The evidence envelope had a clear chain. Every person who had handled the digital file had signed it, timestamped it, and noted what they did. Then a defense attorney pulled the platform audit log and found a gap: at 2:14 a.m. on the night the AI-generated case summary was produced, the system had automatically ingested the original interview recording, transcribed it, and generated a draft. No officer had signed that step. No log entry showed who had authorized the ingestion. No record indicated whether the transcript the AI had produced from the audio matched the audio it received. The chain of custody, the unbroken record of who had the evidence and what had been done to it, had a gap right in the middle of the AI processing step. The defense attorney did not need to argue that the summary was wrong. He only needed to show that the chain was broken. A chain with a gap is not a chain at all. It is two separate pieces of evidence with nothing provably connecting them.

What Chain of Custody Means and Why AI Changes It

Chain of custody is the documented, unbroken record of who had possession of a piece of evidence, when they had it, what they did with it, and how it was transferred to the next person in the chain. In physical evidence handling, chain of custody procedures are built around the object: who collected it, who transported it, who logged it into the evidence room, who removed it for analysis, who returned it. The chain exists to establish authenticity (this is the same object that was collected at the scene) and integrity (the object has not been altered since collection).

Digital evidence introduced a new layer of complexity to chain of custody. When the "evidence" is a recording, a data file, or a digital record rather than a physical object, the questions of authenticity and integrity become questions of bit-for-bit identity: is the file that is being presented in court the same file that was captured at the scene, with no modification, corruption, or unauthorized access in between? Digital evidence platforms address this with hash verification, metadata logs, and audit trails that track every access and every export. The chain of custody for a BWC (body-worn camera) recording includes not just the officer's name and the date it was uploaded but the cryptographic hash of the file that proves the uploaded version is identical to the captured version.

AI introduces a third layer that existing chain-of-custody procedures were not designed to address. When an AI system ingests a piece of evidence (a recording, a transcript, a case file) and produces something derived from that evidence (a summary, a draft report, a transcription), the process creates a new artifact. That artifact is derived from the original evidence. It is not the same as the original evidence. And the process by which the original was transformed into the derivative is now part of the chain that must be documented if the derivative is going to be used in a legal proceeding.

The gap in the opening scenario was not a gap in the physical evidence chain. The recording was properly logged and authenticated. The gap was in the AI processing step: who authorized the ingestion, what did the AI do to the recording, what was produced, and how does what was produced relate to what went in? Those questions were unanswerable from the documentation that existed. The chain broke at the AI step because the AI step was not treated as part of the chain.

Chain of custody does not end at the evidence room door. It extends through every system that touches the evidence, including the AI that reads it to draft a report.

The Specific Chain of Custody Risks AI Creates

Understanding the specific risks requires thinking through how AI systems interact with evidence at each stage of processing. These are not hypothetical risks. They are the documented failure modes that emerge when AI is deployed in evidentiary workflows without chain-of-custody design.

Unauthorized Ingestion

AI-assisted report writing systems typically require access to the raw source material: the BWC recording, the CAD (computer-aided dispatch) entry, the interview audio. In many deployments, this access is automatic: when an officer initiates a report-drafting session, the platform automatically identifies and ingests relevant recordings. The ingestion step, where the evidence file is read by the AI system, is a custody event. Someone or something has accessed the evidence file and processed it. If that access is not logged, attributed to a specific authorized user, and timestamped, it is an unlogged custody event. Every unlogged custody event is a chain-of-custody gap.

The practical requirement is that every ingestion of evidentiary material by an AI system should generate a custody log entry that includes: the specific file or files ingested, the identity of the system that performed the ingestion, the timestamp, and the identity of the human who authorized the ingestion. Automatic ingestion that occurs without explicit human authorization and logging is a custody gap by design.

Transformation Without Documentation

When an AI system ingests a recording and produces a transcript, or ingests a transcript and produces a narrative draft, it has performed a transformation. The input and output are not the same. The output is a derivative work. In any context where the derivative work will be used as evidence or as the basis for a sworn statement, the transformation process must be documented: what went in (with identifying information sufficient to authenticate the input), what system performed the transformation, what version of the AI model was used, and what came out (with identifying information sufficient to authenticate the output and link it to the input).

This documentation requirement is not unusual in forensic contexts. When a forensic examiner processes digital evidence, they document their tools, their methods, their input, and their output. The AI processing step is analogous, but without explicit policy requirements, it is often left unlogged. The RMS (records management system) may capture that a draft was generated, but it may not capture what version of the AI model generated it, what specific audio file was processed, or whether the output was modified before the officer reviewed it.

The Intermediate Artifact Problem

AI-assisted workflows typically produce multiple intermediate artifacts before a final report is submitted. The BWC recording is ingested and produces a transcript. The transcript is used to generate a narrative draft. The officer reviews the draft and makes corrections. The corrected draft is submitted. Each of these is a distinct artifact, and each transition from one to the next is a step that must be documented if the chain is to remain intact.

In practice, many deployments preserve only the original recording and the final submitted report. The intermediate artifacts (the AI-generated transcript, the initial draft, the officer's annotated version) may not be preserved, may not be retrievable, and may not be linked to each other or to the final document in any logged chain. When a defense attorney asks how the final report was derived from the original recording, the answer "the AI produced it and the officer reviewed it" is not a chain of custody. It is a description of a process that cannot be audited.

The Data Integrity Problem

Chain of custody for digital evidence relies on hash verification to establish that the file has not been modified. A cryptographic hash is a mathematical fingerprint of a file: if any bit of the file changes, the hash changes, and the mismatch proves the file was altered. BWC platforms use hash verification routinely. When a recording is uploaded from the camera to the platform, the hash is computed and recorded. If the hash of the uploaded file matches the hash of the original recording, the files are bit-for-bit identical and the recording has not been tampered with.

When an AI system ingests a recording and produces a transcript or summary, hash verification of the original recording does not verify the accuracy of the AI output. The recording may be bit-for-bit authentic, and the AI-produced summary may still contain hallucinated content. The integrity assurance that hash verification provides for the input does not extend to the output. A chain of custody framework for AI-processed evidence needs to address both: the integrity of the original evidence and the verifiable relationship between the original and the derivative product.

Building Custody Into the AI Workflow

The good news is that the chain-of-custody requirements for AI-processed evidence are not fundamentally different from the requirements for other forensic evidence processing. The principles are the same: document what went in, who authorized it, what was done, what came out, and who received the output. The difference is in the implementation, because AI processing is often faster, more automatic, and more invisible than traditional forensic processing, and the gaps tend to form in the automatic steps that no one designed to be custody events.

The Five Custody Log Requirements

For every AI processing step that touches evidentiary material, the custody log should capture five elements.

First, the input identifier: the specific file or data that was ingested, identified with sufficient specificity that the same file can be located and verified after the fact. For a BWC recording, this means the file hash (the cryptographic fingerprint), the camera serial number, the start and end timestamp of the recording, and the case number it is associated with. For a CAD entry, this means the incident number, the CAD system timestamp, and the specific fields included in the ingestion.

Second, the system identifier: the specific AI system or platform that performed the processing, identified by name and version. AI models are updated. A transcript produced by version 2.1 of a transcription engine may differ from a transcript produced by version 2.3 of the same engine for the same input audio. The version is relevant to any challenge to the accuracy or consistency of the output.

Third, the authorization record: who authorized the AI system to access the specific evidence file, and what that authorization was based on. Automatic ingestion is not an authorization. An officer clicking "generate report" may constitute authorization, but only if the system logs that specific click, attributed to that specific officer's authenticated account, with a timestamp. Generic system access logs are not an authorization record for a specific custody event.

Fourth, the output identifier: the specific artifact produced by the AI processing step, identified with sufficient specificity to establish its relationship to the input. If a transcript was produced, the transcript file should have a unique identifier, a hash of its own, and a logged link to the input file that was transcribed. If a narrative draft was produced, the draft should have a unique identifier and a logged link to the transcript or recording it was derived from.

Fifth, the review record: who reviewed the AI output, when, what they found, what they changed, and what they certified. For an AI-generated report draft, the review record is the officer's footage-grounded verification pass, documented in the workflow log described in other lessons in this chapter. The review record is the link between the AI output and the sworn document: it shows what the AI produced and what a human verified and adopted.

The Preservation Obligation

The five elements above are only useful if the documentation is preserved. In many AI-assisted workflows, intermediate artifacts are not preserved by default. The AI-generated draft exists in the officer's workflow session and may be overwritten or discarded when the report is submitted. The AI-generated transcript may be a temporary file in the processing pipeline rather than a preserved artifact. If the intermediate artifacts are not preserved, the chain cannot be reconstructed, and the custody log is a record of a process whose inputs and outputs cannot be verified.

The preservation obligation for AI-assisted evidentiary workflows requires that every artifact in the chain (every input file, every intermediate product, every version of the draft that was reviewed) be preserved in a form that is retrievable, authenticated, and linked to the rest of the chain. This is a vendor contract requirement, not just an officer practice. The platform must preserve the AI-generated draft, not just the final submitted report. The platform must preserve the transcript, not just the recording it was derived from. If the vendor's platform does not preserve intermediate artifacts, the agency is operating with a chain-of-custody gap by design that no individual officer's practice can close.

CJIS and the Agency Obligation

The CJIS (Criminal Justice Information Services) Security Policy administered by the FBI governs how criminal justice information is stored, accessed, and transmitted. CJIS obligations stay with the agency. The vendor's CJIS certification covers the vendor's platform and infrastructure, but the agency remains responsible for ensuring that its use of the vendor's platform complies with CJIS requirements for access control, audit trails, and data integrity.

When an AI system ingests criminal justice information (BWC footage, interview recordings, case files), the access to that information is subject to CJIS audit trail requirements. The CJIS Security Policy requires that accesses to criminal justice information be logged with sufficient detail to reconstruct who accessed what and when. An automatic AI ingestion step that accesses criminal justice information without logging the access, attributing it to a specific authorized account, and preserving the log is not CJIS-compliant regardless of whether the vendor's platform is CJIS-certified. The certification covers the platform's capability; the agency's implementation determines whether the capability is used in a compliant way.

What a Broken Chain Costs in Practice

A broken chain of custody does not automatically result in evidence exclusion, but it provides the foundation for a defense motion to suppress or exclude the derived evidence, and it undermines the credibility of the entire investigative record. The motion does not need to prove that the evidence was tampered with. It needs only to show that the chain of custody was broken, that there was a period in which the evidence's integrity cannot be verified. Once that gap is established, the burden shifts: the prosecution must explain the gap or accept that the derivative evidence will face a credibility challenge.

In the AI processing context, the specific challenge is that the gap is structural rather than incidental. In a traditional chain-of-custody breach, something went wrong: a log was not filled out, a transfer was not documented, a person who handled evidence was not identified. The breach was a failure to follow an established procedure. In the AI context, if the procedure was never designed to cover the AI processing step, the gap is not a failure to follow procedure. It is the absence of a procedure, which is worse. There is no established practice to point to as the standard that was not met. The gap is systemic.

The practical impact is felt at exactly the wrong moment. A defense attorney challenging the chain of custody in an AI-processed evidence case is not looking for a single error. They are examining whether the agency has any documented framework for maintaining chain of custody through AI processing at all. If the answer is no, that fact becomes the foundation for challenging not just the evidence in this case but every AI-processed piece of evidence in every case the agency has built with this system. The systemic gap is more damaging than an individual error because it raises questions about the reliability of the entire class of evidence the AI system has produced.

The Custody Framework for AI Vendors and Contracts

Agency custody requirements need to be reflected in vendor contracts before the system is deployed. Once an agency has signed a bundled, multi-year, sole-vendor contract for cameras, cloud storage, and AI processing (contracts in the range of approximately $45 million over ten years have been reported in public safety procurements), the leverage to require specific custody features is largely gone. The custody framework needs to be a contract requirement, not a deployment afterthought.

The specific contract requirements that address AI chain-of-custody needs include: preservation of all intermediate AI-generated artifacts (transcripts, initial drafts) linked to the original source files; audit logs that capture every AI ingestion event with the input file identifier, the model version, the timestamp, and the authorization record; officer review logs that record the reviewer's identity, the timestamp, and the specific changes made; and export capabilities that allow the agency to produce the full custody chain (from original evidence file through every AI processing step to the final submitted document) in a format reviewable by defense counsel and the court.

Agencies that are still in procurement should evaluate vendors on these capabilities explicitly. Agencies already under contract should audit whether the existing platform provides these capabilities, whether those capabilities are enabled, and whether officers and supervisors are using them. The gap between a platform's capability and the agency's actual implementation is often where the chain breaks.

The King County, Washington, prosecutor who barred AI-written police reports from their cases was responding in part to exactly this documentation gap: agencies could not demonstrate the chain from source evidence to final report, and without that chain, the report was a document of unknown provenance. The lesson is not that AI-assisted reporting is inadmissible. It is that admissibility requires a documented chain, and building that chain is the agency's responsibility.

The Officer's Practical Role in the Custody Chain

Individual officers cannot design the platform, write the contract, or build the logging infrastructure. But they have a specific role in the custody chain that is theirs alone: the review and adoption step. The review-and-adoption step is the human link between the AI output and the sworn document. It is the step where a person with direct knowledge of the incident takes responsibility for the accuracy of the record.

The practical implications of the custody role for the officer are three. First, the review must be real. A cursory read-through that produces a submitted report is not a custody-chain review. A footage-grounded verification pass, as described elsewhere in this chapter, that confirms each factual claim against the source material and documents the verification, is a custody-chain review. The difference is the documentation, and the documentation is what makes the review part of the chain rather than a formality before the chain ends.

Second, the officer must be the identified reviewer. Anonymous review is not a chain-of-custody record. The log should attribute the review to a specific officer's authenticated identity, tied to their department credentials. If the platform does not support this level of attribution, the officer should create a paper or digital record outside the platform that documents their review: what they reviewed, what they found, what they changed, and their signature or authenticated approval.

Third, the officer's adoption statement should be explicit. The standard language for adopting an AI-assisted draft as a sworn account should include: a statement that the AI assisted in drafting the document; a statement that the officer reviewed the draft against the source evidence; a statement of what was corrected and why; and a statement that the officer adopts the corrected document as their sworn account. That language, included in every AI-assisted report, is a custody-chain statement. It identifies the officer as the human who closes the chain from AI output to sworn document.

Key Takeaways

• Chain of custody extends through every system that touches evidentiary material, including AI systems that ingest recordings to produce transcripts and drafts. An AI ingestion step with no custody log is a chain-of-custody gap, and a gap breaks the chain regardless of whether the underlying evidence was properly handled.
• AI processing creates four specific chain-of-custody risks: unauthorized or unlogged ingestion, transformation without documentation, intermediate artifact loss, and a data integrity gap between input authentication and output accuracy.
• The five custody-log requirements for every AI processing step are: input identifier with file hash, system identifier with model version, authorization record tied to a specific authenticated user, output identifier linked to the input, and review record documenting the human verification pass.
• CJIS obligations stay with the agency. A vendor's CJIS certification covers the platform's capability; the agency's implementation determines whether that capability is used in a compliant way. Automatic AI ingestion of criminal justice information without a logged, attributed access record is not CJIS-compliant regardless of the vendor's certification status.
• Broken chains in AI-processed evidence are often systemic rather than incidental: not a failure to follow a procedure but the absence of a procedure covering the AI step. A systemic gap is more damaging than an individual error because it raises questions about every AI-processed piece of evidence the agency has produced.
• Custody requirements must be written into vendor contracts before deployment. Once a multi-year sole-vendor contract is signed, the leverage to require specific custody features is largely gone. Preservation of intermediate artifacts, comprehensive audit logs, and export capabilities for full custody-chain reconstruction should be explicit contract terms.
• The officer's role in the custody chain is the review-and-adoption step: a footage-grounded verification pass with documented corrections, attributed to the specific officer's authenticated identity, with an explicit adoption statement. This is the human link that closes the chain from AI output to sworn document.
• The King County example demonstrates that admissibility of AI-assisted reports requires a documented chain. An agency that cannot produce the chain from original evidence through every AI processing step to final report is in the same position as an agency with no chain: the records are of unknown provenance and subject to challenge or exclusion.