What counts as manipulated video
“Deepfake” usually refers to video in which a person’s face, voice or body has been generated or altered using machine learning so that they appear to say or do something they did not. It sits within a wider range of manipulation, and much misleading video involves no advanced technology at all.
It helps to separate a few types:
- Face swaps. One person’s face is mapped onto another’s.
- Lip-sync or voice edits. Mouth movement or audio is changed to match new words, sometimes using AI dubbing techniques.
- Fully generated clips. A model produces scenes that never occurred, which is the territory of generative video.
- Conventional edits. Cutting, slowing down, cropping or reordering real footage to change its meaning. These are often called “cheapfakes”.
- Mislabelled real footage. A genuine clip presented as being from a different time or place.
The last two are common and need no AI. A real video used out of context can mislead as effectively as a synthetic one.
How these videos are generally made
At a high level, many face-manipulation systems learn from many images or video frames of a person, then produce new frames that imitate their appearance and movement. Generative models can also create footage from text or image prompts. Voice cloning works on a similar principle with audio samples.
Quality varies widely. Some results show obvious flaws, while others are convincing, especially in short, low-resolution clips where compression hides artefacts. Because social platforms recompress uploads, tell-tale traces can be blurred further, a side effect of lossy compression.
Why detection is difficult
Researchers and companies build software that tries to recognise manipulated video, often looking for statistical traces left by generation methods. These tools can be useful, but there are limits worth knowing:
- Detectors trained on one generation method may perform poorly on newer ones.
- Compression, resizing and re-uploading can hide or distort the traces.
- Creators of fakes can adapt to known detectors, so it becomes a back-and-forth contest.
- Results are probabilities, not proof. A tool saying “likely synthetic” or “no manipulation found” should not be treated as a verdict.
- Genuine footage can be wrongly flagged, which can damage trust in real evidence.
For these reasons, many specialists suggest relying on several signals instead of one tool, and avoiding strong claims based on a single score.
Provenance: proving where a video came from
Instead of only trying to catch fakes, another approach records a video’s origin and edit history. This is called provenance. The idea is that a camera, editing tool or publisher attaches signed information, sometimes called content credentials, that states how a file was captured or changed and by whom.
Some industry groups have developed open technical standards for this kind of signed record. Where it is supported, a viewer or platform can check the record and see, for example, whether a file has been edited since capture or whether it was produced with generative tools.
There are limitations:
- Support is not universal, so most video today carries no such record.
- The information can be stripped when a file is re-encoded or shared in a way that discards video metadata.
- A missing record does not mean a video is fake, and a present record only shows what was declared and signed, not that the content is truthful.
- Privacy needs to be balanced against traceability, for example for journalists or activists working in risky settings.
Related ideas include invisible watermarks embedded by generation tools and visible labels. Platform requirements for these vary, and the practice of disclosing AI-generated video is discussed under synthetic media labels.
Practical checks for viewers
No checklist is foolproof, but these habits reduce the chance of being misled:
- Check the source. Who posted it first? Is there an original uploader or a reputable outlet?
- Look for other coverage. Real events are usually reported from multiple angles.
- Search for earlier versions. Reverse image search on key frames can reveal that a clip is older or from elsewhere.
- Check the context. Does the date, place and weather match the claim?
- Listen and watch closely. Odd blinking, mismatched lighting, warped edges or audio that does not fit the mouth can be clues, though their absence proves nothing.
- Pause before sharing. Emotionally charged clips spread fastest.
Newsrooms and fact-checkers use more rigorous methods, which are part of the craft of video journalism.
Responsibilities for creators and platforms
Creators who use synthetic elements can reduce harm by labelling them clearly and avoiding realistic depictions of real people without permission. Depending on the country, impersonation, defamation, privacy and publicity rights may apply, and rules differ, so this is not legal advice. Platforms handle manipulated media through their rules and review processes, as described in the discussion of content moderation, but enforcement is imperfect.
The honest summary is that both creation and detection are changing quickly. Provenance records, careful sourcing and healthy scepticism together offer better odds than any single tool, but none gives certainty.