Researchers in Italy have proposed a new framework for evaluating artificial intelligence tools used in forensic DNA analysis, emphasizing that software validation should reflect the type of forensic conclusion a system may influence.
The review, published on 3 November 2026 in the journal Forensic Sciences, was authored by Paolo Bailo, Lorenzo Franceschetti, Maria Garreffa, Paola Primignani, Alberto Blandino and Andrea Piccinini.
Artificial intelligence is increasingly being explored in forensic genetics for tasks including DNA profile examination, detection of analytical artifacts, biological material classification and assistance with complex DNA mixture interpretation.
However, the researchers argue that these applications should not be treated as though they present identical scientific risks.
A system designed to identify unusual peaks in a DNA electropherogram performs a different function from software that influences conclusions about how biological material arrived at a crime scene.
The authors introduce the concept of validation proportionality, under which testing requirements depend on the forensic question, intended use and potential consequences of errors.
Their framework distinguishes seven functional positions where computational tools may influence forensic genetic interpretation.
These include analytical signal processing, DNA contributor interpretation, probabilistic genotyping and mixture analysis, biological-source identification, transfer-related evidence assessment, activity-level evaluation and expert reporting.
The researchers emphasize that probabilistic genotyping and Bayesian statistical methods should not automatically be described as artificial intelligence merely because they use computational software.
A major concern is the possibility of an AI-generated result being interpreted beyond its scientifically validated purpose.
For example, a system might correctly classify biological material as saliva, but that finding cannot independently establish how, when or why the material was deposited.
Similarly, detecting a person's DNA does not necessarily prove that the person directly handled an object or participated in a particular activity.
To develop their framework, the authors conducted structured literature searches across PubMed, Scopus and Web of Science.
The searches initially identified 6,733 records. After deduplication and the addition of selected guidance sources, the researchers mapped 3,695 records and documents, with a curated core evidence map containing 168 sources.
The authors recommend task-specific testing, representative datasets, documented limitations, version control, auditability and meaningful human oversight.
They also argue that AI may assist with retrieving scientific literature, organizing experimental information and supporting expert-led sensitivity analysis.
However, AI should not independently formulate competing activity-level propositions, generate definitive casework likelihood ratios or reconstruct historical events.
The researchers acknowledge that their publication is a narrative framework review rather than an experimental validation study.
Consequently, the proposed approach has not itself been prospectively validated as a laboratory procedure or quantitative risk-scoring system.
Further evaluation across forensic laboratories would be necessary to establish how effectively the framework supports operational decision-making.
The review highlights an important principle for forensic science: the reliability of an AI system depends not only on whether its calculations are accurate, but also on whether its outputs are used within scientifically justified boundaries.