The Pulse
Magnet Launches AI Research Lab After 91% Forensics Test Score
Magnet Forensics launched a research lab on October 26, 2026, to publish peer-reviewed studies of AI in digital forensics. Its first study found the company's expert system scored 91% on a narrow media-authentication test, compared with a 6
AI.info Team · Written with AI · fact-checked ·
Magnet Forensics’ AI expert system scored 91% on a test of media-authentication results, against a 69% average for human examiners. The company announced its Magnet AI Research Lab on October 26, 2026, using that peer-reviewed study as the new program’s first example of research aimed at testing AI tools used in digital investigations.
A narrow test of media-authentication judgments
The study, published in the Journal of Forensic Sciences on June 11, compared Magnet Verify’s Automated Analysis expert system with human examiners interpreting evidence about video authenticity and provenance. The test covered 14 videos, including camera originals, edited and social-media files, an AI-generated video, and files selected to challenge the system’s reference library. Twenty examiners assessed the material; researchers compared their responses and the system’s outputs with known ground truth.
“Across 20 respondents, the human mean score was 69% as compared to the expert system score of 91%,” write Brandon Epstein, Bertram Lyons, Daniel Fischer and Albert Zlogar in the paper. The authors describe the work as a preliminary study, not a broad evaluation of AI systems or forensic tasks beyond media authentication.
Two inconclusive answers, no incorrect ones
The percentages reflect a point-based scoring rubric, rather than a simple tally of right and wrong classifications. The expert system earned 64 of 70 possible points: it gave a fully correct analysis on 12 items and an inconclusive response on two, with no incorrect analysis. Human participants averaged 47.95 points out of 70, or about 69%.
That distinction matters for interpreting the headline result. The system’s two inconclusive responses received partial credit under the study’s rubric, while the absence of incorrect responses came from this 14-file test—not from evidence that the tool never makes mistakes in real investigations. The paper also reports that human scores did not correlate with years of experience; specialized media-authentication training was more closely associated with performance.
Magnet will support outside research questions
The new lab plans to conduct studies and work with practitioners, agencies and academic researchers. Magnet says it can assist with research design, methodology and publication, or connect examiners with academic partners. The company presents peer review as a way for investigators, agencies and courts to assess evidence about tool performance rather than rely on vendor claims.
The first study’s authors are affiliated with Magnet Forensics, and the paper discloses that they received research funding from the company, which is developing products related to the work. The authors also identify the study’s small participant sample as a limit on statistical analysis. They say broader studies could test other kinds of AI systems and examine how their outputs compare with human interpretations.
Peer review does not settle the deployment question
Magnet’s study evaluates a specific expert system interpreting structured media-authentication data, not a general-purpose chatbot and not every automated method used to identify synthetic media. Its score offers evidence about performance on the files and scoring rules used in the experiment; it does not establish how the system performs across different devices, evidence collections or courtroom settings.
The company says the lab will bring more research questions into peer-reviewed venues, while the initial paper calls for larger studies and additional testing. The immediate question for the field is whether independent teams can reproduce the result on broader samples and under conditions closer to the evidence examiners encounter in practice.