Six Things AI Must Never Do in an Investigation

Most questions about AI in investigative work are questions of local policy. Which tool, which classification, which approval. Those answers differ from organisation to organisation and your own policy governs.

These six do not differ. They hold regardless of platform, licence or local policy, because they go to the integrity of the investigation itself rather than to how your organisation has chosen to configure its software.

1. Write your RIPA reasoning

A necessity and proportionality rationale must be the applicant’s own reasoning, and the authorising officer’s own reasoning. That is the point of it. Generated text is not reasoning, it is text that resembles reasoning, which is a different thing entirely and will be treated as such if it is ever examined.

If you cannot articulate why the interference is necessary and why it is proportionate in your own words, you do not yet have an authorisation worth granting. (We covered declaring AI use in an application in Must You Declare AI Use in a RIPA Application?)

2. Decide disclosure

Under section 7A of the Criminal Procedure and Investigations Act 1996 the duty of disclosure is continuing and under the Attorney General’s Guidelines on Disclosure 2024 the Disclosure Officer must apply the statutory disclosure test. Neither of those is currently delegable to a tool.

A tool can help you find material. It can help you summarise material. It cannot apply the test, and it cannot hold the continuing duty. Those sit with a named person.

3. Author a witness statement

A statement must be the witness’s own words, signed as true. AI-smoothed prose assists one person only, and that person is the cross-examiner. The moment a statement reads better than the witness speaks, you have handed over a line of questioning for free. The same caution applies to witnesses preparing to give evidence see Using AI to Prepare to Give Evidence.

4. Monitor or profile an individual

Systematic use of AI to build a picture of a person or their lifestyle may amount to directed surveillance. The tool will not alert you to this. There is no warning, no threshold notification, nothing on the screen that says you have crossed from research into covert activity.

This is the one that catches people out, because each individual search feels trivial. It is the accumulation and the systematic character that matters, and only you can see it unless a tool is speciically authorised under a Directed Surveillance Authority and the AO is fully sighted on it's capability.

5. Serve as a source of fact

Nothing an AI tool asserts is evidence of anything. It indicates where material may be; the material is the evidence. Every output goes back to the source document, and the source document is what you rely on and what you exhibit.

6. Replace your decision log rationale

The log records why you decided what you decided. That is its entire function. Populating it with generated text defeats the only purpose it has.

Where the Government is heading

None of this means AI is being pushed out of investigative work. The direction is the opposite.

In July 2026 the Government set out reforms bringing AI into the criminal disclosure process, accepting a recommendation to establish a national governance forum for disclosure technology, bringing together policing, the judiciary, prosecutors and government to oversee emerging technologies and ensure safeguards. The stated intention is a framework covering human oversight, accountability, proportionality, legality and safeguards for fair trial rights, with tools that are auditable, evidence-based and subject to bias monitoring.

The Home Office’s policing technology plans announced in January 2026 included £115 million over three years to identify, test and scale AI technologies in policing through Police.AI, the national centre for police AI; the July 2026 disclosure announcement cites £75 million of government backing for that centre in the disclosure context.

The practical test

Before any prompt goes near case material, three questions need settling. What does your organisation’s AI policy actually permit, remembering that corporate approval to deploy a tool is not the same as approval to use it on investigative material. What security classification are your tools approved to handle, and does the material in front of you exceed it. And are investigatory powers engaged, because where AI is used in connection with investigatory powers it falls within the scope of oversight interest, including AI features embedded in commodity software such as Microsoft 365 Copilot.

If the policy is silent on investigations, raise it. Do not work around it.

Free resource — The UK Investigator’s AI Starter Pack. Ten structured prompts for investigative practice, each with the verification step it requires and the legal risk attached to it. Written for investigators, enforcement officers and investigation managers in UK local authorities, the wider public sector and regulated industry.

Email info@bdgtrainingconsultancy.co.uk to request your copy, or visit www.bdgtrainingconsultancy.co.uk.

Sources: Criminal Procedure and Investigations Act 1996, s.7A (legislation.gov.uk); Attorney General’s Guidelines on Disclosure 2024, effective 29 May 2024 (gov.uk); Regulation of Investigatory Powers Act 2000, Part II and the Covert Surveillance and Property Interference Code of Practice; GOV.UK, “AI to speed up justice under major disclosure reforms”, 14 July 2026; Home Office policing technology announcement, January 2026.

This article is a training and awareness summary and does not constitute legal advice — your own organisation’s policy takes precedence.