AI tools can make blurred or low-quality media look clearer, but they can also invent convincing details that were never captured. This article explains how forensic image and video enhancement, deepfakes, and computational photography are blurring the line between authenticity, integrity, and evidentiary reliability.
Real versus fake has always been an oversimplification. It’s never been black and white. But in the AI era, it’s made of many subtle shades of grey.
Since the beginning of our Amped Authenticate training, we refrained from calling a media file “real” or “fake”. Traditionally, the two main kinds of analyses done while doing this activity are “integrity verification” and “authentication”.
Key Takeaways
- Integrity and authenticity are not the same thing. A file can be original yet still fail to represent reality accurately.
- AI enhancement can introduce synthetic details. It may make the media look clearer without truly recovering information that was never captured.
- The boundary between AI enhancement and deepfakes is often blurred. In many cases, the underlying technology is similar, even when the intent is different.
- AI-enhanced evidence should be treated with extreme caution in forensic and legal contexts. If the process is not transparent, repeatable, and scientifically defensible, the result may be misleading.
- Camera-original files are no longer automatically trustworthy. Computational photography and built-in AI processing can alter media before investigators ever receive it.
- The core forensic risk is not only fake media, but fabricated details presented as recovered truth.
- AI risks extend beyond enhancement tools. Computational photography, AI-based editing, and emerging AI compression methods all complicate forensic interpretation.
What Is the Difference Between Integrity Verification and Authentication?
The SWGDE glossary defines “Integrity Verification” as “The process of confirming that the data presented is complete and unaltered since time of acquisition”. “Authentication” is defined as “The process of substantiating that the data is an accurate representation of what it purports to be”.
We can already see that while integrity verification can be performed with purely technical means, authentication is very much content and context-dependent. Interestingly, the two concepts are related but different.
We can do a simple analogy with a physical signature:
- if I sign a paper, it is original (we will use this word as a synonym of a file where integrity is verified) and authentic;
- if I make a photocopy of it, this is not original, but still authentic.
On the other hand, if someone else tries to copy my signature, that would be original but not authentic, and its photocopy won’t be either original or authentic.
In the field of image and video forensics, integrity often implies authenticity: if the file is the original (and thus passes integrity verification), usually it is also authentic. In fact, any change to its content would break its integrity. However, there are a few situations where this is not necessarily true. Let’s see a few practical examples, all seen in real cases:
- Recapture: if we manipulate an image and then take a picture of the result, that would appear as an original file. In fact, we would be analyzing the integrity of the recaptured file, not of the original one.
- Staging: if we stage a scene or event that never happened and take pictures of it, the files would be original, but the content wouldn’t be authentic.
- Misattribution: if we reuse a picture from an event out of its original context as evidence of another situation, that wouldn’t be authentic either. We see this happening a lot on social media.
There’s some debate about whether the analysis of staging or misattribution falls within the responsibilities of media forensics experts. This is because it is often more about context and circumstantial evidence than the analysis of technical features. Even so, this overview shows pretty well how authentication is a complex matter. And notice how we didn’t even write the word “deepfake” yet. This brings us to another level.
What Is a Deepfake?
Definitions of deepfake vary by legal framework, purpose, and jurisdiction, but they commonly refer to “AI-generated or altered media designed to mislead“.
In the United States, a single, official, nationwide legal definition is absent. In contrast, the European AI Act defines it as follows: “‘deep fake’ means AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful“. These definitions underscore that the technology used to create a deepfake is not the sole determinant: the creator’s intent or the effect on the viewer holds greater significance.
Even the transparency obligations in the Art. 50 of the AI Act acknowledge the difficulty of a binary classification. In fact it says “2. Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated”. However, it adds some exceptions among which is “do not substantially alter the input data provided by the deployer or the semantics thereof”. We can discuss hours on what this means. I am copying here an excerpt from my previous article on the AI Act to show how this is tricky.
The photo on the left is an original hi-resolution image of Tommy Lee Jones, the one on the right is a low-resolution version enlarged with AI. While the differences may be subtle, doing a facial comparison on the image processed with AI will likely cause very dangerous errors. The shape of the eyes of Tommy Lee Jones is very peculiar, and changes drastically in the AI-processed image.
Does AI substantially alter data or its semantics here? It depends. The two pictures are very similar. However, if I need the enhanced picture to be used for an identification, it will be very misleading.
The problem in replying to this question is very well exemplified by some of the challenges that many of our expert users are facing now.
“Do You Have Deepfake Casework?”
Until a few years ago, whenever I asked our users how many actual deepfake cases they had, they used to say they were almost none, or very little, but they had to be prepared for the future. And the future is now.
In fact, it slowly changed to a small but growing number of big cases related to politics and propaganda.
Nowadays, we have some of our users being requested every single day about the authenticity of images created with nudifying apps, or plain LLMs, often involving minors. AI-generated CSAM (child sexual abuse material) is a problem we’ve been talking a lot about, but not enough, in the past few years. It was a bomb ready to explode that was somewhat going under the radar. Then, with the Grok issues, early this year, the bomb exploded in the face of everybody. However, this is another huge topic that should require at least another post by itself.
How to Detect AI-generated or Manipulated Images?
Deepfake detection started with an “easy” plan. An image could have been either “real”, generated by a camera, or “synthetic”, generated by an AI model.
But many times, nowadays, the image is not generated entirely by AI: most of the picture is real, but AI may be used to add, remove, or change only part of it. The question becomes “which part of this image is real and which part is fake”? With standard detectors, the classification can depend on the size of the manipulated areas. Detection is complicated by the fact that the manipulation can often be done with tools included in a mobile phone’s camera app. These tools may either keep the same format of the original image or add metadata indicating AI processing in every single image.
In the example below, AI has been used to clean a small part of the image: luckily, good old traditional deterministic methods keep working.
The Problem with AI Image Enhancement
This is just the tip of the iceberg.
More recently, what caught my attention was not intentional manipulation aimed at misleading. Rather, it was the risks of what AI-based enhancement could cause, even with (usually) good intentions.
A few weeks ago, I was having a chat with a long-time customer about deepfakes. They told me that they have the biggest issues with people who are supposedly trying to help them.
More specifically, they were talking about a case in which a witness gave them some photos enhanced with AI to give them a “better” picture. Then they realized something was off and asked for the original, which, of course, was of much worse quality. When they enhanced the original with proper forensic image enhancement methods (without AI), they got a worse result than what was originally provided. And a further big problem is that the prosecutor doubted their skills because it seemed that an amateur got a much better result.
The problem lies exactly here: as Sam Altman said of the early ChatGPT, it gives you “a misleading impression of greatness”. AI provides the impression of very good, trustworthy information, while it is actually something entirely made up, yet presented in a way that seems legitimate and reliable.
The truth is that there’s no clear distinction between an AI-enhanced image and a deepfake. Usually (but not always), the underlying technology is the same.
Inspired by this discussion, I went outside my office and took a picture of a car from far away.
Then I extracted a small part of the image with Amped FIVE.
I asked ChatGPT to enhance the license plate.
To my surprise, it refused to do so, both for privacy reasons and claiming that it wouldn’t be reliable. I was wondering whether this is because it has a lot of context about what I do, or because it is believing in it. So I tried it in a temporary chat without the memory related to what I do, and still struggled to get something.
Of course, not all LLMs are so scrupulous, and Gemini was a click away.
Impressive, isn’t it?
Except it is a completely made-up, but believable plate (well… if you ignore that numbers are not allowed in the first two characters of Italian license plates).
This is the actual license plate from another, higher-resolution image.
Things like these happen quite often. For instance, a user recently shared that police colleagues sent them a video that had been processed using a popular AI video enhancement software. Fortunately, they were using the free version, which superimposed the producer’s logo, making the manipulation immediately noticeable.
What would have happened to the evidence if they hadn’t noticed?
How often does this go undetected?
While these incidents are concerning, courts have already started to express skepticism about such evidence.
One notable case happened in the state of Washington, in the US, in March 2024. During a trial for homicide AI-enhanced video evidence was rejected after a Frye hearing.
AI Enhancement in the News
The web is full of recent examples of wrong AI enhancement use.
At the beginning of 2026, Grok was used to “unmask” the masked ICE agent who shot a lady in Minneapolis. The result? It hallucinated a face and a name, leading to harassment of completely unrelated people.
Same pattern, different case. When surveillance footage of a masked suspect in the Nancy Guthrie kidnapping case was released, X users immediately started asking Grok to “unmask” him, getting back fabricated faces with zero connection to reality.
But there is something even more subtle. While the first example is clearly completely unreliable (despite the X user writing, “Better than nothing“), other people focused on colorizing the infrared video. There is no way to correctly colorize an infrared video because it’s not simply a grayscale version of the visible light; it’s a completely different part of the electromagnetic spectrum. Different versions of the same images had completely different colors, which is a pretty clear indication of their unreliability. And of course, people trying to do this were complaining about the investigators not doing the same.
But it can get even worse. What if we can’t trust even the original coming out of our phones? There’s no doubt that what our smartphone displays on the screen is different from what the camera sensor initially captured. Computational photography is doing a lot to make us appreciate the pictures we have more.
I often get questioned on how this is affecting the authenticity of our smartphone photos, and the reality is that nobody (apart from maybe the producers) knows.
In general, Apple devices seem to be more conservative, or simply, they are more secretive about what their devices do. Samsung has been caught adding teeth to infants, adding way too many details to moon photos, and openly declaring that no photo is real anymore.
With this kind of processing, we don’t expect large details to change, but what about small ones? These are where forensic investigations are more focused on: the mole that could identify a person, a license plate’s letter of a car passing a faraway corner of the image… For now, we could probably put more trust in video than in images, as the processing speed needed to reach 30 or more frames per second limits what can be done. But of course, single video frames are usually of lower quality and affected by higher compression.
Compressing Reality
And speaking about compression, let’s go another step below. What if there’s AI in the way media files are encoded? Video surveillance companies have claimed vague AI compression features for a few years now. JPEG AI is being advertised as an upcoming standard (that, by the way, has very little in common with JPEG apart from the name). This very interesting paper shows very scary examples of what can (and does) happen with AI codecs.
Conclusions
This post has explored the complex and nuanced nature of the concepts of “real” and “fake” today. It’s a spectrum that must be defined and understood on a case-by-case basis, even before undertaking any technical analysis.
For this very reason, within Amped Authenticate, we provide an extensive array of analytical tools:
- some are based on file formats,
- others on the physics of the scene or pixel statistics inconsistencies,
- and even AI-based classification.
Such varied methodologies enable the forensic expert to address diverse questions and inspect the media from different angles, all while ensuring the analyst remains firmly in control of the final interpretation.
When I started, the biggest issue was the so-called CSI effect. When we told people we were doing forensic video enhancement, they expected us to do the impressive things they used to see in the investigative TV series. For years, we told them that it is not possible. Now it is. Except it is still fiction. Unfortunately, a very realistic, very convincing, and very dangerous fiction.