When AI Is a Witness: Admissibility of Algorithmic Forensics in Indian Courts

Ten years ago, the forensic expert in the witness box described what they had personally observed: slides under a microscope, fingerprints on a comparison pad, blood on a laboratory bench. Today a growing share of forensic conclusions is produced by software. Face-recognition systems propose a match from millions of images. Speaker-attribution tools reduce a recording to numeric feature vectors and compare them. Location histories are reconstructed from tower dumps. Usually a human officer certifies the output, but the machine did the deciding.

Evidence law wasn't built for this. Cross-examination, the engine of the adversary system, assumes a witness with perception, memory, and something at stake. A model has none of these. The hard question isn't whether algorithms belong in investigations; they're already there, and used well they surface patterns no human reviewer could find. The hard question is what replaces scrutiny when the inference that decides the case came out of a system nobody in the courtroom has inspected.

Two failure modes human witnesses don't have

The first is opacity. The same output can arise from a genuine signal or from an artifact of training data, and the output alone doesn't tell you which. The second is miscalibrated confidence. The comparison underneath is arithmetic: the engine converts a face into a feature vector, measures its distance from the gallery, and returns every candidate below an operator-chosen similarity threshold. A match is a knob setting as much as a finding, and the ranked output alone never tells you how often the top candidate is wrong on data like this case's. When NIST stress-tested commercial face-recognition engines for its FRVT Part 3 report (NISTIR 8237, December 2019), it found false-positive rates ten to a hundred times higher for Asian and African descent individuals than for European descent individuals on some engines. Newer engines narrowed the gap. None closed it. An algorithm can be genuinely useful on average and systematically unreliable for the specific defendant in the dock.

Courts elsewhere have begun policing this gap by requiring disclosure of training data, error rates, and validation studies before such evidence is heard. India doesn't yet have a leading judgment on facial recognition. The framework will be built over the next few years, one case at a time, by the lawyers and forensic scientists who arrive prepared.

How Indian evidence law copes

The statute, now the Bharatiya Sakshya Adhiniyam, was drafted for documents and witnesses, not model weights, but its mechanisms reach algorithmic output by analogy. The system that produced a machine-generated report is a source of electronic records, so its output enters through the certificate route Section 63 provides. The expert explaining the method testifies under the expert-evidence provisions. And the courts' century-old insistence that scientific techniques be reliable and experts competent, built long before machine learning, does the heavy lifting no single provision does.

In practice, scrutiny runs through three questions. Was the underlying data reliable, collected and stored with the same care as any other evidence? Is the technique validated for this specific use, on populations and conditions like this case's, not merely impressive in a vendor demo? And can the process be examined by an independent expert on the actual artefacts, rather than a summary prepared by the prosecution? Where one of the three fails, the answer isn't reflexive exclusion. It's honest scrutiny of weight: the court may hear the evidence while declining to treat a probability score as an identification.

What to demand before accepting the output

• Validation documentation for the exact tool and version used: model identifier, preprocessing pipeline, and the similarity threshold in force. Not the vendor's marketing material.
• Known error rates, disaggregated where relevant, with the conditions under which they were measured.
• Preservation of the tool itself: models are retrained and pipelines updated, so the version examined at trial must be reproducibly the one that produced the report.
• Access to the actual inputs and outputs for independent expert examination, subject to lawful safeguards.
• A human witness who can explain the method and its limits, not merely read out the tool's conclusion.
• Clarity about what the output is: a lead, corroboration, or an identification. Each carries different weight.

'The software said so' should start an inquiry in court, never end one. The practitioners who learn to ask the three questions above are the ones who'll be able to tell a genuine digital witness from an expensive guess.

Back to all articles