Lesson 1: What AI Assistants Can and Cannot Do in Your Role

Why This Lesson Comes First

AI assistants such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot are already in use across UK policing, fire, ambulance and control rooms. Before you use one at work, you need to know how it produces its answers. If you do not, you cannot judge when to trust it. This lesson covers three things: how these tools work, which of your tasks they suit, and what UK emergency services have actually done with them.

Step 1: Understand How an AI Assistant Produces Text

An AI assistant is built on a large language model. Here is how it works:

1. It was trained on a very large amount of text, from which it learned patterns of how words follow one another.
2. When you type a prompt, it predicts the most likely next word, then the next, and keeps going until it has an answer.
3. It does not look facts up in a verified database unless the tool is connected to search or to your documents. Even when it is, it can still misread or misquote what it finds.
4. It has no idea whether a sentence is true. It only knows whether the sentence sounds like the kind of text that usually follows your question.

Analogy: Think of a probationer who has read every report in the station but never attended a single job. They can write something that sounds exactly like a real report. They can also describe an incident that never happened, in the same confident tone.

When an AI states something false as if it were fact, this is called a hallucination. Its language stays fluent and assured whether the content is right or wrong. Confidence tells you nothing about accuracy.

DO: Next time you read AI output, ask yourself: "How would I check this sentence against a source?"

Step 2: Learn From Real, Dated UK Examples

Example 1: A made-up fact reached a high-stakes decision (West Midlands Police, 2025–26)

• In October 2025, West Midlands Police advised that Maccabi Tel Aviv away fans should be barred from the club's Europa League match at Aston Villa on 6 November 2025.
• The material behind that advice included a reference to a Maccabi Tel Aviv v West Ham match that never took place.
• In January 2026, Chief Constable Craig Guildford apologised and said the error came from the use of Microsoft Copilot. He had earlier told MPs that AI had not been used.
• The Home Secretary called the episode a "failure of leadership" and said she no longer had confidence in him.

What it shows: A hallucination looks like a normal fact. If nobody checks it against a source, it can feed into decisions that affect the public, and it can damage trust in the force and its leadership.

Example 2: AI drafts, a clinician checks (South East Coast Ambulance Service, December 2025)

• In December 2025, SECAmb announced a pilot of an ambient voice transcription tool (Tortus AI) in its Emergency Operations Centres. The pilot was due to run until February 2026.
• The tool listens to clinicians' calls with patients and turns the conversation into structured medical notes.
• Clinicians must check and approve the notes before they are finalised.

What it shows: Used well, AI cuts down admin while a professional stays responsible for the final record.

Example 3: National investment (Police.AI, June 2026)

• On 10 June 2026 the Home Office launched Police.AI, a national centre for AI in policing in England and Wales, with £75 million over three years.
• Its early priorities include help with triaging, disclosing and summarising digital evidence, analysing CCTV and digital media, and transcribing and routing calls.

What it shows: AI use is growing nationally. The focus is on cutting administrative work, not on replacing professional judgement.

Step 3: Sort Your Own Tasks Into Three Groups

Use these three groups for any task you are thinking of handing to an AI.

Suitable for AI support (low risk, and you check the result)

• Rewording a routine internal email or rota note
• Turning your own rough bullet points into a clearer layout
• Making practice quiz questions from guidance you have already read

Needs extra care (useful, but every fact must be checked against the source)

• Summarising policy, guidance or legislation
• Drafting a public safety message
• Translating information for a member of the public
• Structuring a briefing or shift handover

Never for AI (the task needs your own account, judgement or protected data)

• Writing your witness statement or evidential account
• Making operational decisions about people, such as risk, triage, arrest or deployment
• Pasting in personal, sensitive or case data unless your force or service has approved the tool for it
• Treating AI output as intelligence or fact without checking it

Later levels cover approved tools, data protection and evidence rules in detail. For now, remember this: if a task touches evidence, personal data or a decision about a person, slow down.

Step 4: Try It on Your Own Work

1. List five tasks from your last two shifts.
2. Put each one into one of the three groups above.
3. For each "needs extra care" task, write down the source you would check the AI's output against.
4. Keep the list. You will use it again in later levels.

Checklist

• I can explain that AI predicts likely words; it does not check facts
• I treat confident wording as a style, not as proof
• I have sorted five of my own tasks into the three groups
• I can name two dated UK examples and say what each shows

Key Takeaways

• AI assistants produce text by predicting likely words. They can be fluent, confident and wrong.
• A hallucination looks exactly like a real fact. The only safeguard is checking it against a source.
• West Midlands Police (2025–26) shows what happens when an unchecked AI error reaches a decision. SECAmb (2025) shows AI drafting with human approval.
• Your statements, your decisions about people and protected data stay with you.