At the Netherlands Forensic Institute (NFI), we bring forensic expertise to the development of methods for detecting and assessing synthetic media in DETECTOR. We see first-hand how generative AI is changing digital evidence in criminal investigations.
While early discussions around deepfakes often focused on theoretical threats, criminal investigations now regularly involve synthetic or manipulated media. These can range from audio-cloned voice messages used in fraud to manipulated video footage submitted as evidence or raised as part of defence strategies.
Why Detection Alone is Not Enough
When looking at synthetic media from a forensic perspective, the primary focus centres on evaluative reporting and likelihood ratios. The challenge is not only to detect manipulation, but also to understand how reliable a result is and what it means for an investigation. When presenting findings to the court, we must express the weight of evidence using likelihood ratios to carefully weigh competing propositions. Translating technical results—whether derived from sensor noise anomalies, lighting inconsistencies, or specialised detection algorithms—into a statistically sound framework helps judges, prosecutors, and defence counsel evaluate the evidence properly without overstating our scientific certainty.
This approach is important not only for investigators and courts, but also for researchers and technology providers developing future detection methods. Results need to be understandable, measurable and suitable for expert interpretation.
From Research to Real-World Cases
Another challenge is ensuring the technical robustness and validation of the detection methods. While academic benchmarks and public databases provide a useful starting point for testing deep neural networks, real-world case media frequently falls outside those narrow training parameters. Our analysis pipelines must be fully validated, error-quantified, and rigorously tested against adversarial conditions. Because generative techniques evolve at such a rapid pace and creators can quickly adapt once detection signatures become public, we need generalised detection frameworks rather than fragile single-artefact classifiers. Detection methods therefore need to be tested under realistic conditions, with known limitations and error rates. Generative AI is developing so quickly that methods that detect one type of manipulation today may become less effective as generation techniques change.
Protecting Evidence from the Start
Finally, proper handling must begin at the earliest stages of an investigation. Reliable detection also depends on what happens before media reaches a forensic laboratory. Maintaining a strict chain of custody, verifying metadata, and preserving original container structures are essential.
This is why cooperation between law enforcement, forensic experts, researchers and technology providers is essential. Practical guidance can help investigators recognise suspected synthetic media early and preserve it correctly for specialised analysis.
Looking Ahead
DETECTOR aims to bridge the gap between deepfake detection research and its practical use in criminal investigations. The project supports the development of reliable and validated detection technologies that can be used and interpreted in real-world cases. This also requires clear standards for securing and processing digital evidence, so that forensic results can be assessed with an appropriate level of confidence.
By bringing together forensic expertise, law enforcement needs and technological research, DETECTOR can support the responsible use of synthetic media detection in criminal investigations and legal proceedings. DETECTOR will also work with law enforcement agencies, forensic institutions and policymakers to support practical uptake, more consistent practices and better cooperation across borders.
