Sometimes, and the line is clearer than the debate suggests. AI images are fine for abstract, decorative, and textural work where nothing is being claimed. They are wrong for anything that implies a real person, a real place, or a real result. That single rule settles most of the arguments we have on projects.
The reason is not squeamishness about the technology. It is that a website makes promises, and images carry more of that promise than copy does. A generated photograph of a team that does not exist is a false statement, even though nobody wrote a false sentence.
Our position is that AI image tools are genuinely useful and genuinely overused. This article is about where the line sits, what the platforms now require, and what the rules have become in 2026.
Google's position is about quality and intent, not about the tool. Its Search Central documentation is clear that appropriate use of AI or automation is not against its guidelines, and that its systems aim to reward original, high-quality, people-first content regardless of how it was produced.
The limit is scale without value. Google's guidance on generative AI content states that using these tools to generate many pages without adding value for users may violate its spam policy on scaled content abuse. The problem is churning out volume, not using a tool.
That framing maps onto images the same way it maps onto text. A generated background texture that makes a section look better adds value. Forty generated images stuffed onto a page because images are supposed to help SEO does not, and never did.
Where this connects to everything else is trust. Google's quality guidance keeps returning to experience, expertise, authoritativeness, and trustworthiness, which we unpacked in our piece on what E-E-A-T is and how to prove it. Fake imagery undermines exactly the signal you are trying to build.
It depends on where you are and what the image does, and the requirements have tightened. Google does not demand universal labelling on ordinary web pages, but it does say that sharing information about how a piece of content was created can help give your readers more context.
Ecommerce is stricter. Google's guidance states that AI-generated images must contain metadata using the IPTC DigitalSourceType TrainedAlgorithmicMedia tag, and that you should preserve any metadata indicating an image was made with generative AI in the original file. If you sell products online, this is a requirement rather than a suggestion.
The IPTC standard has more than one code, which matters if you want to be accurate. TrainedAlgorithmicMedia covers images created by a model trained on sampled content. CompositeSynthetic covers composites containing synthetic elements. AlgorithmicMedia covers images made purely by an algorithm without training on sampled work.
In Europe the obligation goes further and is written into law. Article 50 of the EU AI Act requires that AI-generated synthetic image, audio, video, or text outputs be "marked in a machine-readable format and detectable as artificially generated or manipulated", with technical solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible".
That same article requires anyone deploying AI to create image, audio, or video deepfakes to disclose that the content is artificially generated or manipulated. We are not lawyers and this is not legal advice, but the direction of travel is not ambiguous.
C2PA is the standard that makes this labelling practical. The Coalition for Content Provenance and Authenticity is an open technical standard for establishing the origin and edit history of digital content, and it describes Content Credentials as functioning "like a nutrition label for digital content".
The membership is why it matters. C2PA's steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok, and Truepic. When the platforms that distribute images and the companies that generate them agree on one format, that format becomes the default.
It is already visible to the public. Google has noted that some images in Google Images may carry data showing whether they were generated with AI or captured with a camera, drawn from C2PA or from SynthID, the digital watermark embedded in some AI-generated images.
The practical consequence for a business site is simple. Assume that whether an image was generated is becoming a detectable fact rather than a secret. Plan as though anyone can check, because increasingly they can.
Abstract and decorative work is the sweet spot. Background gradients, geometric textures, subtle patterns behind a section, abstract shapes in a hero area. Nothing is being claimed, nothing is being faked, and the alternative was usually a generic stock image anyway.
Concept illustration is the second good use. If you need a visual metaphor for a section about data migration, a generated illustration in a consistent style is honest work. Nobody reads an illustration as documentary evidence.
Iteration and exploration is the third, and it is where we use these tools most. Generating twenty directions for an art style in an afternoon, picking one, and then commissioning or building the real thing properly is a genuine improvement to the design process.
The common thread is that in all three cases the image is decoration or explanation, not evidence. The moment an image starts functioning as evidence, the calculation changes completely.
Anything showing people you claim are real. Team photos, headshots, customer portraits next to testimonials, and staff at work. A generated face beside a quote turns a testimonial into a fabrication, and that is the fastest way to lose the trust the page was built to earn.
Product imagery is the second failure. If you sell a physical object, generated images of it are inaccurate by definition, and buyers notice. This is presumably why Google's requirements around AI-generated product images are the strictest part of its guidance.
Anything showing a real place fails too. Generated offices, storefronts, and locations misrepresent something checkable. Local businesses do this surprisingly often, and it is the easiest kind of claim for a customer to disprove by turning up.
Then there is the quiet failure, which is sameness. AI images trend toward a recognisable look, and a site built from them reads as generic in exactly the way the design was supposed to avoid. We wrote about choosing imagery that does the opposite in our guide on choosing photography and images for a website.
The honest answer is that this area is still moving, and we would not pretend otherwise. Ownership of generated images, the status of the material models were trained on, and what indemnities vendors offer all vary between tools and are being actively litigated and legislated.
What we can say is what to do about the uncertainty. Read the licence terms of whichever tool you use, keep a record of what was generated and when, and do not generate anything that resembles a specific identifiable person or a recognisable brand asset. Those three habits remove most of the risk that is actually within your control.
The reputational risk is more immediate than the legal one for most businesses. Being publicly caught using a fake photo of a fake team is a marketing problem that arrives long before any legal question does.
If you operate in the EU, treat the marking obligations in Article 50 as a design requirement rather than a compliance afterthought. It is much easier to preserve provenance metadata from the start than to reconstruct it later.
Not because they are AI-generated. Google's systems care about whether a page is useful, and an image does not become slow or fast based on how it was created. What does hurt is the way generated images usually arrive: enormous, uncompressed, and in the wrong format.
These tools output large files by default, and it is easy to drop a multi-megabyte image straight onto a page. That is a performance problem with a well-understood fix, and it is the same fix as for any other image, which we covered in our guide to optimising images for web performance.
The other real cost is accessibility. Generated images still need alt text that describes them accurately, and describing an abstract generated shape usefully is harder than describing a photo. Getting this wrong is common enough that we looked at it separately in our piece on whether you should let AI write your alt text.
There is one indirect SEO effect worth naming. If your imagery makes the site feel generic and untrustworthy, people bounce, and nothing about your ranking benefits from that. The mechanism is trust, not a penalty.
We ask one question about every image: is this making a claim? If the image asserts that a person exists, a place looks a certain way, a product has a feature, or a result happened, it has to be real. If it is texture, mood, or metaphor, generated is fine.
The second question is whether a real version is available. If a client has actual photography of their actual team, that always wins, even when it is less polished. Slightly imperfect real photos outperform beautiful fake ones, because visitors can tell, and the imperfection reads as evidence.
The third is consistency. A site mixing generated illustration, stock photography, and real photos in three different styles looks assembled rather than designed. If AI imagery is in the mix, it needs to belong to a defined visual system like everything else.
Where we land in practice: generated imagery for abstract and background work, real photography for anything involving people or places, and a written note in the project of what was generated so nobody has to guess in a year.
Audit the images you already have. Go through your site and mark every image as either a claim or a decoration. Anything in the claim column that is not real needs replacing, and that is usually a short list you can fix in a week.
Then set a rule and write it down, because this decision gets made repeatedly by different people over time. A one-line policy about where generated imagery is and is not acceptable saves a great deal of debate later.
If you sell products, check your image metadata now. The IPTC requirement is specific, easy to miss, and easy to satisfy if you handle it before a catalogue is built rather than after.
And if you are not sure whether the imagery on your site is helping or quietly undermining it, we are happy to give you an honest read. Send us the URL at phoenix.studio and we will tell you which images we would keep, which we would replace, and why.
Tell us where you want to go. We'll tell you how we'd get you there.