How duplicate detection actually works.

“The algorithm knows” is not an explanation. Duplicate detection is a set of comparisons, each cheap to run and each looking at a different property of the upload. Some are decisive on their own; most are combined into a score that decides whether a file gets normal distribution, reduced distribution, or a policy review.

1. The exact hash — and why it is the least interesting one

The first check is trivial: have these exact bytes been uploaded before? A cryptographic hash of the file answers it instantly, and it catches the laziest case — downloading a video and re-uploading it untouched.

It is also the easiest to defeat, which is why it is never the only check. Re-encoding the file changes its hash while changing nothing a person would notice. Any tool that markets “MD5 change” as uniqueness is solving the least important layer.

2. Perceptual fingerprints

The second check asks a harder question: does this look like something we have seen? The usual technique is a perceptual hash — reduce frames to a small grid of luminance values, compute relative differences, and store the result as a short bit string. Two videos that look alike produce hashes with a small Hamming distance, even after resizing, compression or a slight colour shift.

Production systems rarely rely on one global hash. They sample keyframes, sometimes compute embeddings per scene, and compare sequences rather than single images. The practical consequence is the same: small geometric and textural changes move the fingerprint more reliably than global colour changes.

3. Audio matching

Audio is often the stricter layer. Music recognition systems have decades of practice matching a recording through noise, compression and partial excerpts. A spectrogram fingerprint survives a video re-encode almost intact, so two uploads with the same soundtrack can be linked even when the visuals differ.

That is why a serious uniquifier does not stop at video filters. Pitch, tempo, waveform alignment, EQ and light modulation change the sample stream while remaining inaudible at small settings. Removing or replacing the audio also works, but it changes the content in a way viewers notice.

4. Metadata and container signatures

Every export carries traces: encoder strings, creation timestamps, device models, bitrate and frame-rate patterns, sometimes a writer application name. None of this proves duplication, but it correlates strongly. A hundred uploads sharing an identical encoder tag and a suspiciously regular creation time look like one pipeline, not a hundred phones.

This is also the layer most “spoofers” actually change, which is why metadata-only tools feel effective in the first week and stop being effective later.

5. Account and behaviour signals

Files are only half the story. Platforms also look at how the upload arrived: new account, no profile history, several posts within minutes, identical captions, the same link in every bio, a device fingerprint shared across accounts. These signals do not care how unique your file is.

This is why “undetectable” is a meaningless promise. A perfectly unique file uploaded by an account that behaves like a farm is still an account that behaves like a farm.

What the patents say

Large platforms have published patents describing parts of this pipeline — scene detection, fingerprint comparison, duplicate ranking — and the general shape matches what researchers describe: multiple weak signals combined into a decision. Patents are not documentation of a live system, but they are a reasonable indication that the multi-signal model is how this class of problem is approached.

What this means in practice

1. A re-encode alone is not enough. It changes the hash and the bitrate, and leaves the visual and audio fingerprints intact.
2. Change all three layers, in small amounts. Geometry and grain for the picture, pitch and timing for the sound, a fresh container and metadata for the file.
3. Make each copy different. A batch of identical “uniquified” files is a batch of duplicates with extra steps.
4. Measure the result. A perceptual distance per copy turns a guess into a number. If the number is low, the preset is weak, not the platform unfair.
5. Do not confuse file uniqueness with account health. Publishing behaviour, rights and audience response are separate problems.

Short answers

Next: what a video uniquifier is, or posting the same video to several accounts.