Two different things people call the same thing
A fully generated image starts from nothing. There is no photograph underneath it, no sensor reading, no moment that happened. Every pixel was produced by a model, and the person in it has no identity to defame because there is no person.
A deepfake starts from something real. There is a photograph or a video underneath, a genuine capture of a genuine moment, and a model has replaced a specific region of it, usually a face, with a rendering of somebody else.
The distinction matters legally, editorially and technically. A generated stock photo of a nonexistent doctor is a licensing question. A photograph of a real doctor with their face pasted onto a body they were never attached to is a defamation question, and possibly a criminal one.
Why the deepfake is harder to catch
A detector reads the statistical texture of an image and reports how much of it looks like something a model produced. When the whole frame is synthetic, that signal is everywhere, and the score is unambiguous.
A face swap inverts the problem. Ninety per cent of the picture is a real photograph, with real sensor noise and real compression history, and the whole-frame average reflects that. A convincing swap can produce a score in the thirties on an image that is genuinely deceptive.
That is not a failure of the model so much as a failure of the question. Asking whether the image is generated has the wrong shape when the honest answer is that most of it is not, and one important part of it is.
The averaging that produced 33 is exactly what hides the one tile that matters.
The region map is the answer to that shape problem. Scoring overlapping tiles separately means a localised replacement is measured against its own neighbourhood rather than diluted across a frame it does not belong to.
The vocabulary, in the order it appeared
| Term | What it means | Is the subject real? |
|---|---|---|
| Deepfake | A real person's likeness moved onto other footage | Yes |
| Face swap | The common form of a deepfake, on a still or video | Yes |
| AI-generated | Produced from a prompt with no source photograph | No |
| AI-edited | A real photo with something added or removed | Usually |
| Synthetic media | Umbrella term covering all of the above | Depends |
| Cheapfake | Real media misled by caption, crop or old date | Yes |
The last row is worth dwelling on, because it is the most common form of visual misinformation and the one no detector addresses. A genuine, unedited photograph from a different country three years ago, captioned as today's news, will score as the real photograph it is.
What to actually do, by case
- You suspect the whole picture is invented. Run a whole-frame check. A high score with a flat map is about as clear as this gets.
- You suspect a person was inserted or swapped. Read the region map first. The number will understate it.
- You suspect a document was altered. Same as above, and expect exactly one hot tile over the changed field.
- You suspect the caption is lying. A detector cannot help. Reverse image search, then check the date and place.
- It is video. Different tooling entirely. Frame-by-frame still analysis misses the temporal artefacts that give real-time swaps away.
Why this distinction protects people
Treating every synthetic image as equally serious produces two failures at once. It wastes attention on harmless things, such as somebody using a generated portrait because they could not afford a photographer, and it under-weights the case that actually damages a person.
The harm is not in the pixels being artificial. It is in a real, named individual being shown doing something they did not do. That is the thing worth escalating, and it is the case where a middling score is least trustworthy as a summary.
Framing an internal policy around identity rather than around technology makes it age better too. The methods will keep changing; the question of whether a specific person is being misrepresented will not.
How a face swap is actually made
Understanding the production helps explain the detection. A swap starts with a target photograph and a set of images of the person whose face will be used. A model learns to render that face under the lighting, angle and expression already present in the target.
The rendered face is then composited back into the original frame, and the seam is the hard part. The two halves have different noise, different compression history and often slightly different colour response, so the blending step smooths the boundary until an eye stops noticing the join.
That smoothing is what a region map reads. Blending removes the high-frequency texture that a sensor produces, and the resulting patch is statistically unlike everything around it even when the join is invisible. The seam is invisible and measurable at the same time.
It also explains why a swap survives a whole-frame check so easily. The manipulated area is small, deliberately blended into its surroundings, and surrounded by an authentic photograph that dominates any average taken across the frame.
The mistake that keeps recurring
People run a suspected deepfake through a detector, see a score in the thirties, and conclude the image is genuine. That is the wrong inference from the right number: the score is correct and the question it answers was never the one being asked.
The habit that fixes it is small. On any image where the concern is about a specific person, open the region map before reading the headline figure, and treat the highest tile rather than the average as the finding worth acting on.