AI use in chain of custody 1171/2025

FOI Reference: 1171/2025

Request

I am investigating how AI-assisted evidence analysis is documented and reported within existing Chain of Custody (CoC) and Audit Trail (AT) structures.

For the sake of clarity, I have included a glossary at the end of this document. CoC is used here as an umbrella term encompassing related concepts such as continuity, chain of evidence, and provenance; while AT refers to any documentation system that captures and records the forensic processing of evidence.

Is the chain of custody (CoC) and/or audit trail (AT) recorded when AI systems are used in processing and analysing forensic audio and textual data, or other forms of evidence?

If yes, please provide:

1. Details of the information that is recorded, which may include but is not limited to: identifier of the evidence, details of the AI model (e.g., version, provider), verified error rates, and identification of human oversight (if applicable).
2. Please indicate whether any paradata is captured during the processing of digital evidence. By paradata, I refer to information generated as a by-product of human interaction with the AI systems during evidence processing (e.g. prompts, edit history, data discards, interaction logs, correction logs, reviewer annotations, parameter adjustments, iterations).
3. Details of any technical or procedural methods used to document CoC and/or AT of AI-assisted evidence analysis, including any templates or forms. Kindly specify if the process is automated, manual, or hybrid.
4. Has your force conducted any internal evaluations or comparisons regarding the efficiency, accuracy, or reliability of automated, manual, or hybrid reporting methods? If yes, please provide copies or references.
5. Any guidelines or practice directions guiding the documentation of CoC and/or AT specific to AI-assisted evidence analysis. Kindly attach or reference related documents.
6. Is the maintenance of a CoC and/or AT managed in-house or outsourced?
7. If in-house, provide details of the responsible role(s) or department(s), including any training materials or internal guidance relevant to this task.
8. If outsourced, please provide the name of the provider(s), a description of their role, and any related contract document(s).

Glossary

* Artificial Intelligence (AI): This refers to a machine that learns, generalises, or infers meaning from input, thereby reproducing or surpassing human performance. The term AI can also be used loosely to describe a machine's ability to perform repetitive tasks without guidance.

* Audit Trail (AT): Refers to the documentation processes that capture, record, and report the sequence of actions applied to digital evidence. An AT may include details such as model configurations, parameters used, error rates, human interventions, and system-generated logs. It functions as a transparency mechanism, enabling the reconstruction and evaluation of how evidence has been processed, analysed, and interpreted.

* Automatic Speech Recognition (ASR): A subfield of AI that processes spoken human language (audio) and converts it into written text, useful for tasks like transcription and speaker diarisation.

*  Chain of Custody (CoC): This is the record of how evidence is handled from the point of collection through processing to presentation in court. This may include the identity of the person handling the evidence, date/time of transfer, storage details, and digital identifiers. Alternative terms include 'continuity', 'chain of evidence' and 'provenance'.

*  Forensic Audio evidence: This refers to recorded sound collected, preserved, and analysed for legal or investigative purposes. This often includes spoken language from sources including body-worn cameras, dashcams, police radio communications, and interview or interrogation recordings.

*  Forensic Textual evidence: This refers to written or encoded content, stored or transmitted in digital form, collected, preserved, and analysed for legal or investigative purposes. It includes unstructured human language

(e.g., transcripts, emails, chat logs, social media posts), semi-structured data (e.g., system or network logs, ASR outputs), and structured data (e.g., metadata, timestamps).

* Large Language Models (LLMs): A type of Natural Language Processing (NLP) system trained on vast amounts of text to perform sophisticated language tasks like summarisation, translation, or text generation.

*  Natural Language Processing (NLP): A subfield of AI focused on the interaction between computers and human (natural) language. It involves programming computers to process, analyse, understand, and generate human language. It is useful for tasks like translation, document generation, summarisation and redaction.

*  Paradata: This refers to information generated as a by-product of human interaction with AI systems during evidence processing. This includes prompts, edit histories, data discards, interaction logs, correction logs, reviewer annotations, parameter adjustments, and iterative outputs. Such records provide traceable documentation of how humans and AI jointly contribute to outputs, supporting transparency and accountability in evidential practice.

I am investigating how AI-assisted evidence analysis is documented and reported within existing Chain of Custody (CoC) and Audit Trail (AT) structures.

For the sake of clarity, I have included a glossary at the end of this document. CoC is used here as an umbrella term encompassing related concepts such as continuity, chain of evidence, and provenance; while AT refers to any documentation system that captures and records the forensic processing of evidence.

Response

I can confirm that Dyfed-Powys Police does not hold the information requested, due to the fact that Custody systems in Dyfed-Powys Police do not use AI through the CoC or Audit Trail. Neither does the Digital Forensic Unit use AI to assist with any of the points asked in this request.

It should be noted that as a result of the systems adopted by Dyfed-Powys Police in relation to the recording of such information that the information released may or may not be accurate.

(This is a response under the Freedom of Information Act 2000 and disclosed on 20/01/2026)