Tell

Getting ready

Check a news photo and its headline

Tell looks for signs that the photo was edited and for a headline that does not fit the picture. It runs entirely in your browser, so the photo never leaves your device. It is a screening signal for deciding what to check by hand, not proof.

The post

Enter to check. Shift and Enter for a new line.0 / 300

Tell runs on your device. Your first check downloads the model once (214 MB); after that it stays in this browser.

Finding

No post checked yet

Add a photo and its headline, or try one of the test cases. Tell gives one of three answers:

Authentic
a real photo with a fitting headline
Edited
the photo has been altered
Mismatch
a real photo with a headline that misrepresents it

    The photo appears here with the evidence drawn over it.

    Amber marks pixels the forensic network scores as edited, on its own 0 to 1 scale. No amber means no strong traces.

    What the decision leaned on

      A gate weighs four kinds of evidence for each post. The shares add up to 100%.

      What the file says

        Find where the photo first appeared

        The surest check is the original. Copy the photo, then paste it into a reverse image search.

        Test cases

        10 photos Tell never saw in training

        Some photos were edited on purpose, some carry a misleading headline. Each one shows the true answer, so you can judge Tell.

          How Tell decides

          Four kinds of evidence

          A ResNet50 reads what the photo shows. DistilBERT reads the headline. A forensic network looks for the noise and seam traces that cutting and pasting leave in the pixels. A cross attention layer checks which parts of the photo each word of the headline matches. A gate decides how much each one counts for this post.

          Trained against caption style

          Reddit posts with edited photos usually carry short joke captions, and the first model learned that instead of looking at pixels. This version was retrained with edits hidden behind normal headlines and with headlines swapped between photos, so the pixels and the fit have to give the answer.

          What it does not check

          Whether a claim is true. Who is in the photo. AI generated images, which were not in its training data: the file check above catches only those that still carry their metadata. Video, and headlines in languages other than English.

          How well it works

          Accuracy of the Tell v2 model on four tests
          TestScore
          Held out Reddit posts (3,446)macro F1, same kind of posts as training0.81
          Stress test (3,354)macro F1, edits behind normal headlines and swapped headlines0.61
          Real Photoshop edits, caption removedmacro F1, only the pixels can tell0.40
          Test cases on this pageright answers out of 104 of 10

          The model is strongest on the kind of post it was trained on and much weaker on new photos. It catches most pasted objects in the stress test (80%) but only half of the regions copied within a photo, and it flags about 3 in 10 real photos as suspicious.

          Heavy compression, screenshots and blur weaken the edit traces. Very short headlines still push it towards Edited.

          This page runs the same model stored at half precision. On 52 check cases it gave the same answers as the original, with every probability within half a percentage point.

          Use it to decide what to check by hand. Never use it alone to label, remove or report a post.

          Privacy

          Tell runs in your browser. The photo and headline are analysed on your device and are never uploaded, logged or used for training. The model is downloaded once from Hugging Face and kept in this browser's storage; clearing the site's data removes it. GPS data in a file is noted but never shown. No cookies, no analytics.