A compressed video is not automatically AI-generated, and a clean-looking clip is not automatically authentic. Compression changes the evidence available for review. It can add blur, blockiness, smeared edges, and altered timing to an ordinary recording, while making subtle clues harder to inspect in a manipulated one. Treat a heavily shared or screen-recorded copy as a lower-quality input—not a verdict.
Quick answer
- Save the URL, uploader, visible caption, and date before the post changes.
- Find the earliest available full version and compare runtime, framing, audio, and context.
- Run a cautious AI video link check on accessible media, then read confidence, checked signals, and limitations together.
- If you lawfully have the source file, compare it with the repost instead of assuming either is definitive.
- For decisions involving identity, safety, money, reputation, or news reporting, seek qualified human review and independent confirmation.
The goal is not to beat compression. It is to recognize when a copy does not contain enough reliable information for a strong conclusion.
What compression changes—and what it does not
| Version you have | What it may contain | Best use |
|---|---|---|
| Camera original or trusted export | The fullest available frame sequence and audio track. | Compare source details and preserve provenance when permitted. |
| First public upload | Platform processing plus useful account and context clues. | Check uploader, description, date, and longer versions. |
| Downloaded or re-uploaded copy | Extra encoding, cropping, or replacement audio. | Use as a lead; find an earlier source. |
| Screen recording or short meme clip | UI overlays, resampling, cuts, and severe detail loss. | Avoid strong conclusions from the clip alone. |
Most online video is encoded again between capture and playback. Exporting, resizing, downloading, cropping, captioning, and screen recording can discard or reshape detail. The picture may show blocks in flat areas, ringing around edges, softened hair and skin texture, or motion that looks less smooth. Those changes matter because screening works from the media available; they do not prove that a face, voice, or event was generated.
Why a compressed copy can create false-looking clues
Fine features such as teeth, eyelashes, fingers, glasses frames, hair, and the edge of a face against a busy background are difficult for an encoder to preserve in motion. A real speaker can appear to have unstable teeth, softened cheeks, or flickering outlines in a low-bitrate repost. Quick edits, speed changes, auto-caption rendering, motion blur, or replacement audio can also make mouth movement look mismatched without proving a synthetic face or voice.
Review several uninterrupted moments, not a paused frame. Compare the same moment in a longer or earlier version when possible. A limited-signals result is useful information: it can mean the available media, access conditions, or quality did not support a fuller review. It is not a hidden “real” result.
Compression can also hide useful evidence
Compression does not only create noise; it can remove fine, temporally consistent information a reviewer would want to compare across frames. Small facial transitions, reflections, subject boundaries, and brief audio cues can be obscured by blur, dropped detail, or hard cuts. A short vertical repost can also omit the beginning, ending, or surrounding conversation that explains a clip.
Read a screening result as a description of available signals and limitations, rather than a determination about a person or claim. A likelihood is not proof of AI generation, deception, consent, intent, or factual truth. An inconclusive or low-signal result does not validate the media either.
Use a source-quality checklist before visual tells
- Is this an original file, first upload, repost, or screen recording?
- Can you identify the full clip, not just a cropped highlight?
- Who uploaded it, and do they link to a source or context?
- Has the title, caption, soundtrack, or language changed across versions?
- Are independent reporting, official statements, or first-hand recordings available?
This separates provenance and context from appearance. Content Credentials are a signed, tamper-evident way to record asset history; they do not decide whether an asset depicts a true event or whether every real-world claim around it is accurate. C2PA: Content Credentials explainer Their absence is not proof of AI generation because provenance may not have been added or may be missing from one copy.
A practical comparison method
Evidence-first review
Compare the same moment, then preserve uncertainty.
Record each URL, account name, post date, title, and duration. Align a distinctive gesture or spoken phrase, watch movement over several seconds, and listen for contextual edits. Screen the strongest lawful input you can access.
Open the deepfake screening guide- Start with context. Note the uploader, date, caption, duration, and whether a full version exists.
- Align the same moment. Match a distinctive gesture or phrase rather than similar-looking frames from different edits.
- Watch movement, not only stills. Ordinary compression can disturb authentic motion.
- Listen separately. A dubbed or clipped soundtrack can explain a mismatch without proving synthetic media.
- Preserve uncertainty. Note which version was checked and what evidence was unavailable.
The evidence report guide explains why result state, likelihood, confidence, checked signals, and limitations belong together. A cropped score screenshot removes the context that makes screening interpretable.
When the original file is worth seeking
Ask for the source file or earliest upload when a clip could materially affect someone—especially if it names a person, alleges wrongdoing, requests money, claims an emergency, or may be used in journalism or moderation. Only request or upload media you have the right and a legitimate reason to handle.
An original can provide a clearer frame and audio sequence, but it is not a certificate of truth. It can still be edited, paired with a false caption, or lack complete provenance. Use it as stronger input for review, then check account history, corroborating sources, and the claim itself.
What not to conclude from a compressed clip
- “The face looks blocky, so it is AI.” Blockiness can be ordinary encoding.
- “The detector did not find enough signals, so it is real.” Missing evidence is not authentication.
- “There are Content Credentials, so the story is true.” Provenance can help establish history and integrity, not every fact.
- “The upload is old, so the event is current.” A genuine recording can be miscaptioned or recontextualized.
Often the most defensible conclusion is narrow: this version is too processed, too short, or too context-poor to support a confident claim. That tells you what evidence to seek next.
Frequently asked questions
Can an AI video detector check a compressed social-media video?
It may screen accessible media and report available signals, but compression, access restrictions, re-uploading, short duration, and missing source context can limit a review. Read the result state and limitations rather than treating one signal as proof.
Does compression make a real video look fake?
It can create blur, blocking, altered edges, and timing artifacts that look unusual in a paused frame. Compare a longer or earlier version before attaching meaning to a visual oddity.
Should I upload an original file instead of using a link?
If you own the file or have permission and a legitimate reason, it can provide clearer input than a heavily processed repost. It still supports screening, not a final accusation or truth determination.
Bottom line
Start with the highest-quality version you can lawfully inspect, then preserve source and context evidence. Let compression lower the certainty of your conclusion, not raise the confidence of an accusation. Use an AI-video detector as a first-pass screening aid and escalate serious claims to qualified human verification.
Sources and update notes
Reviewed September 26, 2026. Product behavior, platforms, and technical standards can change; verify linked primary documentation before relying on a platform-specific workflow.