Welcome: let's start with an honest picture
Hello, and welcome to the course! Before we get into prompts and workflows, we need a clear picture of what an AI assistant actually is and what it does well. We also need to be just as clear about where it lets you down.
You've probably heard both stories already. One says AI will give you back hours every week. The other says it will make dangerous mistakes in people's records. The truth is that both can happen. Let's look at why.
How does a large language model actually write?
Tools like ChatGPT, Gemini, Claude and Microsoft 365 Copilot run on something called a large language model, or LLM.
Here's the plain-words version. An LLM has been trained on a huge amount of text. From that it learned patterns: which words tend to follow which, in which kinds of writing. When you give it a prompt, it builds a reply one small piece at a time. At each step it predicts what is most likely to come next.
Think of a colleague who has read thousands of case notes and can write in that style without effort. Now imagine they weren't in the room for your visit. They can write something that sounds exactly like a good case note. But sounding right and being right are two different things.
That's the key point. The model is built to produce fluent, likely text. It is not checking that text against what really happened. So it can sound calm and confident and still be wrong. When it invents a detail that was never there, we call that a hallucination.
Why does this matter so much in social work? Because a made-up detail in a record isn't just a typo. It can follow a person for years.
Sorting the work: help, partly help, stay human
So what can it do for you? It helps to sort tasks into three groups.
Tasks AI can help with. These are jobs about language and shape, where you can easily check the result. Think of tidying your own rough notes into clear paragraphs, suggesting a structure for a report, or turning a jargon-heavy paragraph into plainer English.
Tasks AI can partly help with. Here AI gives you a first draft or a starting point, and you do the real work of checking and correcting. A summary of a recorded visit is a good example. So is a first draft of a letter to a family, or a summary of a piece of practice guidance. The tool saves you time, but every line still needs your eyes.
Tasks that must stay fully human. Some things belong to you alone: deciding whether someone is at risk, making a safeguarding judgement, building a relationship of trust, weighing up what a person really wants, and signing your name to a professional opinion. AI has no professional accountability. You do.
Here's a quick test you can use. Ask yourself: "If this output were wrong, would I spot it, and who would be harmed if I didn't?" The harder it is to check, and the bigger the harm, the more the task belongs in the human column.
What is the profession actually doing?
Let's look at real, dated evidence. When you read any claim about AI, ask three things: who said it, when, and what exactly they measured.
The regulator's research (January 2026). Social Work England commissioned Research in Practice to study how AI is being used in social work in England. The research was carried out in spring 2025 and published in January 2026. It found that generative AI is now the most common form of AI in social work, used for note-taking, writing documents and transcribing meetings. More than eight in ten respondents thought AI could reduce admin. But 86% of social workers who had qualified in the past five years said their training gave them no specific preparation for using AI.
What does that show? It shows how people feel about AI and how widely it's used. It doesn't prove that AI improves outcomes for the people you work with.
Transcription tools in councils (late 2025). Magic Notes, a tool made by the company Beam, turns recordings of meetings into case notes and summaries. In December 2025 the company said more than 100 local authorities and 50,000 frontline practitioners were using it. It also reported that, in Kent County Council's evaluation, staff saved an average of seven hours a week. Notice the source, though. These are the company's own figures, and they measure time, not accuracy.
Independent research (February 2026). The Ada Lovelace Institute published a report called Scribe and Prejudice? in February 2026. Its researchers interviewed 39 social workers across 17 local authorities. Almost all said the tools brought real benefits. But the report also found risks such as bias and hallucinations that weren't being fully assessed. In one example a summary said a person had expressed suicidal ideation when they never had. The report also found that councils mostly evaluated these tools on efficiency, not on their impact on people receiving care.
See how the pieces fit? Time saved is real for many practitioners. But time saved isn't the same thing as accurate, fair records. That gap is exactly where your judgement comes in.
Where this leaves you
So is AI a helper or a hazard? Honestly, it's both. It depends on the task and on how carefully you check it.
Throughout this course we'll take real tasks from your job and work through them together. Each time, you'll keep the judgement and the AI does some of the typing.
Key takeaways
- An LLM writes by predicting likely next words. It aims for fluent text, not checked facts, so it can be confident and wrong.
- Sort tasks into three groups: AI can help, AI can partly help with your checking, and tasks that must stay fully human.
- Risk judgements, safeguarding decisions, relationships and professional opinions stay with you.
- For any evidence, ask who said it, when, and what they measured.
- Dated evidence from 2025–26 shows wide use and real time savings, alongside documented hallucinations and weak evaluation of the impact on people.