Information that remains in AI-generated images

Images made with generative AI are sometimes assumed to be safe because they were not taken with a real camera.

However, that is too simple.

AI-generated images can involve creation tools, generation dates, editing software, fragments of prompts, filenames linked to account information, service-side histories, and cloud-sharing information.

Also, even if the image itself does not show personal information, the creator or activity may be inferred from prompts, publication text, generation history, or reused source materials.

When thinking about anonymity, check AI-generated images in the same way as ordinary images.

This article organizes information that remains in AI-generated images and the points to check before publication.

Information remains even in AI-generated images

AI-generated images do not necessarily contain camera GPS information.

However, that does not mean no information remains. Metadata may be present in the image file, and generation history, prompts, accounts, IP addresses, payment information, and usage times may remain on the service side.

Information remains not only inside the image file, but also on the service used for generation.

For anonymity, think about the file and the service history separately.

Prompts become clues

In AI-generated images, prompts become important information.

Prompts may include places, descriptions of people, purposes, political claims, organization names, event names, internal circumstances, and the context of planned posts. On services or shared pages where prompts become visible externally, the prompt itself becomes material for correlation.

A prompt is text behind the image.

Even if you think you are publishing only the image, prompts may be visible in a service's sharing feature or history screen.

Before publishing, check whether the prompt is set to be exposed externally.

Generated images also have metadata

AI-generated images may also retain file metadata.

This includes creation software, editing software, generation tools, dates and times, color profiles, XMP information, and in some cases provenance information related to AI generation.

ExifTool is a representative tool that can be used to check image-file metadata. It can check a wide range of information, including EXIF, GPS, and XMP.

URL : https://exiftool.org/

However, what ExifTool can check is mainly file-side information. It cannot check histories, accounts, payment information, or IP addresses remaining on the generation service side.

Check the image content too

Even in AI-generated images, the content itself can become a clue.

For example, a background that resembles a real region, a composition that suggests a specific workplace or school, a symbol of an activity group, or an art style similar to an icon used in the past.

Even if an AI image is "fictional," the interests and activity range of the person who made it can appear in it.

For anonymity, do not judge only by whether an image is photographic or generated.

Check the sharing method

When sharing AI-generated images, look not only at the file but also at the sharing method.

What the other person can see differs between a public page on a generation service, cloud sharing, a social media post, and a chat message.

For AI-generated images, it is easier to manage risk if you save the file and then create a copy for publication.

If you paste a service's shared URL as-is, always check what the other person can see.

Consider generation-service account history

For AI-generated images, history also exists outside the file.

A generation service may retain prompts, generation dates and times, created images, editing histories, accounts, payments, and login histories. Even if you download the file and remove metadata, this does not necessarily disappear from the service side.

If you use generative AI for anonymous activity, include the generation-service account in the threat model.

If you use an image made with a real-name account on an anonymous account, correlation is created.

Be careful with reused materials

AI-generated images may use photos you have or reference images.

If that material is an image used on a past account, an image in a real-name cloud account, or an image taken at home or work, features may remain in the image after generation.

Also, when you put reference images or unpublished materials into an external image generation service, images, prompts, usage times, account information, and connection information may be handed to that service at that point. For high-risk materials, first consider not only whether the appearance changes after generation, but whether it is a service you should give that input to.

Processing something with generative AI does not necessarily erase the original information.

Treat materials used as input as part of the publication. Materials handed to an external service may remain in service-side histories or logs even if they have disappeared from the published image.

Create a publication copy

Before publishing AI-generated images, create a copy for publication.

Instead of pasting a service's shared page as-is, save only the necessary image and check the filename, metadata, and image content. If needed, change the size or format and recheck whether unnecessary information remains.

This process is inconvenient, but it is important for anonymous activity.

By not mixing the image generation service, cloud, and posting service, you can reduce entry points for correlation.

Summary

AI-generated images are not safe simply because they are not photographs.

File metadata, generation tools, generation dates and times, prompts, service histories, account information, shared URLs, and filenames affect anonymity.

Image content can also show regular activity area, interests, activity groups, and correlation with past accounts.

Before publishing, check file-side metadata, generation-service sharing settings, prompt visibility, image content, and filenames.

For AI-generated images too, it is important to check "appearance," "internal information," and "sharing route" separately, just as with ordinary images.

Related tools

Related articles