Synthetic media has crossed a line. Deepfake videos, cloned voices and fabricated photographs now routinely surface in criminal investigations, insurance claims and civil disputes. What was once a novelty has become a systemic risk to truth itself. The world is facing an evidence crisis: when anyone can forge reality, the burden of proof becomes unmanageable.
We have decided to confront that problem head on. Over the past year we began developing a forensic artificial intelligence system capable of detecting, classifying and explaining synthetic media. The project is still in early development, with the first controlled tests showing promising (though cautious) results. It is difficult work at the intersection of technology, law and ethics, but the goal is simple: to make digital authenticity provable beyond reasonable doubt.
Why authenticity is collapsing
Courts have long relied on the visual and auditory power of evidence. A video or a voice recording can outweigh hundreds of pages of testimony. Yet the rise of generative AI has shattered that hierarchy. Today, a single person with a consumer GPU can fabricate a witness, a confession or a crime scene that passes ordinary human scrutiny.
Legal systems everywhere are struggling. Judges are not forensic analysts. Prosecutors must prove that an image or recording is genuine, while defense lawyers now question the authenticity of almost everything. Expert witnesses report exponential growth in authenticity disputes. Each side brings its own experts, each with different tools and uncertainty models. Trials slow down. Truth itself becomes procedural.
The legal foundations we build on
European law already offers a skeleton for digital truth, though it was never designed for deepfakes.
- The EU AI Act introduces transparency duties for synthetic content. Deepfakes must be disclosed, but disclosure is not proof. Courts need something verifiable, not declarative.
- The Digital Services Act pushes platforms to mitigate systemic disinformation. Yet even platform labels or watermarks do not carry legal presumptions.
- The eIDAS Regulation provides those presumptions: qualified electronic seals and timestamps that courts across the EU must recognize as proof of origin, integrity and time.
Our concept builds on that third pillar. If a forensic AI can output a signed, time-stamped and sealed report under eIDAS trust rules, its results gain evidential weight by law. The challenge is to make the technical chain match the legal chain.
The architecture of our forensic AI
We are developing a modular AI platform that analyses media from two converging perspectives: provenance (how the file came to be) and forensic signal evidence (what the file contains). Each result is cryptographically signed and embedded into an evidence credential.
1. Provenance and metadata
When a file enters the system, we extract any remaining metadata, signatures and known provenance markers such as C2PA content credentials, camera attestations or watermarks. This information forms the first layer of the authenticity model. It is cryptographically hashed and stored in an immutable ledger.
2. Physical and signal-level forensics
Our engineering team implemented a series of low-level detectors that measure inconsistencies invisible to the naked eye:
- Camera noise fingerprinting (PRNU) to link an image to a physical sensor.
- Lighting and shadow consistency using physically constrained models.
- Frame correlation analysis to detect duplication, interpolation or blending in video.
- Audio electric-network-frequency tracing to verify temporal continuity and detect splicing.
These methods are grounded in established forensic science. They provide structured probabilities, not binary judgments.
3. Learned AI detectors
Above the physics layer runs our core detection AI: a multimodal neural ensemble trained to recognize synthetic artifacts in images, videos and audio streams.
- It learns across modalities (vision, acoustics and text) to capture cross-modal inconsistencies.
- It is adversarially trained against a red-team generator fleet that simulates current and next-generation forgery models.
- It constantly recalibrates as new generative architectures emerge.
The system never outputs a simple “real/fake” label. Instead, it expresses a likelihood ratio, the probability that a piece of media would appear as it does if it were genuine versus if it were artificially generated.
4. Calibration and validation
Every model version is validated on held-out datasets, with measured false-positive and false-negative rates. We perform calibration so that confidence intervals correspond to empirical reliability. Each result is versioned, signed and time-stamped so that future courts can reproduce the inference under identical conditions.
5. Explainability
Each finding must be explainable. The AI generates visual or auditory maps highlighting the features that led to its conclusion, areas of the image where blending is detected, spectral bands where cloning artifacts appear, or metadata inconsistencies. Explanations accompany numeric scores so that human experts can audit them.
The early tests and first signs of hope
Our first laboratory evaluations were limited in scope but instructive. On a mixed dataset of real and synthetic video clips, the system identified 96 percent of deepfakes correctly, with a 2.3 percent false positive rate. In controlled audio tests it detected cloned speech with slightly lower confidence, around 90 percent, largely due to codec interference and compression artifacts. These figures are not courtroom-ready yet, but they demonstrate feasibility.
More importantly, the confidence calibration held: when the system claimed 90 percent certainty, it was correct approximately nine times out of ten. That statistical honesty is critical. Courts can tolerate uncertainty; they cannot tolerate hidden error.
We remain cautious. Generative models evolve monthly, and what works today may fail tomorrow. Robustness to new forgeries is the hardest challenge. But the combination of provenance, forensics and AI detection already reduces ambiguity dramatically.
The legal chain of custody
Each report the AI produces is treated as a digital affidavit. It contains:
- The original file hash and capture metadata.
- The cryptographic signature of the analysis software version.
- The qualified timestamp proving when the analysis was performed.
- A signed JSON structure with model outputs, likelihood ratios and explanations.
- A qualified electronic seal linking the report to our organization’s legal identity.
This package becomes a qualified electronic record under eIDAS. In legal proceedings, it carries a presumption of integrity and origin. Opposing parties may challenge it, but the burden of proof shifts, they must now disprove a chain that is mathematically and legally verifiable.
The governance framework
A project like this cannot rely solely on code. It requires governance.
- AI management system: we follow ISO/IEC 42001 guidelines to document every decision, dataset, model update and validation.
- Data ethics: all training data are licensed, consented or synthetic; no biometric scraping.
- Privacy by design: the system never stores personal data unless legally necessary for evidence.
- Auditability: every inference and every model version is reproducible, sealed and stored with a qualified timestamp.
We also cooperate with legal scholars to align our documentation with forensic reporting standards such as ISO/IEC 27037 and 27042, ensuring the format of our reports matches what courts expect.
Why this is technically hard
Detecting deepfakes is a moving target. Generative models learn to hide their fingerprints; codecs blur residual artifacts; social media platforms strip metadata that could have helped. Even forensic signals like PRNU can be spoofed or destroyed by resampling. Moreover, courts demand explainability and stability over time, while machine learning models thrive on continuous retraining. Balancing those demands stability versus evolution) is one of the deepest engineering and legal challenges of our time.
Our internal rule is conservative: if a method cannot be validated, documented and explained, it cannot appear in a court report. That slows development, but it is the only way to earn judicial trust.
The road ahead
The next phase of development focuses on field validation. We are partnering with investigative agencies, media organizations and legal experts to test the system in real-world workflows. The goal is to evaluate how the AI performs on messy, real evidence, compressed files, partial audio, edited footage, and how human reviewers interact with its explanations.
Parallel work is underway on an evidence credential API that packages each analysis result into a verifiable token. This will allow courts, insurers and media outlets to verify authenticity instantly using open standards.
If the pilot results continue to hold, we will apply for certification as a qualified trust service under eIDAS. That status would grant our signatures and timestamps direct legal effect across the European Union.
A cautious optimism
We are under no illusion that this will solve misinformation overnight. The task is technically demanding and legally complex. But the early signs suggest that a structured combination of cryptography, forensic science and transparent AI can make authenticity provable again.
Our ambition is not to police the internet but to rebuild confidence in evidence. Every image, every recording, every proof should be traceable to a verifiable origin. When the next generation of deepfakes arrives (and it will) the courts must already have tools that tell them, with evidence not opinion, what is real.
This is the beginning of that work.