AI Image Generators vs Free Stock Photos: When to Use Which

The short answer

Use free stock photos when the image has to show something real or when you need a clear licence. Use an AI image generator when the picture you need doesn’t exist yet, such as a concept, a stylised illustration or a background in exact brand colours.

Most pages ranking for this question are published by companies that sell AI generators, and they conclude that you should generate everything. Few mention who owns the result, what audiences think of AI imagery, or the disclosure rules now taking effect. This guide covers those, with a decision table, a scenario table and a hybrid workflow you can use today.

One framing question settles most cases: does this image claim that something happened, or that someone exists? A team photo, a customer story, a product shot and a news image all make that claim, and they need real photography. A blog header about cash flow, a pattern behind a headline, or a fantasy scene for a campaign don’t, so a generator is fair game.

How do AI images and stock photos compare?

Stock wins on ownership clarity, realism and trust. AI wins on specificity and speed to an exact concept. Cost is close to a draw at the free tier.

Who owns the image?

With stock, you get a licence from someone who owns the copyright. With a purely prompted AI image, there may be no copyright for anyone to own, which means you can use it but often can’t stop others copying it.

AI images and copyright

The US Copyright Office’s January 2025 report on copyrightability concluded that AI outputs can be protected only where a human author has determined sufficient expressive elements, and that “the mere provision of prompts” isn’t enough (US Copyright Office). Human selection, arrangement and modification of an output can still be protected. On 2 March 2026 the US Supreme Court declined to hear Thaler v. Perlmutter, leaving the human authorship requirement in place (Holland & Knight).

The tools’ own terms point the same way. OpenAI’s Terms of Use assign you its rights in output but warn that output “may not be unique across users”. In practice, a competitor can generate something very close to your hero image, and you may have nothing to enforce. We go deeper on this in Are AI images copyrighted?

In Australia, the law on AI output hasn’t been settled by a court yet. On training data, the federal government ruled out a text and data mining exception for AI training on 26 October 2025, and set up a reference group to consider licensing and whether to clarify how copyright applies to AI-generated material (Simpsons Solicitors).

Training data risk

The other legal question is how the model was trained. In the UK, Getty Images’ case against Stability AI ended in November 2025 with the High Court dismissing Getty’s remaining secondary copyright claims and upholding only “extremely limited” trademark findings (Cleary Gottlieb). That’s good news for generator users, though cases in other countries continue. If you want the lowest-risk option, Adobe says its Firefly models are trained on licensed content such as Adobe Stock and on public domain material, not scraped from the open web (Adobe).

Stock licences

Free stock licences are short and readable, and they come with specific limits. The Pexels licence bans selling unaltered copies, showing identifiable people in a bad light, and implying that people or brands in the image endorse your product. The Unsplash licence bars compiling its images into a competing service. Our guide to free stock photo licences compares the major ones.

Rule of thumb: If you need to stop other people using the image, neither free stock nor a raw AI output gives you that. Commission it, or add enough human design work that the final piece is clearly yours.

What does each option cost?

Both start free. Free stock has no generation cap, while every free AI tier is rate-limited and the limits change often. Here’s what each provider’s own pages said when we checked in October 2026.

We couldn’t confirm a current free tier for Midjourney on its own pages, so it isn’t listed.

Two details are easy to miss. Gemini users can now turn off the visible watermark on images, but Google said in August 2026 that the invisible SynthID watermark and metadata remain (PCWorld). And Ideogram’s free images are public by default, so don’t generate client concepts there before launch.

Free stock has no equivalent cap. Unsplash, Pexels and the other libraries in our free stock directory let you download as many images as you like, and our list of free stock sites for commercial use with no watermark covers the safest picks for business use.

Watch out: Counting cost per image flatters AI. A usable marketing image often takes several generations plus cleanup, so a daily allowance of 15 goes less far than it sounds. Upscaling and background removal add more steps. Our round-up of free AI upscalers and background removers covers the free options.

Realism, people and likeness

Stock photos of people are real people who usually signed a model release. AI people have no release, and they can drift towards resembling someone real.

Generators are now good at photorealism, and they still slip on hands, small text, jewellery and how objects are held. Those errors matter most where people look closely: product pages, team pages and anything printed large.

People are the riskier case for a second reason. A generated face that happens to resemble a celebrity or a private person can create a likeness or misleading endorsement problem. The UK Advertising Standards Authority said in June 2026 that AI depictions of real people must not mislead consumers into thinking they endorse a product, and that its rules apply “regardless of how ads are created” (Lewis Silkin).

Common mistake: Using an AI-generated “customer” next to a real testimonial. Even if the quote is genuine, the picture implies a real person said it, which turns an honest review into a misleading one.

What do audiences think of AI imagery?

Audiences are more sceptical of AI images than marketers assume, especially when the images show people or products, and most want AI images labelled.

The strongest data points we found:

Sources: Getty Images newsroom, IAB, NielsenIQ, WSU, Digital Camera World, University of Melbourne.

A few details sharpen the picture:

• The gap is widening. In the IAB study, the share of Gen Z consumers feeling negative about AI ads rose from 21% in 2024 to 39%, and 30% of Gen Z respondents described brands using AI ads as inauthentic, against 13% of Millennials.
• Subject matters. Getty’s full report says sentiment skews more negative towards brands using AI to create people or products, and is more positive for buildings, structures and landscapes.
• Disclosure isn’t fatal. In the IAB study, 73% of consumers said knowing an ad was AI-created would make them more likely to buy or make no difference, and clear disclosure ranked third among the things that drew their attention to an ad.
• Vendor interest. Getty and Storyblocks both sell human-made stock, so they have a stake in these results. The IAB, NielsenIQ and university studies don’t, and they point the same way.

Do you have to disclose AI images?

Sometimes, and the rules are tightening. The EU now requires machine-readable marking of AI output, Google and Meta read AI metadata, and Australia has voluntary guidance backed by consumer law.

Sources: Baker Botts, Google Merchant Center, Meta, Department of Industry, Science and Resources, ACCC.

The EU’s provider obligations have a grace period to 2 December 2026 for generative systems already on the market. For Australian businesses, the federal guidance published on 28 November 2025 recommends three mechanisms: visible labels, watermarks and metadata, noting labels are more credible when backed by the other two. The ACCC’s rule is that claims must be “true, accurate and based on reasonable grounds”, and it counts “images and descriptions of what is offered” among the things that must be accurate. An AI image that makes a meal look bigger or a room look newer can breach that, labelled or not.

Quick check: Before publishing an AI image, ask whether a customer would feel misled if they found out. If yes, change the image. If you’re unsure, label it.

Do AI images affect SEO and AI search?

Google doesn’t penalise images for being AI-generated, but it does read AI metadata, and original images serve both readers and search better than generic ones.

Google’s guidance on generative AI content focuses on whether content helps users. It warns that generating many pages “without adding value for users” may breach its scaled content abuse policy, and recommends adding image metadata so readers know how content was made. Google Images supports the IPTC Digital Source Type field, including trainedAlgorithmicMedia for generated images and compositeWithTrainedAlgorithmicMedia for inpainted or outpainted photos (Google Search Central). Its “About this image” panel can also show C2PA data revealing whether AI tools made or edited an image (Google).

Three practical points follow:

1. Don’t strip metadata. Many image optimisers remove IPTC and C2PA fields to save bytes. For AI product images in Merchant Center, that breaks a requirement.
2. Pick representative preview images. Google’s image guidelines advise using an image that’s “relevant and representative of the page” and avoiding generic ones.
3. Originality earns attention. Nielsen Norman Group’s eyetracking research found readers study photos of real people, such as a company’s own staff, and ignore decorative stock filler (NN/g).

Worth knowing: AI answer engines pick sources they can trust, and a page with original, properly attributed images reads as more credible than one padded with generic visuals. As an AI SEO consultant in Melbourne, Saijo George treats intact image provenance as one small part of being worth citing.

Use stock when, use AI when

Stock is the default for anything that represents reality. AI is the default for anything that represents an idea.

Which to use for common jobs

The framework applied to the images small teams make every week:

A hybrid workflow: stock first, AI second, labelled

Start with a real photo, use AI only to adapt it, then keep and declare what you changed. This gets you specificity with far less trust and legal risk than a fully generated image.

1. Find a real base image. Search the libraries in our directory that ban AI uploads, such as Unsplash and Pexels, so your starting point is genuine.
2. Check the licence for edits. Most free licences allow modification. Watch for endorsement and likeness limits if people appear.
3. Make targeted AI edits. Extend the canvas to fit a banner, remove a distracting object, swap a background colour or match a brand palette. Avoid altering faces or adding people.
4. Keep human design on top. Add your own typography, layout and graphics. That improves the result and strengthens any copyright claim in the final design.
5. Tag the file. Set the IPTC Digital Source Type to compositeWithTrainedAlgorithmicMedia and keep any Content Credentials your editing tool adds.
6. Label where it matters. A short caption or alt text note such as “Photo by [photographer], extended with AI” covers both credit and disclosure.
7. Record the trail. Save the source URL, licence, tool used and date in your asset notes.

Rule of thumb: Let AI change the frame around a real moment, never the moment itself. Extending a sky is fine. Adding a crowd to an empty shop isn’t.

A 60-second checklist before you publish

Run these questions on any image, stock or generated:

• Does the image imply a real person, place, product or event? If yes, it must be real.
• Do you have a licence or tool terms that allow this use?
• Could a viewer mistake an AI person for a real customer or staff member?
• Does the image show the product or result accurately?
• Is the AI status in the metadata, and is it intact after optimisation?
• Does the brief, client contract or platform require disclosure?
• Would you be comfortable if a customer asked, “Is this real?”

If an image passes all seven, it’s ready to publish.

Questions you may have

Saijo George

SEO Consultant · Melbourne

Curator of AllTheFreeStock and a handful of other tools for marketers. Spends most of the week on technical SEO and the rest arguing about licence terms.

TwitterLinkedInWebsite