Stock, Real or AI Images: What Each Costs You in Rights and in Trust
The choice between stock photography, commissioned work and generated images is normally settled on budget. That is the wrong axis. The three options differ mainly in what rights you end up holding and what risk you carry, and those differences do not correlate neatly with price.
They also stopped being purely commercial questions in August 2026, when the European transparency rules on synthetic content became enforceable. Most coverage of that change misstates who owes what, which is worth getting right before anyone rewrites a content policy.
Three options, three risk profiles
| Stock | Commissioned | Generated | |
|---|---|---|---|
| Cost | Low per image | High upfront | Very low |
| What you hold | A licence, with conditions | Rights, if assigned in writing | Usually no copyright at all |
| Exclusivity | None; competitors may use the same shot | Complete | None, and unenforceable |
| Main risk | Licence scope exceeded | Rights never assigned | Provenance and disclosure |
| Trust effect | Neutral to negative if recognisable | Strongly positive | Negative when detected |
The row that surprises people is exclusivity on generated images. Because copyright generally requires human authorship, an image produced by a prompt is unlikely to attract protection in its own right. You can use it; you probably cannot stop a competitor using the identical output. For a decorative blog illustration that is irrelevant. For a brand asset it is a serious defect.
Who actually has to mark AI content
This is the part most articles get wrong, and the distinction decides what you have to do.
Under the EU AI Act, the obligation to mark synthetic output in a machine-readable format sits with the provider of the generative system: the company operating the image model. The rules became enforceable on 2 August 2026, and systems already on the market before that date have until 2 December 2026 to comply. In practice this means the tool embeds the marking, not you.
The obligation that falls on you, as the business publishing the image, is narrower and different: where content is generated or manipulated so that it resembles real people, places or events and would appear authentic, it has to be disclosed as artificially generated. That is the deepfake provision, and it does not turn every illustrative graphic into a labelling exercise.
So the practical position for a typical business site is this. A generated abstract background needs no label. A generated photograph of a person who looks like a real customer, presented as if it were one, does — and probably should not exist in the first place, for reasons that have nothing to do with the AI Act, as covered in what actually reassures a visitor.
Stock licences, and the four conditions people miss
A stock licence is not ownership and its conditions are enforceable. Four of them cause almost all of the trouble.
- Editorial use only. A large share of the most striking images on stock libraries are editorial: usable in news and commentary, not in advertising or promotion. Using one on a landing page is a licence breach regardless of how it was labelled in the search results.
- Model and property releases. An image of a recognisable person without a release cannot be used commercially. The same applies to some buildings, artworks and trademarks in frame.
- Standard versus extended. Standard licences typically exclude merchandise, resale, large print runs and use in templates you distribute. If the image will end up on packaging, check.
- “Free” libraries with unclear provenance. The permissive licences on some free sites are only as good as the uploader’s right to grant them, and a takedown notice reaches you rather than them.
The mitigation is unglamorous: keep a record. Source, licence type, date, download reference, and who approved the use. A spreadsheet is fine. The cost of not having it appears years later, when nobody remembers where an image came from and the person who chose it has left.
The trust cost of a recognisable image
Beyond rights, images carry a credibility signal, and both stock and generated imagery can carry a negative one.
The recognisable stock photograph is the older problem: the same smiling handshake appears on a competitor’s site, and a visitor who notices concludes the business is generic. This matters most on pages where the claim is about people — team pages, testimonials, service pages — and least on abstract or conceptual illustration.
Generated images have their own tell, and detection is improving faster than generation is hiding. Hands, text within the image, reflections, jewellery and the specific plastic quality of skin remain the common giveaways. The commercial risk is not legal but reputational: a customer who spots an invented office or an invented staff member reasonably concludes that other things on the site may also be invented.
The line we apply is straightforward. Anything that makes a factual claim about your business — premises, staff, products, results — must be real. Anything decorative or conceptual can be whatever is cheapest and looks right.
Indemnities, and what they are worth
Several generative image vendors now offer some form of intellectual property indemnity: a commitment to defend or compensate a customer whose output attracts a claim. It is a real product feature and it is narrower than the marketing suggests.
Three conditions recur across the market and are worth checking in the specific terms rather than assuming.
- Tier. Indemnities are normally attached to business or enterprise plans, not to the consumer subscription somebody in marketing pays for personally.
- Default behaviour. Cover typically applies only where the safety filters were left in place and no third-party imagery, brand or likeness was supplied as input. Prompting with a competitor’s product photo voids it.
- Scope. The commitment is usually about copyright claims over the output, not about publicity rights, trademark, or the separate question of whether you may depict a real person.
An indemnity is therefore useful as a backstop and useless as a policy. It does not make an image yours, and it does not answer the exclusivity problem at all.
Where generated images are the right answer
Nothing above is an argument against using them. It is an argument for using them where their weaknesses do not matter, and there are several such places.
Abstract and conceptual illustration, where no factual claim is made and exclusivity is irrelevant. Placeholder and prototype work, where the alternative is delay. Variations on an image you already own, for testing layouts. Internal material that never faces a customer. And illustration for volume content where commissioning would be disproportionate and stock would be worse.
What they are not suited to is anything load-bearing: the hero image of the brand, the photographs on a team page, product imagery, or anything that will be registered, printed at scale or defended.
Where commissioned work still pays for itself
Photography is the expensive option and the only one that produces an asset rather than a licence. It earns its cost in three specific places.
Product photography, where the buyer is trying to see the actual item and a generic substitute answers nothing. This connects directly to the coverage and scale points in what resolves doubt on a product page.
People and premises, where the entire value is that these are the actual people and the actual place. A single afternoon with a photographer produces enough for a team page, an about page and a year of social posts.
Anything used repeatedly. A hero image that appears on every page has a cost per use approaching zero, which makes it the easiest place to justify the spend.
Where budget is genuinely tight, the hybrid works: commissioned photography for the pages that make claims, stock or generated for the rest, and a written rule about which is which so the decision is not remade weekly.
The practical policy, in five lines
- Record the source and licence of every image at the moment of use, not later.
- Never use stock or generated images for people, premises or results. Those must be real.
- Check editorial-only status before anything reaches a commercial page.
- Self-host everything rather than hotlinking, which is both a licence question and a performance one.
- Write alt text describing the image, not the page. This is an accessibility obligation rather than an SEO tactic, and it is part of the same body of rules discussed in what accessibility law requires.
What to do with the images you already have
Most sites carry an accumulated library with no provenance records at all. A full audit is rarely worth it; a targeted one is.
Start with the pages that matter commercially: the homepage, the top service or product pages, anything used in advertising. For each image, establish whether you can name the source. Where you cannot, and the image shows a recognisable person or is doing promotional work, replace it rather than researching it, because replacement is usually cheaper than the investigation and always cheaper than the dispute.
Then apply the recording discipline going forward. The library will remain partly undocumented for years, and that is acceptable as long as the commercially important part is clean and the new part is recorded.
The short version
Stock, commissioned and generated images differ in what you hold and what you risk, not primarily in price. Generated images usually attract no copyright, so you cannot stop anyone reusing them, which makes them unsuitable as brand assets. Under the EU AI Act the machine-readable marking duty falls on the provider of the generative system, enforceable from 2 August 2026 with existing systems given until 2 December 2026; your own duty as a publisher applies to content resembling real people, places or events. Stock licences fail on editorial-only restrictions, missing releases, standard-versus-extended scope and free libraries with unverifiable provenance. Keep anything that makes a factual claim about your business real, use cheap imagery for decoration, record source and licence at the moment of use, and audit the commercially important pages rather than the whole library.







