The Ghibli AI Filter Went Viral - Here's What It Taught Us About Photo Privacy

The Ghibli filter phenomenon

In early 2025, the internet collectively decided to turn itself into a Studio Ghibli movie. ChatGPT's image generation got an update, and within days, millions of people were uploading selfies, family photos, pet pictures, and baby portraits to see themselves rendered in Hayao Miyazaki's distinctive animation style.

The results were genuinely charming. Your toddler as a Spirited Away character. Your dog looking like it wandered out of My Neighbor Totoro. Your wedding photo reimagined as a scene from Howl's Moving Castle. It was the kind of viral moment that felt harmless and wholesome.

But underneath the cute illustrations was something worth paying attention to. Every one of those uploads was a personal photo - often containing faces of children, family members, and friends - being sent to an AI company's servers. And almost nobody stopped to ask what happens to those photos after the Ghibli version gets saved to your camera roll.

The scale was staggering. OpenAI's CEO Sam Altman posted that ChatGPT was generating roughly a million images per hour at the trend's peak. That's millions of personal photos flowing into OpenAI's infrastructure in a single day.

What actually happens to photos you upload to AI tools

I went and read the privacy policies that most people skip. Here's what I found across the major AI image tools.

OpenAI (ChatGPT, DALL-E)

OpenAI's terms state that inputs - including uploaded images - may be used to "improve" their services. For free-tier users, this effectively means your photos can be used as training data unless you specifically opt out in settings. Paid users get slightly better terms, but the data still passes through OpenAI's servers and is retained for up to 30 days for abuse monitoring. That selfie you uploaded? It sat on OpenAI's servers for a month.

Other AI image generators

The pattern is consistent across the industry. Midjourney's terms grant them a broad license to use your inputs. Lensa AI - the app behind the earlier AI portrait trend - was found to retain uploaded photos and use them for model improvement. Many smaller AI filter apps that pop up during viral trends have even vaguer policies, or no meaningful privacy policy at all.

The common thread: when you upload a photo to an AI tool, you're almost certainly granting that company some right to use your image. The specifics vary, but the default across the industry is "we can use your uploads to make our AI better."

The metadata problem

It's not just the pixels in your photo. Most phone photos contain EXIF metadata - GPS coordinates, timestamp, device information, camera settings. When you upload a photo to an AI tool, you're often sending all of this along with it. That Ghibli-fied family photo also told the AI company where you live, when the photo was taken, and what phone you use.

The real privacy risks most people miss

The Ghibli filter trend revealed something bigger than one viral moment. It showed how normalized it's become to upload personal photos to AI services without a second thought.

Children's photos are the biggest concern

A huge portion of the Ghibli trend was parents uploading photos of their kids. These are minors who can't consent to having their faces processed by AI systems. Once those images are in a training dataset, there's no practical way to remove them. Your child didn't agree to become part of an AI model's understanding of what children's faces look like - but that's potentially what happened.

Aggregate data is powerful

One photo might seem harmless. But AI companies are collecting millions of photos with associated metadata. At scale, this creates detailed datasets about how people look, where they gather, what their homes look like, and how families are structured. The individual photo is a data point. The collection of millions of photos is a surveillance dataset.

You can't un-train a model

This is the part that makes AI uploads fundamentally different from, say, posting on social media. If you delete an Instagram post, the photo is gone (at least from public view). But if your photo was used to train an AI model, that influence is baked into the model's weights permanently. There's no "delete my photo from your AI" button that actually works. Model unlearning is an active research area, but it's nowhere close to being practical at scale.

The terms can change

Even if a company's current privacy policy seems reasonable, terms of service change. Companies get acquired. Startups run out of money and sell their data assets. The AI filter app that promised not to use your photos today might update its terms tomorrow - or get bought by a company with very different values.

How to use AI photo filters safely

I'm not saying you should never use AI filters. They're fun, and sometimes the results are genuinely great. But there are ways to participate in these trends without giving away more than you need to.

• Strip metadata first: Before uploading any photo to an AI tool, remove the EXIF data. On iPhone, you can do this in the share sheet by tapping"Options" and toggling off location. On Android, most gallery apps have an option to remove location data before sharing. This at least prevents the AI company from getting your GPS coordinates.
• Use paid tiers when possible: Paid plans on services like ChatGPT and Midjourney generally have better data retention policies. OpenAI's paid tier, for example, doesn't use your inputs for training by default. It's not perfect, but it's a meaningful improvement over free tiers.
• Avoid uploading children's faces: This is my strongest recommendation. Whatever you decide about your own photos, think twice before uploading photos of your kids to AI tools. They can't consent, and the data can't be removed from trained models. Use photos of objects, landscapes, or pets instead.
• Read the opt-out settings: Most major AI platforms have settings to opt out of training data collection. On ChatGPT, go to Settings and turn off"Improve the model for everyone." On Midjourney, check the privacy settings in your account. These options are often buried, but they exist.
• Use locally-running AI when available: Some AI image tools run entirely on your device. Apple's Image Playground, for example, processes images locally on newer iPhones and Macs. No upload means no data sharing. The results might not be as impressive as cloud-based tools, but the privacy trade-off is zero.
• Check the company behind the app: During viral trends, dozens of copycat apps appear. Many are from unknown developers with no track record and no meaningful privacy policy. Stick with established companies where you at least know who has your data, even if their terms aren't perfect.

A different approach to sharing photos

The Ghibli trend highlighted something interesting about why people share photos. The appeal wasn't really the AI transformation - it was sharing a fun version of a meaningful photo with people who care about it. The AI was just the excuse.

But there's a growing gap between how people want to share photos and what actually happens to those photos behind the scenes. You want to show grandma a cute photo of the kids. You don't want that photo training an AI model. You want to share vacation pictures with friends. You don't want those photos stored on servers indefinitely for"service improvement."

This is exactly why platforms like Viallo exist. Viallo doesn't process your photos through any AI - no facial recognition, no object detection, no style transfer, no content analysis. The only automated processing is reading GPS metadata for location-based organization and generating thumbnails for fast loading. Your photos are stored on EU servers (Cloudflare R2, Europe), and they're never accessed, analyzed, or used for anything beyond showing them to the people you choose to share with.

You can share albums through links that don't require recipients to create an account, add password protection, and see who's viewed your shared albums. It's photo sharing without the AI tax. For more on how mainstream platforms handle your data, see our deep dive into Google Photos and AI training.

To be clear, Viallo isn't end-to-end encrypted - services like Ente and Proton Drive offer stronger technical privacy guarantees. But if your main concern is sharing photos without feeding them into AI systems, Viallo gives you that while still providing a good sharing experience with gallery views, interactive maps, and no-account access for recipients.

Frequently Asked Questions