Policies and training on the use of Generative AI Tools

Response

This request relates specifically to generative AI / large language model tools. It does not cover other forms of AI such as facial recognition technology, predictive analytics, or automated data matching, which are outside the scope of this request.

I request disclosure of the following recorded information:

1. Policies and governance
(a) Copies of all policies, standard operating procedures, guidance documents, or directives ‘ whether issued at force level or adopted from national guidance ‘ that govern the use of generative AI tools by officers or staff, in force since 1 January 2023.
(b) If no such policy or guidance exists, please confirm this.

 

All disclosure documents provided in response to your request have been combined into a single PDF file. This includes the following policy and guidance documents held in relation to this question, listed by page number within the PDF:

 

• Disclosure 1 – Pages 1 -14 – Microsoft Copilot Chat Acceptable Use Policy – January 2026

 

• Disclosure 2 – Pages 15 – 17 – Using Artificial Intelligence (AI) – intranet guidance – published 06/08/2025

 

• Disclosure 3 – Pages 18 – 20 – Using Copilot Chat – intranet guidance – published 06/08/2025

 

Our intranet guidance also contains a link to an NPCC document titled Responsible AI Checklist for Policing, available via the link provided.

 

We also have a draft Artificial Intelligence Policy, which I have not considered to be within the scope of your request as it is not a finalised policy.

 

Within the documents provided for this question and for question 2, names have been redacted under the following exemption:

 

Section 40(2) – Personal Information.

 

Section 40 is a class based absolute exemption and there is no requirement to consider the public interest in this case. To disclose the requested information would identify individuals, and information that relates to them which would breach principle 1 (lawfulness, fairness, and transparency) of the Data Protection Act 2018.

 

Within the document, ‘Using Artificial Intelligence (AI)’, further redactions have been made to withhold the list of generative AI tools used by Avon and Somerset Police. These redactions have been made under Sections 24 and 31. Please see our response to question 3 below, for further explanation including the harm and public interest test.

 

 

2. Risk assessments and compliance
   (a) Copies of any Data Protection Impact Assessment (DPIA) completed in relation to the use of generative AI tools by officers or staff for any purpose, including administrative, investigative, or operational tasks.
   (b) If no DPIA has been completed for any generative AI tool, please confirm this.

 

Please see attached the following DPIA held in relation to this question:

 

• Disclosure 4 – Pages 21 – 46 – DPIA for the Proof of value of Co-Pilot

 

As explained above in our response to question 1, names have been redacted under Section 40(2) relating to personal information.

 

3. Deployed tools and licensing
   (a) A list of any generative AI tools (including products bundled within existing software suites, such as Microsoft Copilot within Microsoft 365) that have been made available to officers or staff, and the date from which each tool was made available.

 

Avon and Somerset Police make use of generative AI functionality within existing corporate systems used to support efficiency or administrative tasks. A complete list of tools used has been withheld – with relevant redactions made where the documents provided in question 1 list such tools – under the following exemptions:

 

• Section 24(1) – National Security

• Section 31(1) – Law Enforcement

 

Harm (Sections 24(1) and 31(1))

As sections 24 and 31 are qualified and prejudice based exemptions there is a requirement to evidence the harm in confirming or denying that information is held and also to consider the public interest.

Harm (Sections 24 and 31):

While we have confirmed that generative AI functionality is used within corporate systems for administrative and efficiency purposes, we are withholding a detailed list of specific tools.

Disclosing a complete list of generative AI tools would reveal information about the force’s technological environment and capabilities. Even where tools are used for routine purposes, their identification could enable inferences about their potential use in more sensitive areas, including counter-terrorism, serious and organised crime, covert intelligence, or digital forensics.

 

This information could be exploited by hostile actors to identify vulnerabilities, understand system capabilities, and adapt their behaviour to evade detection. In addition, disclosure could contribute to a wider “mosaic effect” when combined with information from other forces, undermining national security and law enforcement effectiveness.

 

Disclosure would therefore be likely to prejudice national security (section 24(1)) and the prevention and detection of crime (section 31(1)).

 

Public Interest Test:

There is a public interest in transparency about the use of AI in policing, which supports accountability and public confidence.

However, there is a stronger public interest in protecting national security and law enforcement capabilities. Disclosure of specific tools would provide useful intelligence to those seeking to undermine policing activity, enabling them to exploit weaknesses or avoid detection.

Balance Test:

The public interest in transparency is outweighed by the risk of harm. Avon and Somerset Police therefore considers that sections 24(1) and 31(1) apply, and the requested information is withheld.

 

(b) Whether access to any generative AI tool has been suspended or restricted since 1 January 2023, and if so, which tool, on what date, and any recorded reason for the suspension or restriction.

 

No information is held in relation to this question.

 

4. Training
   (a) Whether any training, e-learning, briefing, or written guidance on the use, limitations, or risks of generative AI tools has been provided to officers or staff, and if so, the date it was first made available.

 

I can confirm that training, e-learning, briefings and written guidance on the use of AI tools, including their limitations and risks, have been provided to officers and staff. The date these were first made available is not recorded. However, we have sought to assist by providing publication dates for the current guidance and policy documents listed above in relation to question 1.

 

(b) Whether completion of such training is mandatory before officers or staff may use generative AI tools, or whether it is optional or advisory.

 

None of the training identified has been mandatory. However, the expectations and requirements of users are set out in the acceptable use policy and guidance documents provided above.

 

(c) The number of officers and staff who have completed such training, if this figure is recorded, and the total number of officers and staff who currently have access to one or more generative AI tools.

 

This information is not recorded.

 

(d) If no training has been provided, please confirm this.

 

Not applicable

 

 

5. Incident recording and quality assurance
   (a) The number of occasions, if recorded, on which output from a generative AI tool used by officers or staff was subsequently identified as containing factual inaccuracies, fabricated information, or material errors ‘ whether identified internally or by an external party ‘ since 1 January 2023.

This information is not held. Avon and Somerset Police does not hold a centrally maintained record of the number of occasions on which output from a generative AI tool was subsequently identified as containing factual inaccuracies, fabricated information, or material errors.

 

The force does not currently operate a central monitoring or assurance mechanism capable of systematically identifying, validating, or quantifying all such instances across user activity.

 

(b) If no such records are held, please confirm this.

 

No centrally maintained records are held for the purposes described above.

 

Force guidance relating to the use of Microsoft Copilot and generative AI tools requires users to verify outputs and not rely on generated content without appropriate human review. Such tools must be used to support professional judgement, not replace it.

 

 

6. Audits and reviews
   (a) Copies of any audit, review, or evaluation ‘ whether conducted internally or by an external body ‘ of the force’s use of generative AI tools.

 

No information is held in relation to this question. No audit, review or evaluation has been conducted internally or externally that focuses specifically on Avon and Somerset Police’s use of generative AI tools. However, in relation to the use of generative AI tools in policing more generally, relevant information is available via the following links:

 

NPCC AI Covenant

 

Responsible AI Checklist

 

(b) If no audit or review has been conducted, please confirm this.
Please see our response above confirming this.

 

In addition to our responses above, I can neither confirm nor deny (NCND) whether any additional information is held in relation to questions 3 – 6, by virtue of the following exemptions:

• Section 24(2) relating to National Security

• Section 31(3) relating to Law Enforcement

As sections 24 and 31 are qualified and prejudice based exemptions there is a requirement to evidence the harm in confirming or denying that information is held and also to consider the public interest.

Harm (Sections 24 and 31):

 

Having disclosed high‑level information about routine, non‑sensitive uses of Artificial Intelligence (AI), Avon and Somerset Police will neither confirm nor deny whether it holds any further information that would indicate whether AI is, or is not, used in connection with counter‑terrorism, serious and organised crime, covert intelligence activity, or specialist digital forensics supporting those functions. The duty to confirm or deny is disapplied by virtue of sections 24(2) (national security) and 31(3) (law enforcement) of the Freedom of Information Act 2000. This response should not be taken as an indication of whether any additional information is held.

 

Confirming or denying whether AI is used by Avon and Somerset Police in counter‑terrorism (CT), serious and organised crime (SoC), covert intelligence or specialist digital forensics would enable hostile actors to infer the presence, absence or maturity of particular policing capabilities. This type of capability insight is of practical value to adversaries who routinely adapt their behaviour to exploit perceived weaknesses, overwhelm less technologically equipped forces, or adjust methods to avoid detection. Even a single confirmation or denial could be combined with publicly available information from other forces to build a detailed national picture of capability. This mosaic effect represents a realistic and actionable risk, providing information that could materially undermine national security and the effectiveness of law‑enforcement operations.

 

Public Interest Test – Section 24(2) (National Security)

Factors in favour of confirming or denying:

 

There is a strong public interest in openness and transparency about the use of AI in policing. Confirming or denying the existence of information could contribute to informed public debate about how emerging technologies are deployed, the safeguards in place, and whether they align with legal and ethical standards. Disclosure can enhance accountability, promote trust, and reassure the public that AI use is subject to appropriate oversight.

 

Factors against confirming or denying:

There is an overriding public interest in protecting national security. Confirming or denying the existence of information about AI deployment in CT, SoC, covert intelligence, or linked specialist forensics would provide adversaries with useful insights into operational capabilities and priorities. Even a simple confirmation or denial could enable hostile actors to adjust methods, exploit potential gaps, or test thresholds. The resulting risk to national security would be real and more than trivial, and is not mitigated by partial or local confirmations, because the cumulative mosaic effect across multiple forces would still disclose sensitive capability at a national level.

 

Public Interest Test – Section 31(3) (Law Enforcement):

Factors in favour of confirming or denying:

 

There is a general public interest in understanding how AI is used to support policing functions, including the efficiency and effectiveness of core processes, and in ensuring that novel technologies are used lawfully and proportionately. Confirmation or denial could support external scrutiny and promote confidence in policing practice.

 

Factors against confirming or denying:

 

Confirming or denying the existence of information relating to AI use in CT, SoC, covert intelligence, or specialist digital forensics would be likely to prejudice the prevention and detection of crime, the apprehension or prosecution of offenders, and the operation of law enforcement tactics and methodologies. Knowledge of whether specific capabilities exist, or do not exist, enables offenders to adapt behaviour, exploit capability gaps, frustrate investigations, and increase the cost and complexity of counter‑measures. The public interest strongly favours maintaining an NCND stance to preserve the effectiveness of law enforcement and the safety of the public.

 

Balance Test:

For the reasons set out above, Avon and Somerset Police will neither confirm nor deny whether it holds information indicating AI is used in CT, SoC, covert intelligence activity, or specialist digital forensics supporting those functions.

 

The public interest in maintaining the NCND position under sections 24(2) and 31(3) outweighs the public interest in confirmation or denial. This response should not be taken as an indication of whether any relevant information is held.

 

A separate response has already provided high‑level information about routine, non‑sensitive AI use. To go further would risk disclosing sensitive capability by inference.