Does Google Photos Train AI on Your Photos?

What Google's own help page says

The answer is written down, and it is narrower than most articles on this question suggest. Google's privacy hub for Gemini features in Photos says, word for word: "We don't train any generative AI models outside of Google Photos with your personal data in Google Photos." The same page opens with "Your personal data in Google Photos is never used for ads." and then qualifies that sentence: "unless you choose to make data from Google Photos available to third party or other Google services that support ads."

Three qualifiers do the work in those sentences:

• "Generative". The promise covers models that produce text or images. It does not describe the recognition models Photos runs to sort your library, which are covered below.
• "Outside of Google Photos". Models that live inside Photos, such as the ones behind search and Ask Photos, are not what this sentence rules out.
• Connected services. Further down, the page says that when Photos is connected to another service, "Google may train on things like summaries, inferences, and generated media that are based on the contents of your Photos library." The imagery itself stays excluded; what is said about it does not.

That last clause is recent. The Internet Archive's copy of the page from 6 December 2025 carries an older sentence, "We don't train any generative AI models outside of Photos with your personal data, which includes other Gemini models and products.", and says nothing about summaries. The copy from 22 February 2026 has the current wording and the summaries passage.

For European readers the clause has nothing to apply to yet. Google's list of countries and territories where Photos can be connected to the Gemini apps has no European Union member on it, no Switzerland, no United Kingdom, and the United States only with exceptions.

Analysis is not training, and Photos does a lot of it

A photo nobody trains on can still be read by software every time you open the app. Google Photos groups faces, recognizes objects for search and builds Memories, and all of that runs on your pictures. Google's face grouping help page describes it plainly: "algorithms are used to create face models that numerically represent the images of faces", and "these models may be considered biometric data in some jurisdictions."

So the question most people mean, whether a machine looks at their family photos, has a different answer from the one Google's training sentence gives. Yes, constantly, for the features you use. Whether that bothers you depends on whether you value the search box more than the absence of a face model of your children.

The two Google Photos settings that change anything

The recognition itself has one switch: face grouping can be turned off, and Google says what that deletes. For object and scene recognition the help pages document no equivalent.

1. Face Groups. On a computer, open photos.google.com/settings, expand "Group similar faces" and switch Face Groups off. Google's page lists what goes with it: your face groups, the face models used to create those groups, and the face labels you created.
2. Gemini features. The privacy hub linked above is where the Gemini features in Photos are managed, including the Remember List of saved facts, each of which you can review and delete.

And the connection itself: if you never link Photos to the Gemini apps or another service, the "summaries and inferences" clause has nothing to work on.

How other photo services answer the same question

Each row below is what the service's own page said on 26 October 2026, and the link beside it is the page it came from; Viallo's row comes from its source code instead. A company's statement about its own product is a source for what it commits to and nothing more. Ente and Proton make no training statement in prose, and their rows say so rather than implying one. Meta is not in the table because its answer is different in kind; it is covered in whether Facebook uses your photos for AI training.

For keeping a library nobody but you can read, Ente or Proton Drive is the better choice because the provider never holds a key. For sharing an album with people who will not install anything, they are harder work, and that is the job Viallo was built for.

What Viallo's upload pipeline does to a photo, measured

Viallo is a private photo sharing platform: full-resolution storage, albums that open from a link with no account for the viewer, and no AI scanning or training on what you upload. The pipeline that would have to break that promise is in the repository, so the claim can be run. On 26 October 2026 a 12 MP photo, built from a NASA public-domain image with macOS sips and exiftool, went through the same steps the API runs, taken from its source, and these are the values that came back.

The perceptual hash is the closest thing to "looking at" the photo, so it is worth knowing what it is. The display copy is shrunk to 32 x 32 grey pixels and reduced to 64 bits of light-and-dark pattern; there is no model and nothing in it names a face, an object or a place. A copy resized to 2016 x 1512 at JPEG quality 60 (392,320 bytes) produced the identical 64 bits. A small, heavily recompressed copy (756 x 1008, 34,653 bytes) came out 2 bits away. A different photograph, an aurora seen from orbit, came out 34 bits away. The code treats two photos within 4 bits as a candidate and then compares the pixels before calling them copies. No label, caption or face field is filled in by the pipeline; a caption on a Viallo photo is one somebody typed.

An iPhone upload arrives as HEIC, which sharp cannot decode, so the API converts it first. The 48 MP HEIC in the same run (6,922,145 bytes) came out of that conversion step as a 12,220,491-byte JPEG, and its display copy was 2048 x 1536 at 363,766 bytes. The file grows on upload, and the HEIC itself is not stored: the converted JPEG is kept and relabelled image/jpeg.

Two things are read from the file's metadata rather than its pixels: when it was taken and where. The coordinates group photos into places on the album map, and to give each place a name, its centre point (a latitude and longitude, never the image) is sent to a geocoding service: Geoapify first, then LocationIQ, then OpenStreetMap's Nominatim. If that matters to you, remove location before uploading, on the phone or with a photo metadata remover that runs in the browser.

What this does not give you

Viallo is not end-to-end encrypted. The server makes the thumbnails, the display copy and the map, which means it can read the files, and "no AI on your photos" is a statement about what the code does, not a lock that stops it. Deleting the account removes it at once and queues erasure of every album and photo it owned. If you need a guarantee that the provider cannot read your photos even in principle, use Ente or Proton Drive.

If you decide to move the photos that matter

Most people asking this question do not want to leave Google Photos entirely; they want the albums they share with family to sit somewhere with no face models and no connected services. Keep the library where its search is useful, and move the shared albums. A link that opens a gallery without the viewer needing an account replaces the shared album without asking grandparents to sign in to anything. For the Apple side of the same comparison, read Google Photos vs iCloud, and for a ranked list of options, private alternatives to Google Photos.

How was this checked?

On 26 October 2026 the help, privacy and policy pages of Google, Apple, Ente and Proton were fetched with Python's standard library, and every cell in the comparison table kept with the exact sentence that supports it, archived copies of Google's page included. The same day the test photo was rebuilt from a NASA image with macOS sips and exiftool, a second NASA photograph was downloaded, and both went through the pipeline functions taken from apps/api/src/photos, with sharp 0.33.5, blurhash 2.0.5 and heic-convert 2.1.0; the measurement takes about five seconds. Ente and Proton publish no prose sentence about training on photos, which is what their rows say. Every value is in dataset.csv and dataset.json, and check.py re-fetches the pages and re-measures the files.

Questions people ask about Google Photos and AI