Mississippi Artificial Intelligence Network
AI Policy Supplement for Law Enforcement
Reports, transcription and records
Decide what a tool may produce, how people will check it, and which records the agency must be able to preserve and retrieve. Use alongside MAIN’s Local Government AI Policy and Guidance Template.
- What this is
- Adaptable planning guidance with sample clauses, review checks and a blank approval record.
- Who it serves
- Mississippi police and sheriff’s offices, reporting personnel, supervisors, records staff, security owners, counsel and prosecutors.
- Start here
- Choose one documentation task. Resolve its authority, data, verification and records plan before permitting use.
01 · Baseline
Start with the agency’s governing rules
Use the Local Government AI Policy and Guidance Template for general governance, procurement, personnel, public records, accessibility, security and human authority. Its public-safety section supplies the baseline. This supplement adds decisions for AI-assisted documentation, including AI features embedded in existing software. [1]
Apply it first to a defined report, transcription or summarization task. A campus or state agency may find the questions useful, but its own institutional authority and adopted policies govern. The relevant MAIN guide remains a planning reference.
- Official requirements
- Applicable law, court rules and orders, records schedules, security requirements, agreements and adopted agency policies control. Determine their applicability to the actual data, system, proceeding and agency.
- Voluntary guidance
- NIST AI guidance and professional policing resources inform the review questions. They do not approve a product or agency use. [2] [8]
- MAIN recommendations
- The proposed boundaries, clauses, checklist, fictional cases and approval record below are MAIN’s synthesis for local adaptation.
Scope boundary: this supplement provides no authority for identification, surveillance, predictive policing, forensic interpretation, warrants, arrests, detention, use of force or emergency-dispatch decisions. It supplies no evidence-admissibility determination or equipment-operating procedure.
02 · Permission
Define the exact task and permission
A tool may perform more than one function. Record which functions the agency is considering and what each can change. None of these categories creates automatic approval.
- Writing assistance
- Spelling, grammar or formatting changes to existing text. Check whether the tool transmits information and whether a change alters meaning.
- Transcription
- A text rendering of recorded speech. Treat speaker labels, wording, timestamps and uncertain passages as items to verify against the recording.
- Summarization
- A shortened account of selected material. Identify its intended purpose and what information may be omitted.
- Narrative generation
- A newly composed draft assembled from inputs. Require a separate decision about eligible documents, source material and review before it enters an official workflow.
Set allowed document types and exclusions before testing. Distinguish routine administrative writing from case-related reports, interview material and statements. Route any proposed affidavit, sworn statement, charging document or other consequential submission for separate legal and supervisory review; general report-writing permission does not include it. FPI’s model policy offers professional reference points for restrictions, supervision and disclosure. [8]
- Agency approver: accepts the defined use and its conditions.
- Reporting person and supervisor: verify the content and document unresolved discrepancies.
- Records/evidence owner: establishes preservation, access, retrieval and correction processes.
- Security owner and appropriate CJIS authority: review the actual data and environment.
- Prosecutor or agency counsel: advises on case use, disclosure, holds and legal limits.
Record next: one accountable owner, the exact function, permitted documents, exclusions, reviewers and stop authority.
03 · Data and vendors
Resolve the data path before uploading
Identify the recordings, notes, reports and other information the tool would receive. Confirm authority to use and disclose each source for this purpose; possession alone does not approve vendor processing. Include attachments, background speech, location information, telemetry, linked systems and people who could access them. Removing names alone may leave sensitive or identifying information. NIST’s Generative AI Profile addresses privacy, information integrity and third-party risks. [3]
Applicable security requirements: the FBI’s Criminal Justice Information Services (CJIS) Security Policy v6.1, dated June 25, 2026, applies to entities with access to criminal justice information (CJI) or operating systems used to process, store or transmit CJI. Have the agency security owner and appropriate CJIS authority determine the requirements for this data flow. Sensitive agency information may need protection even when it is not CJI. A vendor’s compliance claim does not constitute agency authorization. [4]
Vendor questions: retain evidence for each answer
- Where does information go? Document the selected account, storage, connected systems, subprocessors, support access and remote access.
- How may it be reused? Check contracts and settings for training, human review, secondary uses and sharing.
- Who can act? Identify who may upload, view, edit, export, delete or automatically submit material; confirm how access ends.
- What can be retrieved? Demonstrate an authorized export of the material and history identified in the records plan, including after a contract ends.
- What changes or fails? Check feature/model change notices, incident contacts, outage procedures and deletion commitments.
- What proves suitability? Retain applicable agreements, security-review findings, a data-flow description, an export example and unresolved limitations.
These questions are MAIN’s adaptation of NIST and applicable CJIS review considerations, not a federal checklist or product certification. Use current vendor documentation for exact behavior. [3] [4]
04 · Verification
Verify each output against its sources
Generated language can be plausible and wrong. A transcript can omit speech or misattribute a speaker, and a summary may leave out a qualification that changes meaning. NIST and IACP identify confabulation, information-integrity risks and over-reliance as reasons for contextual evaluation and human review. [3] [9]
Report-review checklist
- Confirm the input. Check case/document identifiers, source files, recording coverage and gaps. Establish what the tool could actually observe or process.
- Check the facts. Verify names, dates, times, locations, quantities, sequence and speaker attribution against appropriate sources.
- Check meaning and omissions. Look for lost negatives, qualifications, conflicting accounts, unsupported characterizations, uncertainty and facts the draft leaves out.
- Separate types of information. Distinguish the reporting person’s observations, another person’s account, inferences and generated wording. Do not present generated text as a verbatim quotation or an observation without establishing its basis.
- Resolve uncertainty. Mark unclear audio or missing information through the agency’s process. Seek qualified language or accessibility assistance when needed; do not invent words to complete a passage.
- Document the review. Correct discrepancies, preserve the required history, identify reviewer and supervisor, and route unresolved issues before submission.
Keep the boundary: an AI transcript or narrative does not replace the original source or the agency’s established evidence and reporting procedures. Review completion does not itself establish admissibility.
05 · Records and handoff
Make the record explainable and retrievable
Before approving the use, identify which originals, generated versions, prompts or settings, edits, review records and logs need preservation. Record where each will be stored, who can retrieve it, the authority for its retention and disposal, and how holds are communicated. Protect sensitive material in the approved system rather than duplicating it into a public worksheet.
Mississippi records context: MDAH’s Local Government Records Office publishes retention schedules and guidance and can assist with unlisted records. Determine the schedule and authority that apply to this agency and record series. Do not assume every draft is disposable, that a vendor’s deletion default is an approved schedule, or that one retention period applies to all recordings and reports. [5]
Keep the connection between the source material, the generated item, material corrections and final approved work. Record relevant tool/configuration information available to the agency and the people who prepared and reviewed the item. Verify retrieval with an authorized test export; keep originals intact under the agency’s evidence process. This is MAIN’s proposed documentation practice, not a declaration that every intermediate item has the same legal status.
- Internal disclosure
- Tell the reviewing supervisor which tool and function were used, the source material and any material limitation or correction. Follow the agency’s approved labeling process.
- Criminal discovery
- Route AI involvement and relevant records questions to the prosecutor or designated counsel. The court’s posted Mississippi Rule 17 addresses disclosure and discovery; its provisions distinguish court and case settings. Counsel confirms the current rules, applicable duties and orders, and the appropriate handoff. [6]
- Public-records release
- Route requests through the records custodian and counsel under the agency’s public-records process. A discovery decision, an internal label or a vendor setting does not by itself determine public access. [1]
This supplement sets no universal AI disclosure label, retention period or discovery deadline. An internal recommendation to document AI use is distinct from a legally required disclosure. Confirm the current rules, relevant court orders and agency-specific requirements with counsel.
06 · Training and response
Test the review process before relying on it
Begin with fictional or otherwise authorized test material. Train staff to find inaccurate speech, missing context, unsupported conclusions and altered meaning. Include realistic noise, multiple speakers, difficult passages and language needs relevant to the proposed task. A product demonstration does not establish local suitability. NIST recommends context-specific evaluation and monitoring. [2] [3]
Set acceptance criteria before testing: error types and consequences, omissions, correction effort, total review time, retrieval success and reviewer workload. Record failures as well as useful results. Define how staff can complete the work using the approved non-AI process when the tool or a required reviewer is unavailable.
For an unauthorized upload, incorrect submission or other suspected incident, follow the agency’s incident and correction procedures. Notify the designated owner, preserve relevant evidence safely, assess affected records and recipients, and route legal and security decisions to the responsible people. NIST SP 800-61 Rev. 3 supports incident preparation, response and recovery; it does not impose a universal reporting deadline for this supplement. [7]
Reassess after a material change in the model, features, vendor terms, data access, document type, reviewer capacity or observed errors. Set a calendar review date as well. These are recommended review triggers. MAIN’s AI courses can support general literacy; completion does not authorize agency or case use.
07 · Local decision
Document the local decision
Walk through one proposed use with the people responsible for its data, content and records. The decision may be to approve a defined use, revise it, hold it or decline it. An incomplete record is a request for review, not permission.
Blank approval record
Copy this record into the agency’s approved system. Use references to protected case material rather than pasting sensitive content into an unapproved tool or public form.
- Use and exclusions
- [Agency / task / transcription, summary, narrative or writing assistance / eligible documents / excluded documents, cases and actions]
- People and authority
- [Accountable owner / agency approver / reporting person / supervisor / records owner / security and appropriate CJIS contacts / prosecutor or counsel / backup / stop authority]
- Tool and data path
- [Provider / product, available version and features / approved account / integrations / data categories / recipients / reuse and training terms / permissions]
- Requirements and evidence
- [Applicable security, records, court and agency requirements / source and date checked / responsible reviewer / contracts and review findings / unresolved questions]
- Verification and test results
- [Source material / review steps / test conditions / consequential errors and omissions / correction and total review time / acceptance criteria / limitations]
- Preservation and handoff
- [Material to preserve / storage and access / links between originals, drafts, corrections and final work / export test / retention authority and holds / disposal owner / internal, prosecutor and public-records routes]
- Fallback and correction
- [Stop conditions / reporting route / approved non-AI process / affected-record review / restoration evidence and authority]
- Decision and review
- [Approve defined use / revise and resubmit / hold / decline] · [Rationale / conditions / approver / date / next review date and change triggers]
Ready for the agency’s decision?
- Staff can explain exactly what the tool may produce and which documents are excluded.
- Data access, material vendor terms and required security review are resolved.
- The assigned reviewers can check the work against its sources.
- The agency can preserve and retrieve the material its records plan requires.
- The supervisor, records owner and legal contacts know the handoff and correction routes.
- Staff know when to stop and how to complete the work without the tool.
Source notes
Sources and status
Sources checked October 25, 2026. The clauses, examples, checklist and record are MAIN’s proposed adaptation. No source endorses this supplement or approves an agency’s use. Recheck the relevant source before adopting or revising local rules.
- MAIN Local Government AI Policy and Guidance Template. Companion planning baseline; public safety, records, procurement and human authority. It is not adopted local policy.
- NIST AI Risk Management Framework 1.0, January 26, 2023. Final voluntary framework for governance, contextual evaluation and monitoring.
- NIST AI 600-1: Generative AI Profile, July 26, 2024. Final voluntary companion guidance on confabulation, privacy, information integrity, security and third-party risks. Not an evidence-admissibility standard.
- FBI CJIS Security Policy v6.1, June 25, 2026. Section 1.2 identifies scope; section 1.3 addresses local security policy. Applicable controls and implementation schedules require agency/CJIS review. The resource center links the current version. Neither is a blanket certification of AI products.
- MDAH local-government records guidance and retention schedules. Official Mississippi guidance; identify the applicable jurisdiction, record series, approved schedule and other controlling authority. Court records may require a different authority.
- Mississippi Rules of Criminal Procedure, Rule 17. The judiciary’s current rules index links this compilation, labeled effective July 1, 2023. Its 2023–2026 amendment registers, checked October 25, 2026, list no later Rule 17 amendment. Rule 17.1 distinguishes court/case coverage; Rule 17.10 addresses municipal and justice courts. Counsel determines applicable duties and court orders for the actual case.
- NIST SP 800-61 Rev. 3: Incident Response Recommendations, April 3, 2025. Final voluntary guidance supporting incident preparation, response and recovery. Does not supply case-specific correction or disclosure instructions.
- Future Policing Institute resources. Public model generative-AI policy and implementation checklist; professional references for restrictions, supervision, transparency and evaluation. Voluntary models, not Mississippi requirements; review provisions before adopting any of them.
- IACP: Generative AI in Policing. Professional primer discussing confabulation, privacy, bias, over-reliance and information integrity. Supplemental context, not a legal rule or product-performance guarantee.
For organizational assistance, contact MAIN. Mississippi Artificial Intelligence Network is Mississippi’s statewide AI initiative.