The Future of Forensic Science: AI, Databases, Rapid Methods

Pattern comparison disciplines (fingerprints, toolmarks, handwriting, bitemark, tyre marks) share a common problem: the analyst makes a categorical call based on experience and training, but there is no agreed quantitative threshold separating identification from exclusion, and error rates are poorly characterised. Machine learning offers a different approach. Train a deep convolutional network on hundreds of thousands of labelled image pairs and it learns a similarity function that can be applied to new pairs with a score rather than a category.

The most developed application is fingerprint comparison. Systems such as NIST's Fingerprint Vendor Technology Evaluation (FpVTE) test suites compare commercial AFIS algorithms on large datasets of controlled impressions. The best-performing algorithms now have false non-match rates below 0.1% at a false match rate of 0.01% on the test datasets. Those are better numbers than most human examiners produce on the same test materials. But the test materials are controlled, and latent prints from real crime scenes are not. A model trained on rolled inked prints may degrade on a smeared partial latent on an irregular surface.

Two structural problems slow adoption. First, many ML models are black boxes: the network produces a score but cannot easily explain which features drove it. In court, where the analyst must be able to describe their reasoning, that opacity is a serious obstacle. Second, model performance is tied to the training data. If the training set over-represents one demographic group or one instrument, the model may perform worse on images that look different from what it was trained on. Transfer performance between datasets is the key validation test, and it is frequently omitted from published evaluations.

Rapid DNA instruments, commercially available since around 2012 from companies including ANDE Corporation and Thermo Fisher Scientific, automate the entire STR profiling workflow into a sealed cartridge. A buccal swab is inserted, a reagent cartridge is loaded, and roughly 90 minutes later the instrument uploads a profile. The FBI's Rapid DNA Program began piloting the connection of these instruments to CODIS in 2019, with the operational standards for booking-station uploads taking effect in September 2020, allowing arrestee profiles to be searched against the national database without the sample ever going to a laboratory.

For its intended use, single-source, clean reference samples from buccal swabs, the technology performs reliably. In that context, the error rate is low and the profile quality matches laboratory STR analysis. Disaster victim identification teams have used rapid DNA instruments to generate profiles from recovered remains for comparison against family reference samples in the field, reducing turnaround times from days to hours.

The oversight question is particularly sharp because rapid DNA at the booking stage shifts DNA collection from post-conviction (where it is widely accepted) to pre-conviction, creating a database of profiles from people who are eventually acquitted or never charged. Retention and expungement policies for these profiles vary enormously between jurisdictions, and many have not been updated to address the rapid-DNA scenario.

The Combined DNA Index System (CODIS) in the United States held approximately 22 million profiles as of 2024. The UK's National DNA Database (NDNAD) held roughly 6 million individuals (7.2 million subject profile records, including duplicates). These are the largest national forensic databases, but the category is growing: fingerprint databases (AFIS systems), facial recognition repositories, mobile phone location databases, and vehicle recognition networks all operate on the same logic of cataloguing identity information to enable future matching.

• Demographic imbalance: in many countries, Black and Hispanic individuals are over-represented in DNA databases relative to their share of the population, a direct consequence of differential arrest rates. This means their relatives are also more likely to be reached by familial searching, creating a compounding disparity.
• Familial searching: searching a database for partial matches to identify relatives of an unknown contributor is legal in some US states, England and Wales, and a handful of other jurisdictions, but prohibited in others. The Golden State Killer case (2018) used a commercial genealogy database rather than CODIS for this purpose, prompting new debates about the boundaries of forensic database use.
• Retention and expungement: profiles collected from people who are acquitted or never charged are retained in many national databases, sometimes indefinitely. The European Court of Human Rights ruled in S and Marper v United Kingdom (2008) that indefinite retention of profiles from unconvicted individuals violated Article 8 of the ECHR.
• International data sharing: Interpol's DNA Gateway and the Prüm Convention (EU) enable cross-border database searching. The governance frameworks for these exchanges have not kept pace with their technical capabilities.

Portable analytical instruments have become genuinely powerful. A handheld Raman spectrometer can identify many common narcotics through a sealed plastic bag in seconds. A portable X-ray fluorescence (XRF) analyser can determine the elemental composition of a glass fragment in the field. Near-infrared instruments can screen for food adulteration, explosive precursors, and soil composition at a scene. A field mass spectrometer can identify chemical warfare agent precursors without sending samples to a distant laboratory.

The validation problem is structural. Every instrument is validated for specific matrices (the type of sample), specific conditions (temperature, humidity), and specific reference libraries. Field conditions are none of these: samples are irregular, contaminated, mixed, or degraded. The reference library was built on pure compounds, not street-level adulterated powders. A false-negative in the field means a genuine threat is missed. A false-positive in the field means a person is detained based on an incorrect reading that may not survive a confirmatory laboratory test.

Before continuous probabilistic genotyping software, laboratories applied a binary threshold to mixed DNA profiles: if the number of contributors was too uncertain, or the stochastic effects (dropout, drop-in, pull-up) too severe, the profile was reported as 'inconclusive'. A large proportion of crime-scene samples fell into this category and were not further interpreted. Probabilistic genotyping software such as STRmix (New Zealand Institute of Environmental Science and Research), TrueAllele (Cybergenetics), and ArmedXpert (Forensic Science International group) changed this.

These systems model the entire electropherogram probabilistically, using Markov chain Monte Carlo sampling to estimate the most probable genotype combinations for each contributor and the stochastic parameters that explain the observed peak heights and ratios. The output is a likelihood ratio for each contributor of interest rather than a binary call. Profiles that would previously have been excluded from analysis can now produce LRs in the millions for high-quality three-person mixtures.

The broader lesson is that computational power does not resolve underlying uncertainty; it makes that uncertainty explicit and tractable. A well-validated probabilistic genotyping system provides better answers than a binary threshold precisely because it quantifies what is unknown. But 'well-validated' requires substantial ongoing investment in proficiency testing, software auditing, and inter-laboratory comparison, investment that smaller laboratories often cannot sustain.

The NAS (2009) and PCAST (2016) reports reached a common conclusion: forensic methods reach courtrooms before their reliability has been rigorously established. The mechanisms that validate pharmaceuticals or aviation components before deployment do not exist in most forensic disciplines. There is no pre-market approval, no mandatory efficacy threshold, no systematic post-deployment surveillance for errors. A laboratory can adopt a new method, conduct internal validation, and begin testifying within months.

Several bodies are working to close that gap. In the United States, the Organisation of Scientific Area Committees for Forensic Science (OSAC) at NIST publishes discipline-specific standards and best-practice documents covering validation requirements, uncertainty reporting, and proficiency testing frequency. ISO/IEC 17025 accreditation, which is required for testimony in many jurisdictions, mandates method validation and uncertainty estimation. The UK's Forensic Science Regulator publishes Codes of Practice that have legal effect for police and public forensic providers since the Forensic Science Regulator Act 2021.

• Independent validation: methods should be validated by parties without a commercial or institutional interest in the outcome, using datasets that are representative of real casework conditions, not idealized test materials.
• Transparency and reproducibility: software, algorithms, and population databases used in casework should be subject to independent audit. Closed-source forensic software used for criminal testimony is a governance problem, not just a technical one.
• Post-deployment surveillance: once a method is in use, error rates should be monitored through blind proficiency testing and case review. A baseline published error rate from a validation study may not reflect the lab's ongoing performance.
• Judicial gatekeeping: Daubert standards in the US federal system and equivalent admissibility tests in other jurisdictions require judges to assess scientific validity before testimony is admitted. The PCAST report recommended that judges apply these standards more stringently to pattern-comparison evidence.

The tools covered in this topic are, in principle, opportunities to reach more accurate answers more quickly, and to analyse evidence that previous methods could not. Whether they fulfil that promise depends on the rigour with which they are introduced, validated, and governed. Forensic scientists who adopt new methods without adequate validation, or who testify beyond what the science supports, bear direct responsibility for the outcomes.