Your AI Photo App Is Uploading Your Photos. Mine Doesn't.

Your AI Photo App Is Uploading Your Photos. Mine Doesn't.

Point Google Lens at something, and that photo goes to Google's servers. Edit a selfie in Facetune, and your facial geometry — jaw shape, eye placement, estimated age — gets extracted and stored. Upload a photo to ChatGPT, and OpenAI uses it to train their models by default. Only 22% of users know that setting exists, let alone that they can turn it off.

None of this is a fringe edge case or a bug. It's the default behavior of most AI-powered photo apps. The moment you open your photo in one of them, it's usually no longer just yours.

I build iOS apps, and I made a different bet with PhotoDrift and SnapDraft: every analysis, every recommendation, every piece of AI inference happens entirely on your iPhone. Nothing leaves your device. There's no account. There's no server to breach, because there's no server.

This post is about why that matters more than most people realize — and exactly how it works under the hood.

What most apps actually do

It's worth being specific, because "the cloud" sounds abstract until you look at what individual companies do with photos.

Google Photos does not train generative AI outside of Google Photos — but when AI features connect to other Google services, the broader Gemini privacy policy applies, which states that user content may be used to improve Google's AI systems. Google Photos photos are not end-to-end encrypted: Google can access them. In 2022, Google settled a class action over its face grouping technology for $100 million after collecting biometric data without proper consent.

Google Lens sends every analyzed image to Google's servers — no on-device option. The App Store privacy label lists collection of user content, usage data, location, contacts, search history, identifiers, and financial information. Results are saved to Web & App Activity by default.

Gemini (free tier) uses uploaded images for model training by default. You can opt out via settings — but conversations reviewed by human trainers are retained for up to 3 years even after you delete them. New chats are saved for 72 hours even with history disabled.

Amazon Photos uploads photos to AWS and uses Amazon Rekognition — a commercial facial recognition service — to process them. Rekognition is also marketed to businesses and government agencies for surveillance. A class action lawsuit (currently active) alleges Amazon collected biometric facial geometry from uploaded photos including photos of children, without consent.

ChatGPT uses uploaded images to improve OpenAI's models by default. A 2024 EU audit found only 22% of ChatGPT users were aware the opt-out existed.

Facetune settled a class action in November 2024 for approximately $4.5 million under Illinois' Biometric Information Privacy Act (BIPA) for collecting facial geometry data — jawline, eye placement, nose shape, estimated age — without written consent.

And then there's "Ever" (Everalbum), the most instructive case of all. The app secretly enabled facial recognition on all uploaded photos — including after users deleted their accounts. The FTC intervened in 2021 and ordered the company to delete every facial recognition algorithm it had built from user data. That set a precedent worth remembering: training AI on user photos without consent isn't just bad manners. It's a deceptive trade practice, and regulators will make you destroy what you built with it.

Why your camera roll is worth protecting

The average iPhone user has 2,400–3,500 photos on their device (Photutorial, 2024). Mine isn't a curated portfolio, and yours probably isn't either. It's faces, GPS coordinates embedded in metadata, screenshots of bank statements, photos of prescriptions, and moments that were never meant for anyone else. A camera roll is closer to a diary than a gallery.

Now multiply that by everyone. About 1.94 trillion photos were taken globally in 2024, with smartphones accounting for 92–94% of them. At that volume, even a small percentage of improperly handled photos is an enormous amount of private data.

Regulators have noticed. Under GDPR, photos containing faces become Special Category data when biometric processing is applied — the same legal tier as medical records, religious beliefs, and sexual orientation. That requires explicit consent, not a buried terms-of-service checkbox. The Netherlands Data Protection Authority fined Clearview AI €30.5 million in 2024 for processing facial biometric data without consent; France's CNIL fined the same company €20 million two years earlier.

Users have noticed too, even if they can't always name the culprit. 82% of US consumers say AI data loss-of-control is a serious personal threat (Relyance AI, December 2025), and 75% say they won't purchase from companies they don't trust with their data (Cisco, 2024). The anxiety is real and measurable. What's missing is a clear picture of which apps are the problem — and proof that there's another way to build them.

There is. Apple shipped it.

What Apple built so I don't have to

On-device photo analysis isn't magic, and I didn't invent it. It's Apple's Vision framework — 31 machine learning request types that run entirely on the Neural Engine, GPU, or CPU of your iPhone, with no network call required.

Here's what Vision can do locally, right now, in iOS 18:

Two of these deserve a closer look.

VNCalculateImageAestheticsScoresRequest became a public API in iOS 18 at WWDC 2024 — it's the same API Apple's own Photos app uses internally to pick Memory candidates. Before it went public, developers had to approximate it with custom Core ML models. Now it's built into the OS.

VNGenerateImageFeaturePrintRequest is the one PhotoDrift leans on for duplicate detection. It runs an image through a neural network and produces a high-dimensional vector; similar photos produce vectors with low distance between them. At a threshold of 0.35, the API reliably separates true duplicates — including lighting variations, crops, and minor rotations — from different shots of the same subject.

And it's fast: 20–40ms per image on an iPhone with an A18 Pro chip. A library of 2,500 photos analyzes in seconds. One developer processing 35,000 photos with this approach deleted 12,000 of them and freed 150 GB of storage — all without a single byte leaving the device.

No API cost. No rate limits. No data shared. Zero.

How PhotoDrift puts it together

PhotoDrift is a photo cleaner. You swipe left to delete, right to keep. The Tinder mechanic is deliberate — it makes each decision feel lightweight. You're not committing to a big organizational project; you're making one small call at a time.

Under the hood:

1. On first launch, PhotoDrift requests Photos access and analyzes your library using VNGenerateImageFeaturePrintRequest and VNDetectFaceCaptureQualityRequest.
2. Similar images are grouped. Blurry shots are flagged. Screenshots and utility images are separated from memorable photos.
3. You swipe through the flagged groups. Deleted photos go to Recently Deleted — Apple's 30-day recovery bin — not into the void. You have a month to change your mind.
4. Nothing is transmitted anywhere. PhotoDrift has no backend, no account creation, no email field, no analytics SDK phoning home.

That 30-day undo window matters more than it sounds. In my experience, the reason people don't clean their camera roll isn't laziness — it's the fear of accidentally deleting something irreplaceable. The Recently Deleted buffer removes that risk, and the on-device architecture removes the other one: the photos you're sorting through never become someone else's training data.

The same principle carries over to SnapDraft, my AI drafting tool, where the stakes are words instead of pixels. It runs its language model locally using Apple's Foundation Models framework (announced WWDC 2026) — so when you're drafting an email about a medical situation or a reply to a difficult conversation, the prompt never reaches a server, because there isn't one. Sensitive text deserves the same treatment as sensitive photos.

The honest part: this is harder to build

I won't pretend on-device is the easy path. It isn't.

Cloud is easy. You write a REST call, you get results. You benefit from constantly improving server-side models. You can process arbitrarily large data without caring about device constraints. And you charge a subscription to cover the server bill.

On-device means:

• Working within Apple's framework APIs instead of choosing from the broader landscape of cloud ML services
• Optimizing for the most constrained devices in your target range, not for infinite compute
• Writing more careful code to avoid blocking the main thread during analysis
• No revenue from recurring usage — so you have to charge upfront

That last point looks like a constraint, but it's actually the whole business model inverted. On-device apps have no running costs that grow with usage, which is why they can charge $9.99 once instead of $X/month forever. The architecture and the pricing are the same promise: your usage doesn't cost me anything, so I have no reason to want anything from you except the purchase.

Where this is heading

Apple has been telegraphing this direction for years. "What happens on your iPhone, stays on your iPhone" has run as a campaign since 2019. Apple Intelligence, announced at WWDC 2025, uses on-device AI as its primary privacy differentiator against Google and Microsoft. The Foundation Models framework released at WWDC 2026 hands on-device generative AI to every iOS developer.

So the picture is complete. The infrastructure exists. The legal pressure is growing — ask Everalbum, Facetune, or Clearview. And the consumer verdict is already in: 84% of users would abandon or restrict companies that are opaque about AI data use (Relyance AI, 2025).

The apps that win the next five years won't be the ones with the most powerful cloud infrastructure. They'll be the ones users actually trust — and trust isn't a privacy policy paragraph. It's an architecture. Either your photos leave the device or they don't.

Mine don't. That's the whole pitch.

Vadym Suprun is an iOS developer. PhotoDrift is available on the App Store. SnapDraft is coming soon.