Launched this week
DeepFakeCheck
Evidence-first AI checks for video, audio, image & text
10 followers
Evidence-first AI checks for video, audio, image & text
10 followers
Review suspicious video, audio, images, and text in one place. DeepFakeCheck combines locally verified C2PA provenance when available, metadata clues, and AI-based analysis, then returns a clear risk rating with uncertainty notes instead of claiming proof. Video checks sample up to 10 frames across the full timeline. Free checks, no signup; uploads are not retained after the request completes.
honestly really appreciate that you surface uncertainty instead of just slapping a true/false label on things, that alone sets you apart from most detectors. one thing i'd love to see is a side by side comparison view where i can see the specific frames that triggered the risk score, basically the evidence right next to the rating, so i can judge it myself instead of taking the score on faith
@uleku9t
Thank you, Şule — that’s exactly the direction we want. I’ve attached a current image-analysis example showing how we surface the original input, uncertainty notes, and localized evidence. For video, DeepFakeCheck already samples up to 10 timestamped frames and links evidence to the relevant frame; when reliable coordinates are available, we also mark the suspicious region. Your point is that this should sit directly beside the risk rating and be easy to compare, not buried in an evidence list. We’ll make that clearer. The goal is to let people inspect the evidence, not simply trust a score.
One thing I'd love to see is a side-by-side comparison view that overlays the frame timing analysis with the audio waveform sync, so I can spot lip-sync drift at a glance. Would make the verification feel more transparent and help me trust the risk rating.
@berkayamdau0wc Thanks, Berkay — you’re right that this would make the evidence much easier to inspect. One important caveat is that lip-sync drift isn’t proof of AI on its own: dubbing, editing, variable frame rates, capture latency, and platform transcoding can also cause it. We’re going to test a synchronized audio-visual check that compares mouth motion with speech activity and highlights timestamped drift windows, while treating it as supporting evidence rather than automatically raising the verdict. The frame + waveform view would make both the signal and that uncertainty visible. Appreciate the thoughtful suggestion.
A reverse image and audio search would be a great add, even if it pulls from a small curated set of fact checking sources. It would help catch clips that are old or out of context, which feels like the most common type of misinformation I see in family group chats.
@halimez69340 Thank you, Halime — we took this seriously. Your suggestion pushed us to prioritize it immediately, and we shipped a first version during launch day. DeepFakeCheck now adds Public source traces to image checks, looking for full and partial public matches and the pages where an image appeared.
In this test, the model stayed UNCERTAIN at 74%, while source search found 10 full public matches — exactly the kind of extra context you described. I’ve attached the result. Matches are supporting clues, not proof, and reverse audio search is still on our roadmap.
One thing that would make this even more useful is a side by side comparison view showing the specific frames or audio segments that triggered the risk flag, so users can see exactly what the analysis picked up rather than just reading a score.
@ceylinuhat Thanks, Ceylin — this is already available for image and video checks. On desktop, the evidence snapshot sits beside the risk rating, and video findings are linked to timestamped frames so you can inspect what contributed to the assessment. Segment-level audio visualization isn’t available yet, and all signals remain supporting evidence rather than proof.
Behind the launch: DeepFakeCheck is built for situations where a single confidence score isn’t enough—suspicious clips, voice notes, images, or text that need inspectable evidence. We combine C2PA provenance when present, metadata clues, and multimodal analysis, then keep uncertain cases uncertain. Our current image regression set includes 5 AI-generated and 7 real images. All 12 pass our release guardrails; in the first 10-image run, all 5 AI images stayed high risk, while none of the 5 real images were falsely marked high risk (3 difficult real images remained uncertain). Video regression currently covers 4 manipulated and 6 real clips. These are small regression guardrails, not a claim of universal accuracy.
the no signup and no retention thing is honestly what got me to try it. ran a video through and liked that it showed uncertainty notes instead of a flat "real/fake" answer, kind of refreshing honestly.
@elaarnavutwvsm Really glad that stood out — privacy and honest uncertainty are two things we care deeply about. Thanks for actually testing it with a video. Was there any part of the evidence that felt unclear or missing?