OSINT Techniques for Verifying AI-Generated Media
How to verify AI-generated images and video using metadata, visual artifacts, reverse image search, and detection tools.
A year ago, identifying AI-generated media was largely a matter of looking for obvious tells. Hands with too many fingers, text that dissolved into nonsense, faces with an uncanny waxiness. Those tells still exist in lower-effort generations, but the top tier of image and video generation tools has improved to the point where casual visual inspection is no longer a reliable filter. For OSINT practitioners, that shift matters. Verifying whether an image or video is AI-generated is no longer a niche skill. It is becoming a standard part of source evaluation, and it connects directly to the disinformation tradecraft covered in recent posts on this publication.
This post covers where the verification process stands right now: what metadata can and cannot tell you, what visual artifacts are still worth checking, why reverse image search behaves differently with AI content, and what the emerging detection tools can and cannot be trusted to do.
Metadata Analysis
Metadata is often the first thing practitioners check and the first thing that gets stripped or altered, intentionally or not. Most images shared on social media have already lost their original EXIF data through platform compression and re-encoding, which means the absence of metadata tells you very little on its own. A real photograph shared through a messaging app and then re-uploaded to a forum will often arrive with no more metadata than an AI-generated image.
What metadata can sometimes provide is a positive signal in the other direction. Some AI generation tools embed identifying information, either through the C2PA Content Credentials standard, which a growing number of platforms and generation tools support, or through tool-specific metadata fields that survive at least the first round of sharing. If an image carries C2PA credentials indicating it was generated or edited by a specific AI tool, that is strong evidence. The catch is that this metadata is easy to strip, either deliberately or as a byproduct of how an image gets re-saved, screenshotted, or re-compressed during sharing. Its presence is meaningful. Its absence is not.
For video, container-level metadata can sometimes reveal the software used to create or render a file, which is occasionally useful when that software is itself an AI generation tool rather than standard editing software. This is inconsistent enough that it should be treated as a secondary check rather than a primary method.
The practical takeaway is that metadata analysis is worth doing because it occasionally provides a clean answer, but a negative or empty result should not be read as evidence that the content is authentic. It just means the metadata did not help.
Visual Artifacts and Inconsistencies
Despite the overall improvement in generation quality, certain categories of artifact remain useful, particularly in video and in images with complex scenes.
Hands, teeth, and ears remain disproportionately difficult for image generators to render consistently, especially in scenes with unusual poses or partial occlusion. This is less reliable than it was, but in a meaningful percentage of generated images these details still show subtle errors: an extra joint, asymmetry that does not match how the body actually works, teeth that do not align with the jaw structure.
Text rendering within images has improved substantially but still produces errors under scrutiny, particularly with text on signage, labels, or background elements that were not the focus of the generation prompt. Letters that are almost but not quite correct, logos that are close approximations of real brands but subtly wrong, and text that becomes incoherent at the edges of the frame are all still common.
Reflections, shadows, and lighting consistency are some of the more durable signals because they require the generator to maintain physical consistency across the entire frame, which is harder than getting any single element right. Check whether shadows fall in directions consistent with the apparent light source, whether reflections in glass, water, or mirrors show the same scene that is visible directly, and whether lighting on a subject’s face matches the lighting on their surroundings.
For video specifically, temporal consistency is the area where artifacts are most likely to appear. Background elements that subtly shift or warp between frames, hair or fabric that moves in ways inconsistent with physics, and faces that show small inconsistencies frame to frame even when any single frame looks convincing. These artifacts are often invisible at normal playback speed and become apparent when stepping through frame by frame or slowing the video down significantly.
Edge artifacts around the boundaries of people or objects, particularly where a subject meets a complex background, can show telltale blurring, color bleeding, or unnatural smoothness that differs from the texture of the rest of the image.
None of these signals is individually conclusive. The practical approach is the same as with any other OSINT verification task: look for a cluster of small inconsistencies rather than expecting one obvious tell, and weigh the cumulative picture rather than dismissing a single anomaly or treating a single anomaly as proof.
Reverse Image Search Limitations
Reverse image search is one of the most reliable tools in the OSINT toolkit for verifying real photographs, and it behaves very differently with AI-generated content in ways that practitioners need to account for.
When a real photograph has been used before, reverse image search can often find the original context, the original date, or prior instances of the same image being used in a different setting. This is the core mechanism behind a large portion of image verification work.
AI-generated images generally will not return prior instances through reverse image search, for the simple reason that the image did not exist before it was generated. A practitioner running a reverse image search and getting no results might interpret that as the image being original and therefore more likely authentic. With AI-generated content, the opposite can be true: a lack of any prior instances is consistent with both an authentic original photograph and a newly generated image, and reverse image search alone cannot distinguish between those two explanations.
There is a secondary use case that remains valuable. If a reverse image search does return a result showing the same image, or a very similar one, associated with an AI image generation platform, a prompt-sharing community, or a stock AI-image marketplace, that is strong evidence of AI origin. The search is not finding the “real” version of the image. It is finding evidence of the image’s existence within AI-generation ecosystems.
The broader point is that reverse image search needs to be understood as a tool that answers a different question for AI-generated content than it does for real photographs, and practitioners need to adjust their interpretation of both positive and negative results accordingly.
Emerging AI Detection Tools
A growing category of tools is being built specifically to detect AI-generated content, and the practical reality is that these tools are improving quickly but remain inconsistent, particularly across different generation models.
Detection tools generally work by analyzing statistical patterns in the image data that are characteristic of generation models, looking for the kinds of artifacts that generation processes tend to leave behind even when they are invisible to the human eye. Some tools focus on specific generation models and perform well against those models while performing poorly against others. Detection accuracy also tends to degrade significantly once an image has been compressed, resized, or re-encoded, which describes the vast majority of images encountered in real investigations.
The practical guidance for using these tools is to treat their output as one data point among several rather than a verdict. A detection tool returning a high confidence score that an image is AI-generated is meaningful and worth weighing heavily. A detection tool returning a low confidence score is much weaker evidence of authenticity, because it may simply reflect a generation model the tool was not trained to detect, or an image that has been processed in ways that obscured the relevant signals.
For video, detection tools generally analyze either individual frames using image-based methods or look for temporal artifacts across frames. The same caveats apply, with the added complexity that video compression is typically more aggressive than image compression and can further degrade the signals these tools rely on.
The detection tooling landscape is moving quickly, and tools that perform well against current generation models may perform poorly against models released six months later. This is not a space where a practitioner can learn a tool once and consider the skill set complete. Staying current with what tools exist and what their actual performance looks like against current-generation AI media is part of the ongoing work.
Why This Matters for Disinformation Tradecraft
The disinformation investigations covered in the previous post on this publication rely on establishing the authenticity and origin of content as a foundational step. A narrative built around a fabricated image or video is a different investigative problem than a narrative built around a real image used in a misleading context, and getting that distinction right at the start shapes everything that follows.
As AI-generated media becomes more difficult to distinguish from real content through casual inspection, the verification step becomes both more important and more demanding. The practitioners who treat this as a checklist item, running one tool and accepting whatever it returns, are going to miss things. The practitioners who understand what each method can and cannot tell them, and who weigh multiple imperfect signals against each other, are the ones whose assessments will hold up.
This is not a problem that gets solved once. The generation tools keep improving, the detection tools keep adapting, and the gap between them keeps shifting. Building the habit of methodical, multi-method verification now is what keeps a practitioner’s tradecraft relevant as that gap continues to move.
Some links in this post may be affiliate links. If you purchase through them, I may receive a small commission at no additional cost to you. This helps keep the blog running and the content free.