutterstock merger agreement on July 7, 2026, after the UK CMA required Shutterstock's editorial business to be sold. The two generators remain separate products with separate terms.
Getty is the wrong pick for a team that needs volume or style range. It generates stock photography and will not match a brand illustration system, and per-generation pricing punishes iteration.
Shutterstock: The Loser, Because You Cannot Read the Terms
Shutterstock announced AI image indemnity in 2023, extending it to enterprise customers on request, with requests fulfilled through a human review of the images. We found no current, public document stating the cap, the exclusions or the review turnaround. An October 2026 practitioner survey of 14 generators' commercial terms reaches the same place: coverage needs a qualifying plan and applies only to outputs a person has reviewed and approved, and the general licence caps are not a guarantee for unreviewed AI images.
A per-image approval step is a workable control for a stock buyer licensing a dozen images a month. For 200 campaign assets a quarter it is a bottleneck, and a legal lead cannot compare an indemnity they have not seen.
Shutterstock is the wrong pick for anyone who has to show their general counsel the indemnity language before the purchase order.
What No Indemnity Does: The Human Review Step Stays
An indemnity pays defence costs and damages after a claim is made. It does not stop a takedown notice, pull a paused campaign back into market, or repair the coverage of a brand that ran a lookalike of a famous photograph. Every vendor here also excludes the cases you are most likely to face in marketing: Google, OpenAI and Adobe's partner cover all exclude trademark claims, Adobe's partner cover excludes likeness, and every contract voids cover if you "knew or should have known".
That clause turns a review step into a condition of the cover. If your reviewer spotted a logo in the background and shipped anyway, you have given the vendor its exclusion. Keep a human check on four things before an image is published: recognisable people, logos and product packaging, famous artworks or photographs, and text rendered inside the image. Log who approved it.
Content Credentials and the 2026 Disclosure Rules
Every vendor here marks its output, and two laws now make the label your problem as well as theirs. OpenAI says images from ChatGPT and its API carry both C2PA metadata and a SynthID watermark, and warns that most social platforms strip metadata on upload. Google embeds C2PA metadata in Nano Banana Pro images generated on Vertex AI alongside SynthID.
C2PA is an open standard for a signed record, inside the file, of how an image was made. A watermark like SynthID sits in the pixels and survives a screenshot; the C2PA record does not.
In the EU, Article 50 of the AI Act has applied since August 2, 2026. Providers must mark output in machine-readable form, and deployers, meaning you, must clearly label deepfakes; the Commission's final Code of Practice, published June 10, 2026, suggests the EU icon or an equivalent label. A campaign image of a realistic person who does not exist is the case to check with counsel. In California, AB 853 moved the AI Transparency Act's operative date to August 2, 2026; its duties fall on large generative AI providers, so the practical step for you is not to strip their provenance data in your export pipeline.
The Decision: Criteria That Predict Regret
The buyers who regret this purchase usually assumed the cover was wider than it was. Four checks catch it:
- Is the indemnity in your contract, or only on the vendor's website? Adobe's cover applies only if your agreement references the product description; Getty's uncapped cover is enterprise only. A seat bought on a card has none.
- Can you lock the models your team uses? If designers can switch to an unlisted partner model, your cover depends on a dropdown.
- Does your workflow modify or composite outputs? If yes, Google's, OpenAI's and Adobe's partner indemnities all cover less than you think. Archive the raw generation.
- Who owns the trademark and likeness review? No vendor does.
What changes the answer: if your images come from an engineering pipeline, Vertex AI moves to first. If you publish under ten images a quarter and want a dollar figure in writing, Getty moves to first. If Shutterstock publishes its AI indemnity terms, re-run this comparison.
This Week: Pull the contract for whichever image tool your brand team uses and find the sentence that grants indemnity. If you cannot find it, assume you have none.
This Month: Limit Firefly partner models to the ones your contract covers, and add the four-point human review to your creative approval workflow, with a named approver.
Before Q1 Close: Set up an asset record that stores the raw generation, the prompt, the model name and the C2PA manifest next to every published AI image, so you can answer a claim or an Article 50 question with evidence.
The Bottom Line
Stock photography went through this in the 2000s: royalty-free licences came with indemnity caps, model releases and editorial-only restrictions, and agencies learned to read them. AI image indemnities are the same instrument with a new set of carve-outs. Buy Firefly with an enterprise contract, lock the model picker, and keep a person checking logos and faces before anything ships.
Continue Reading
- Training Data Provenance: Indemnity Stops at Your Fine-Tuning Data
- Sony Sued Anthropic. Re-Prompting Voids Your Indemnity.
- AI Vendor Security Review: 6 Questions That Change the Answer
- Adobe's Topaz Pledge Names Apps and Models but Never the API
- Mistral Bought Pimento for Vibe, Leaving 180 Brands No Roadmap
- Gartner: Information Integrity Now Risk Leaders' #1 Concern