We are asked fairly often what snapWONDERS is working on next. Rather than answer it piecemeal, here is the shape of where forensic analysis is heading — and, because it matters just as much, where it is not.
Where we are today
Analysis already covers the ground most people arrive for: metadata across every section a file carries, embedded thumbnails, quality and compression characteristics, and the structural traits that identify how an image was produced.
In June we added detection for AI-generated imagery. It looks at the properties of the image itself rather than trusting anything the file claims about its own origin, which is the only approach that survives contact with a file that has been stripped, re-saved or deliberately laundered.
What we are working on
Face manipulation. Detecting that a face region has been edited is a different problem from detecting a fully synthetic image, and needs its own treatment. This is the nearest piece of work.
Video. Most forensic tooling — ours included — has historically been far stronger on stills than on moving images. Extending genuine analysis to video rather than sampling a frame and hoping is a substantial piece of work, and one we think matters increasingly.
Stylised and illustrated imagery. Cartoons, renders and heavily stylised photographs confound checks that assume a photographic origin. Treating them as their own category makes the rest of the analysis more honest, because it stops a drawing being assessed as though it were a photograph.
Provenance. C2PA Content Credentials let a file carry a signed record of how it was made and edited. We already read that record, and where it contains an AI declaration we treat it as definitive — a file that says it was generated by a machine is not something we need heuristics to second-guess.
Images from Google Gemini are a good working example: they ship with Content Credentials embedded, so the file states its own origin and we can report that plainly rather than inferring it.
Which brings us to the caveat, and there is a live demonstration of it on this very page. The illustration at the top of this article was generated by Gemini and did carry Content Credentials. Preparing it for the web — resizing it and re-encoding it as a JPEG, as any publisher would — discarded them. Nothing sinister happened; ordinary image processing simply does not carry that metadata across.
So the credential is worth everything when present and nothing when absent. An image with no Content Credentials tells you only that it has no Content Credentials. Surfacing that distinction clearly, rather than letting an absence read as a clean bill of health, is the actual work.
Format coverage. Less exciting, and the thing that most often decides whether analysis is useful to you: the list of formats we can genuinely take apart keeps growing.
What we will not do
Two commitments are worth stating plainly, because a roadmap without limits is a wish list.
We do not infer protected attributes. No race, no age, no gender, no eye colour. The techniques exist and would be straightforward to bolt on. We consider them out of scope for a forensic tool, and adding them would change what this platform is for.
We say when we cannot tell. Face-swap deepfakes are a good example: they are a genuinely hard problem, and where our confidence is low we would rather report that than manufacture a verdict. A forensic tool that is confidently wrong is worse than one that admits the limits of what it can establish, because someone will act on the answer.
On timing
Deliberately, none of the above carries a date. Roadmaps with dates tend to become either fiction or a reason to ship something half-finished. Each of these lands when it is good enough to trust, and we will write about it here when it does.
If there is a forensic capability you need that is not on this list, we would genuinely like to hear about it — that is how several items got here.
