When you click on a sponsoring school or program advertised on our site, or fill out a form to request information from a sponsoring school, we may earn a commission. View our advertising disclosure for more details.
“What I’m afraid of is that authentic videos are going to be called deepfake videos. Just imagine you go into court, and you see a video you don’t like, and then you start to claim it is AI. This is why we need digital forensic professionals. We need to separate the real one from the deepfake one.”
Halil Dursunoglu, Faculty Specialist in Computer Science and Cybersecurity, Western Michigan University
A deepfake is video, audio, or imagery that artificial intelligence has generated or altered to show something that never happened. The first widely circulated examples gave themselves away. Faces flickered at the edges, subjects blinked at odd intervals, and the mouth and the audio drifted apart just enough for a careful viewer to catch it.
Those tells are disappearing. Generative video models released in the last two years produce clips that hold together frame to frame, and the flaws that once made detection a matter of looking closely are largely gone.
Fraud investigators saw the shift first. The FBI’s Internet Crime Complaint Center devoted a section of its annual report to artificial intelligence for the first time in the report’s nearly 25-year history, documenting roughly $893 million in losses across 22,364 complaints tied to AI-enabled fraud in 2025. The report describes criminals using voice clones, fabricated identification documents, and believable video of public figures and loved ones.
Courts are moving more slowly. The federal judiciary’s Advisory Committee on Evidence Rules has drafted a proposed Rule 901(c) that would create a tougher authentication standard for evidence a party claims was fabricated by generative AI, and the committee has now declined twice to publish it. A Federal Judicial Center survey of federal trial judges found that of the 931 who responded, only 15 had encountered a deepfake challenge in their courtrooms.
Halil Dursunoglu, who researches deepfake detection at Western Michigan University, thinks the courtroom problem will arrive from the opposite direction: “Right now I’m not afraid of the deepfake videos. They can be detected. We have so much research and so many resources for that, especially in the United States. What I’m afraid of is that authentic videos are going to be called deepfake videos,” he says. “This is why we need digital forensic professionals. We need to separate the real one from the deepfake one.”
Keep reading to learn more from Dursunoglu about what makes current deepfakes harder to catch, why detector scores are not proof, and what forensic training programs need to add to keep pace.
| Featured Cybercrime & Fraud Programs | ||
|---|---|---|
| Arizona State University | Digital Forensics (BS) | |
| Purdue Global | BSA - Auditing & Forensic Accountancy | |
| Southern New Hampshire University | BS in Accounting - Forensic Accounting & Fraud | |
| American Military University | Cybercrime (Grad Cert) | |
| American Military University | Cybersecurity (Bachelor's) | |
| American Military University | Digital Forensics (Undergrad Cert) | |
| Stevenson University Online | Online Master's in Forensic Accounting | |
| sponsored × When you click on a sponsoring school or program advertised on our site, or fill out a form to request information from a sponsoring school, we may earn a commission. View our advertising disclosure for more details. |
||
Meet the Expert: Halil Dursunoglu
Halil Dursunoglu is a faculty specialist in the Department of Computer Science at Western Michigan University, where he teaches in the university’s cybersecurity program. His research covers AI-generated media security and deepfake detection, focusing on audiovisual synchronization, whether a speaker’s lip movements, facial expressions, and timing match the audio.
In 2026, he published a review of synchronization-based detection methods in the International Journal of Computer Techniques, and he has also published research on human activity monitoring using wearable sensors.
Dursunoglu holds a master’s in computer science from Western Michigan University and a bachelor’s in computer science engineering from Yeditepe University in Türkiye. He works as a senior software engineer alongside his faculty appointment.
What Makes Today’s Deepfakes Different
Deepfakes are not new. The technology has circulated for five or six years, but the pace of change makes that stretch feel longer. “A year in computer science is more like 100 years of the human race,” explains Dursunoglu. “It is moving rapidly.”
The earliest deepfakes to reach a wide audience swapped one face onto another, and they left evidence behind. Edges warped, blink rates fell outside a normal range, and skin texture shifted between frames. Examiners could work from those artifacts. Generative video models do not produce them the same way.
“So far, in the deepfake videos you see, the old generation, we say there is always a fault in the frames that can be seen by the human eye easily. It is a blinking problem or a facial problem. When you see it, you understand,” he continues. “Rather than this kind of fault, I am focusing on the generative AI deepfake videos. Those are usually more consistent. I am not saying all of them cannot be detected by the human eye, but mostly the human eye cannot detect that fake version.”
A 2025 systematic review in Human Behavior and Emerging Technologies pooled 40 studies of people asked to sort authentic footage from fabricated footage and found average detection accuracy running between roughly 58 percent and 75 percent. Performance depended heavily on how the video was made. Participants caught face-swapped videos 91.3 percent of the time but identified lip-synced videos only 52.6 percent of the time, which is barely better than guessing.
Lip sync is Dursunoglu’s territory. Instead of hunting for visual artifacts, he compares signals that should agree in any authentic recording. “Every video, every image, every audio is a signal. Multimodal signals come together. Right now I am talking to you. If my facial expression is not matching my lip movements, or my lip movements are not matching the phonetics, or my hands are not matching what I say, then the digital signals need to be examined,” he says.
The speed that improved the fakes also ages the tools built to catch them. Detection models learn the signatures of the generators they trained on, and a new generator does not leave the same signatures. “One year ago we were talking about deepfake videos, but they had those facial faults and other things people could understand. Recently what we are seeing are purely deepfake videos that cannot be detected by detectors at all,” notes Dursunoglu. “We are trying to use yesterday’s detectors to detect tomorrow’s deepfake videos, so it is not going to work.”
Why a Detector Score is Not Proof
Deepfake detection tools exist, and forensic examiners will use them. Dursunoglu’s objection is not to the tools but to what people expect from their output. A detector returns a probability, and a probability is not a finding. “I am not going to say that this or that detector is perfect, which it is not. None of them is perfect,” says Dursunoglu. “Even though a detector says there is an 85 percent chance it is fake, it does not prove that it is fake.”
Federal researchers reached a similar conclusion. The National Institute of Standards and Technology’s November 2024 report on synthetic content found that detectors trained on output from one generator perform measurably worse on output from a different one, with cross-generator testing producing accuracy as low as 61 percent in one study after routine post-processing. The report further states that none of the techniques it reviews offers a comprehensive solution on its own.
“Detecting AI is adding another layer to digital forensics. Let’s say right now we have six or seven layers. It is going to be eight layers,” Dursunoglu explains. “So they are going to use the traditional way. Then they start to use the detectors. That’s not enough. They need to understand the audiovisual synchronization. They need to check the lip movements and facial expressions. Then you can say it is a fake or real video.”
The older layers do not go away. Examiners will still need to establish where a file came from, whether the source can be trusted, what the metadata shows, and whether deleted files can be recovered, all before a detector enters the picture. Origin carries more weight than any detector score, and it can settle the question on its own. “It doesn’t matter if it is a deepfake or a real video. If it is coming from an untrusted source, or untrusted provenance, it is not acceptable,” he adds.
The federal courts are working through a version of the same problem. Proposed Rule 707, published for public comment in 2025, would subject machine-generated evidence offered without a supporting expert to the same reliability screening that applies to expert testimony. The Advisory Committee on Evidence Rules noted in its commentary that those reliability standards will be difficult, and sometimes impossible, to meet without expert testimony, which is where a bare detector output would run into trouble.
Dursunoglu’s own standard is that a video is “evidence, not a verdict,” and that a close call belongs with more than one examiner: “In this case they need to be examined by a third expert or a fourth expert, because we don’t want to judge anyone or create a verdict because of a deepfake,” he says.
When Real Video Gets Called Fake
Synthetic media cuts both ways, and Dursunoglu is more worried about the second: “Deepfakes create two main problems right now. Fake media can be presented as real, and real media can be dismissed as fake,” he explains. “People cannot believe what they see anymore.”
Legal scholars named the second problem before the technology could deliver it. Writing in the California Law Review in 2019, Robert Chesney and Danielle Citron called it the liar’s dividend: the advantage a wrongdoer gains once the public knows convincing fakes are possible and can dismiss any authentic recording as one.
More recently, a 2025 study in the American Political Science Review ran five survey experiments with more than 15,000 American adults, showing them real political scandals followed by a range of responses from the politicians involved. When the politician claimed the evidence was misinformation, the share of participants who believed the scandal had actually happened dropped from 44 percent to roughly 35 percent.
However, video has held up better than text. Claims of misinformation moved support for politicians caught in text-based scandals but had negligible effect when the scandal arrived on video, which suggests audiences still grant footage more credibility than a transcript. That advantage lasts only as long as the video stays hard to fake convincingly.
Skepticism carries its own cost. The same 2025 systematic review that measured human detection rates found that warning people a video might be manipulated made them doubt everything they watched, authentic clips included.
Judges appear to be bracing for the problem. In the Federal Judicial Center survey, 82 percent of the judges who had never encountered a deepfake claim said they would want some initial showing before entertaining an argument that a piece of evidence was fabricated.
Dursunoglu’s example is a head of state appearing on video to declare war on the United States. If the footage is not live, no one watching can rule out that a model produced it, and he points to social media as where a clip like that would spread before anyone authenticated anything. He does not think examiners have much time to prepare. “It is going to be very, very hard in the future. When I say future, I am not talking about 20 years from now. I am talking about next year or the year after that,” he says.
What Needs to Change in Forensic Training
Curriculum moves slower than the technology. Getting a single course changed at a university takes months of review. To prepare students for this new future, digital forensics training needs to build deepfake work into its core: “We have to stop treating deepfakes as a single lecture at the end of a digital forensic course. It should be an entire course,” Dursunoglu explains.
“Students should encounter synthetic evidence throughout the training. If they are learning image forensics, we need to include AI-generated images. If they are studying audio, we need to add cloned voices. If they are studying video authentication, we need to include manipulated and synthetic videos along with the unedited ones. Our goal is to make authenticity verification part of normal forensic thinking.”
The lab exercises he designs work like a crime scene without the crime scene. Students get three videos and no labels, one untouched, one lightly edited, and one generated. They examine the patterns, trace where each file came from, and write a report defending their conclusion.
Advice for Students Entering the Field
Students entering the field now are studying a subject that changes between semesters. Here is Dursunoglu’s advice for students who are considering this field.
Choose a Program With a Current Curriculum
“I hope they choose a university with an updated curriculum, because that is the most important part,” advises Dursunoglu. A program that folded generative media into its labs in the last two years graduates examiners ready for current casework. One that has not is training students against a threat model that has already changed.
Read the Papers Nobody Assigns
Coursework lags publication, and the gap is where the current methods live. “The things that are newly published, we don’t teach in the classes. There are so many newly published papers and articles that we don’t teach them in class. Students need to find those by themselves,” he says.
Build a Deepfake Yourself
Dursunoglu plans to have his own students generate deepfake videos this semester, under clear ethical limits, so they understand the process from the production side. Making one shows where a generator cuts corners and what it leaves behind, which is what an examiner has to recognize later.
Think Like the Adversary
“I teach all of them how to hack the system. Without hacking the system, you cannot understand the solution,” shares Dursunoglu. “They need to think a little bit like a criminal. If you don’t think like a suspect, you cannot get the solution.”
Kimmy Gustafson
WriterKimmy Gustafson’s expertise and passion for investigative storytelling extends to the world of forensics, where she brings a wealth of knowledge and captivating narratives to readers seeking insights into this intriguing world. She has interviewed experts on little-known topics, such as how climate crimes are investigated and prosecuted, and has written for ForensicsColleges.com since 2019.