Editing photos with AI: what works and when to label it

What does this look like in everyday life?

Fictional example: Niamh lives in Cork and rents out a small holiday cottage in Kerry. After a walk by the lake she wants to print a photo for her grandmother. There is a yellow digger on the shore and a couple she doesn't know at the edge of the frame. The AI eraser in her photo app removes the digger in seconds. She saves the result as a copy, so the original stays in her library.

She would like to do the same for the cottage's booking page, where a building site is visible from the window. She decides against it: guests book on the strength of the picture, and an edited view would promise something that isn't there. She uses the real photo and mentions the building work. On the private print she doesn't remove the strangers but makes their faces unrecognisable.

Which AI edits change what a photo says?

Photo apps offer two kinds of AI tools. Some improve what is already in the picture, such as light, colour or red-eye. Others create new parts of the image: they fill the gap left by a removed object, extend the edges or add something based on a text prompt.

On 20 July 2026, the European Commission adopted guidelines on the transparency rules of the AI Act. They are not legally binding, but they draw a useful line: what matters is whether an edit significantly changes the content or its meaning.

This line matters first for the companies behind the tools: their software must mark results in a machine-readable way unless it only assists with standard editing. For you, a rule of thumb: the more an edit changes what the picture says about reality, the more a label belongs with it.

How do you edit a photo without losing the original?

The steps come from Google's and Apple's help pages. Labels can differ depending on the app version and the Android phone.

If you did overwrite the original, tap ‘Edit’ and then ‘Revert’ in Google Photos. According to Google, changes to photos saved as a copy can't be undone. Keep the unedited picture in your backup too; Where are your family photos? An overview before you tidy up helps you find them all.

On iPhone the feature is also called Clean Up and requires Apple Intelligence. Apple's help page from April 2025 lists the iPhone 15 Pro, iPhone 15 Pro Max and the iPhone 16 models; in the EU, iOS 18.4 or later is needed.

Tip: You don't have to erase strangers: Apple notes that brushing over a person's face in Clean Up may blur it with a pixelated effect. Before sharing, read Sharing a group photo: ask before it goes public.

Where do AI photo tools reach their limits?

• Device: most Google Photos features need Android 8.0 or higher and at least 3 GB of RAM.
• Account, age and region: to edit just by asking in Google Photos, you must be 18 or over and use a personal Google Account, not a Workspace or school account. The feature isn't available in all regions.
• Language and region: Clean Up requires Apple Intelligence, which isn't available everywhere.
• Invented detail: whatever the AI puts in the empty spot is calculated, not photographed. Zoom in and check edges, lettering and reflections.
• No videos: the editing tools in Apple Photos can't retouch videos.

For some purposes, AI-edited pictures are the wrong choice: damage photos for an insurance claim, evidence for a complaint or photos in a sales listing should show what is really there. The EU guidelines explicitly give an AI image of a product in advertising or on packaging that misrepresents the product as an example of a deepfake.

When should you disclose a change?

Article 50 of the EU AI Act splits transparency duties between two roles. Providers of the tools must mark AI output in a machine-readable way. Anyone who uses a tool under their own authority is a deployer and, since 2 August 2026, must disclose deepfakes.

Under the AI Act, a deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic or truthful. According to the guidelines, no intention to deceive is needed.

Private or business use?

If you use AI purely for personal, non-professional purposes, the deployer duties don't apply. According to the guidelines, that includes deepfakes you share on social media in a personal capacity. Any activity that regularly brings an economic benefit counts as professional, for example an online shop, a holiday let or paid posts.

If a picture is then a deepfake, the label must be clear and noticeable without special tools, at the latest when someone first sees it. The provider's invisible marking is not enough, the guidelines say. The voluntary EU code of practice, as explained by Germany's Bundesnetzagentur, provides for an EU icon showing the letters ‘AI’ or an equivalent symbol. For evidently artistic, satirical or fictional works, a lighter duty applies.

Even without a duty: say what you changed

If a picture's meaning changes, a short note such as ‘person removed with AI’ is fair even in private. Platforms have their own rules: YouTube asks creators to disclose realistic content made or meaningfully altered with AI, but not beauty filters, colour adjustment or background blur. You set this in YouTube Studio by selecting ‘AI use’ under ‘Attributes’.

What does the image file reveal about AI editing?

Some apps record AI editing in the file. Google Photos shows this in the ‘How this was made’ section, for example ‘Edited with AI tools’ for photos edited with generative AI such as Magic Eraser, or ‘Edited with non-AI tools’ after cropping and rotating. It is based on Content Credentials following the C2PA standard.

Such a record is no substitute for a visible label. Google notes that Content Credentials aren't available for every photo, aren't shown on Google Photos web and don't tell you whether a photo is real or fake. What labels and watermarks mean for viewers is explained in AI labelling explained: notices, watermarks, provenance.

What applies in which country?

The AI Act applies directly in all EU countries, and some add their own rules. Switzerland is not an EU member, so the AI Act does not apply there automatically. Platform rules apply wherever you post.

If an edited picture humiliates or harms a real person, that person's rights are at stake; an advice service or a lawyer can assess the individual case. If you are the target of a fake yourself, see Deepfake of you online: secure evidence, report it, get it removed.