AI
What Image Enhancement Can and Can’t Recover
Enhancement makes existing detail easier to see. AI upscaling invents detail that was never there. Those are completely different things and only one of them belongs anywhere near a case.
The distinction isn’t academic. It decides whether what you produce is evidence or a picture of something a model guessed.
| Legitimate | Invention |
|---|---|
| Brightness, contrast, levels | AI upscaling to add pixels |
| Sharpening within existing data | “Face restoration” on a blurred face |
| Noise reduction | Generative fill of obscured areas |
| Colour and white balance | Reading a plate that’s four pixels wide |
| Rotation, perspective correction | Anything described as “imagining” |
Why upscaling is not recovery
If a license plate occupies six pixels, the information required to read it does not exist in the file. An AI upscaler doesn’t recover it. It generates a plausible plate — characters that look like what a plate looks like — and it will produce a different plausible plate if you run it again.
That isn’t evidence of anything. Presented as evidence it’s worse than useless: it’s a fabrication with your name on it.
Work from the original file, never a copy of a copy
Every re-save of a JPEG loses information. A screenshot of a photograph, sent through a messaging app, saved again, has almost nothing left to enhance.
Ask for the original file. Check what you have with the EXIF viewer — if it carries camera metadata it is probably close to original; if it’s stripped, it has been through a platform.
Hash it before you touch it
Record the SHA-256 of the file as received, before any processing. Then work on a copy.
If the image ends up mattering, you need to be able to show the original was unmodified and that your working copy derives from it.
Adjust, don’t generate
Brightness, contrast, levels, curves, sharpening, noise reduction, rotation. These operate on data that’s present.
Anything offering to “restore”, “enhance faces”, “upscale” or “fill” is generating. Don’t use it on evidential material.
Record every step
Write down what you did, in order, with the software and version. “Levels adjusted, black point 12, white point 240; unsharp mask radius 1.2”.
The test is reproducibility: someone else with the original should be able to follow your steps and get your result. If they can’t, it isn’t enhancement.
Present the original alongside the worked version
Never show only the processed image. Show both, state what was done, and let the reader judge.
The image comparison tool puts them side by side and will show how much actually changed.
Where AI genuinely helps with images
Not on the evidential file — on the search around it.
- Describing what’s in a photograph so you know what to search for
- Reading text that’s legible but awkward — a sign at an angle, handwriting
- Identifying a vehicle model, a uniform, a landmark, a plant that indicates a region
- Sorting a large set of images by content so you can find the three that matter
All of that generates leads. None of it goes in the report as a finding without independent confirmation.
If a case may end up in court, ask before you process. Some jurisdictions and some clients have specific requirements about handling of digital images, and doing the work first and asking afterwards can taint an exhibit that was fine when it arrived.
Go deeper
AI Essentials for Investigators
Hands-on with AI tools that improve and unblur images for clearer evidence and stronger leads, plus using AI for data analysis and social media intelligence. Only basic internet skills needed. $49.99.
See the course