Police Records and AI: Where the Line Sits

The demo was polite. A chatbot took a typed complaint in Hindi, classified it as a cognizable offence, filled a form, and offered a button labelled 'register FIR'. The room clapped. A DCP who had spent a decade on a range did not clap. He asked who would stand in the witness box when the button was wrong — when a non-cognizable matter was forced into an FIR, or when a cognizable complaint was quietly dropped as 'not a fit'. The vendor said the officer was in the loop. The loop was a checkbox that defaulted to yes.

This is an opinion, and the opinion is narrow. Police records are not a normal departmental corpus. They contain allegations, medical facts, locations, minors, and the raw material of liberty. An agent in this building is a clerical aid or it is a mistake. There is almost no useful product in the middle, and the middle is where vendors live.

It is written on 17 August 2026. Criminal procedure in India now lives primarily in the Bharatiya Nagarik Suraksha Sanhita, 2023, with related codes. Read the current India Code text and your state police standing orders. This article will not cite a fake section that 'allows AI FIRs'. If you cannot find the permission, you do not have it.

Where the line actually sits

On one side of the line: retrieval of published police circulars and gazette notifications; draft translations of a complainant's statement for an officer to adopt; completeness checks on forms the officer still owns; internal search over circulars the force has declared searchable. Those jobs are boring. They are also how a range saves hours without touching liberty.

On the other side: deciding whether information will be registered; classifying an offence as the legal act rather than as a suggested label; writing a case diary; recommending arrest; scoring a person; facial search against a scrapable gallery; sending an FIR, a medico-legal opinion, or a witness list to a public model API. Those jobs look like 'AI for policing' on a conference slide. They are how a government loses a High Court matter and a generation of trust.

The line is high because the harm is not a delayed subsidy. The harm is a person in a lock-up, or a person who could not get an FIR, or a leaked statement that gets a witness killed. MeitY's November 2025 governance guidelines do not lower that line. A DPIIT certificate does not lower it. A Mumbai region does not lower it.

Jobs a home department can discuss, and jobs this opinion refuses.

Job	This opinion	Why
Retrieve published circulars for a duty officer	Discuss	Low liberty risk if the corpus is published and logged
Draft a typed version of a dictated complaint	Discuss	Officer adopts the text; audio retention is a separate decision
Auto-register or auto-refuse an FIR	Refuse	Statutory function; exclusion and inclusion both destroy rights
Suggest an IPC / BNS section as a sticky note	Only as a non-binding draft the IO can ignore	A default section becomes the case theory
Person risk score or predictive hotspot as enforcement	Refuse as a casual add-on	Different legal and legitimacy design, not a plugin
Public-API summary of a case diary	Refuse	Uncontrolled copies of investigative material

CCTNS is not your lake

State crime records sit in governed applications for a reason. Copying them into a departmental vector store 'so the LLM can see everything' is how you create a second, worse CCTNS with none of the access control and all of the leakage. If NCRB or your state CID has not given you a connector for a purpose, you do not have a connector. Screen-scraping a police application is not integration. It is a breach with a sprint name.

DPDP still matters. Complainants and witnesses are Data Principals. Some processing may sit on a Section 7 function-under-law limb. That is a conversation with counsel, recorded per workflow. It is not a poster in the control room that says 'lawful because police'.

Two ranges after the same home-department workshop

Objections you will hear — and what to do with them

Other forces already use AI for FIRs

Show you the instrument. A press note that a helpdesk chatbot exists is not a permission to auto-register. Copying another state's demo is how bad standing orders spread.

Officers are in the loop

Then put the act in the system of record, with a default of no-register until a named officer acts. A pre-ticked loop is not a loop. It is the model deciding, with a tired human as decoration.

We will fall behind on modern policing

You will fall further behind if the first writ freezes the whole programme. Start with circular retrieval. Earn the right to discuss harder jobs with a governance body that includes the DGP's office, counsel and the DPO, not only the vendor.

Anonymised FIRs can train the model

FIRs are often re-identifiable by place, relationship and rare facts. Training rights on investigative text are a separate, high legal design. They are not a schedule in a clerical-aid contract. Default refuse.

A four-week playbook that stays on the safe side of the line

1. Week 1: write a one-page prohibited list — auto-FIR, person scores, public APIs, shadow copies of case diaries. Get the DGP's office or the competent home-department authority to initial it.
2. Week 2: inventory published circulars. Build retrieval only on that corpus. Prove refused egress.
3. Week 3: if you add dictation-to-draft, keep audio and draft retention on a written schedule. The officer adopts the text.
4. Week 4: tabletop a writ petition. If the only exhibit is a chat, you are not ready to go live even on the clerk.

File note you can paste

Subject: Limits on agentic systems that touch police records.

This department will not implement automatic registration or refusal of information as a police case. No agent shall be the system of record for that act. No case diary, medico-legal packet or witness list shall be sent to a public model API. No shadow copy of CCTNS or equivalent shall be built as a vector store without a written authorisation that names the purpose. Permitted work is retrieval of published circulars and clerical drafts that an officer adopts. This note does not create a new police power.

This note is an internal aid. It is not legal advice.

Prcept AI will walk away from a police file that needs the button. We would rather lose a deal than invent a permission. If that sentence is inconvenient, we are not the vendor for the job — and you may not need a vendor for that job at all.

Minors and medical facts inside the complaint

Complaints are not clean administrative forms. They contain children, sexual assault facts, medical details, and neighbourhood names. A clerical agent that stores full text 'for quality' is building a second crime record with worse access control. Default to the shortest draft the officer needs, with a retention clock, and with a hard ban on using that text as pre-training, eval publication, or a demo in the next state.

If a vendor asks for a corpus of 'real FIRs to tune the classifier', the answer is no unless counsel, the DGP's office and a written ethics path exist — and even then this opinion would still refuse a public-facing classifier. A hackathon pack of redacted FIRs has a way of becoming identifiable. Place and relationship leak. Do not donate that leak to a slide.

The governance body in this cluster should see any police-adjacent workflow before a VLAN is opened. If the body has not met, you do not have a police AI programme. You have a vendor in a range office.

This article is informational field guidance and opinion for Indian public institutions, not legal, criminal-procedure or procurement advice. Confirm against the criminal-procedure law in force, state police standing orders, DPDP, CCTNS / NCRB access rules, and counsel before you file it.