What Data Does Raven Actually See When You Upload a Photo?

"We take your privacy seriously" is a sentence that measures nothing. It appears on the policy page of services that delete everything within seconds and services that keep every upload for a decade, which makes it useless as a way of telling them apart. The only version of that promise worth reading is a specific one, so here is the specific one: what technically happens to an image file from the moment you pick it on your device to the moment Raven's server has finished with it.

Nothing below is unusual engineering. That is rather the point. The privacy properties come from ordinary decisions made in a particular order, and any of them could be checked by watching the traffic and the storage layer rather than by trusting a slogan.

When you choose a photo, the browser reads it as a byte stream — the same underlying data as the file sitting in your camera roll, unchanged. That stream travels over HTTPS, the same encrypted transport used for online banking, so the file is unreadable to anyone sitting between your device and the server. This step is entirely standard. Every web application that accepts a file does it this way, and there is nothing Raven-specific about it.

One consequence worth naming: whatever is inside the file goes with it. If the photo carries Exif metadata, that block crosses the wire too. It is not read by the analysis, but it is not magically removed either, and the same is true of every upload form on the internet.

The browser attaches a label to an upload saying what kind of file it thinks it is sending. That label is a claim, not evidence, and a naive server that believes it can be persuaded to process something it should not. So the bytes are inspected directly instead. A JPEG opens with the sequence FF D8 FF; a PNG opens with a fixed signature of 8 bytes; the full catalogue of these is documented in the public list of file signatures. If the leading bytes and the claim disagree, the request stops there.

This is a security check rather than a privacy one, but the two are related. A service that will not carefully validate what it accepts is unlikely to be careful about what it retains.

This is the part that carries all the weight, so it is worth being blunt about what is absent. There is no upload directory. There is no temporary file with a cleanup job that might one day fail to run. There is no row keyed to your account with a pointer to an object in cloud storage. The bytes are converted into a text-safe encoding in the same memory they arrived in, sent to the model, and dropped when the handler returns.

That absence is what makes the promise durable. A retention policy is a decision that can be revised in a settings panel; a system with nowhere to put an image has nothing to revise. It is the same reasoning laid out across the wider AI and photo privacy picture: the safest data is the data that was never written down.

Photos routinely carry an Exif block, a format standardised in 1995: camera make and model, exposure settings, a timestamp and, when location services are enabled, coordinates precise to a few metres. Raven does not consult it. That is not merely a policy choice — it is what makes the product interesting. Reading a stored coordinate would be a lookup, not a guess, and there would be nothing to explain afterwards.

Stripping metadata before sharing photos remains a reasonable habit generally, particularly for services you know less about. The practical version of that habit, for trips and public posts, is covered in traveling without oversharing your location.

Once metadata is off the table, everything the model has is what you can also see: rooflines and building materials, the script on a shop sign, the species of tree at the edge of the frame, kerb paint, plate proportions, the length and hardness of shadows. It reads them together and returns a probable region with its reasoning attached.

• —Signage and script, which usually narrow a photo to a language region before anything else is considered.
• —Built environment, meaning roof pitch, window proportion, balcony style, wiring, kerbs and paving — the ordinary construction habits of a place.
• —Vegetation and terrain, which act as a proxy for climate band rather than for a country.
• —Traffic conventions, including which side of the road is used, sign geometry and number-plate shape.
• —Light, where shadow length is a genuine measurement of sun angle rather than an impression.

That list also explains what an ordinary holiday snap gives away without any technical trickery at all, which is the subject of what your vacation photos might reveal. The same evidence a model reads is available to any patient human being with an internet connection.

It would be dishonest to claim that nothing persists. Accounts persist, because that is what an account is. The free-guess flag persists, because otherwise a free trial could not be a trial. Operational logs persist, because a service nobody can debug is a service that breaks quietly. Each of those is boring, and being clear about the boring parts is what makes the interesting claim credible.

The same philosophy runs through Geospy AI, Raven's free iOS sibling: analyse a photo in the moment, answer the one question, keep nothing. Neither product needs to remember an image to say where it looks like it was taken, and once that has been said there is nothing left worth holding. The design boundary this creates — curiosity rather than monitoring — is argued out in the ethics of AI photo analysis.

If you would rather see the reasoning as a game than as a policy question, the comparison in GeoGuessr vs AI photo geolocation tools covers the same visual deduction from the opposite end: humans doing it for sport rather than a model doing it on request.

So the short answer to the question in the title: Raven sees an image for a few seconds, reads what is visible in it, and then forgets the whole thing. Everything above is simply the long form of that sentence, written so it can be checked rather than believed.