Using AI

AI tools can be incredibly useful. With many different types and categories, and functions they can perform, they’re evolving all the time.

The most commonly used AI tools are “generative” (they “generate” content such as text, images, videos, music, etc) and the majority of those are an LLM.

Prompts example Hallucinations other AI tools links

LLM

A Large Language Model is designed to simulate human conversation. Using neural networks (machines loosely modelled on neurons in the brain), they’re trained on massive amounts of data to understand, generate, and manipulate human language. Examples include:

Gemini (by Google)

ChatGPT (by OpenAI)

Claude (by Anthropic)

Meta (by Meta)

Grok (by xAI)

These all work in the same way and have continuously evolving strengths. Many other AI tools, with nice interfaces and additional features, are built on top of an LLM.

Prompts

A “prompt” is how you communicate with AI. Learning how to prompt is essential to get useful responses and there are simple guidelines for a well-crafted prompt.

Talk to an LLM like it’s a conversation as they’re designed and trained to recognise speech patterns.

aim - what you want it to do e.g:

create a text response, document, image or video

explain a problem, soution or concept

brainstorm an idea

context (this is critical) - give relevant background and information e.g:

who are you (so it knows how to talk to you)

who’s it for (so it knows what level to pitch the response)

why you want it (which can change the style of the response)

use examples (so it has information to work with)

rules - set limits, format and style e.g:

100 words, 3 pages or as long as it takes

text only, include a table, just bullet points or in the style of a blog

pupil language, formal, friendly or provide a example so it can copy that style or tell it what role to play (teacher, consultant, etc)

These don’t have to be in order but be specific - if the prompt is vague, AI will have to guess how to respond. AI is not always predictable so the more specific you are with your prompt the more likely you are to get a useful and accurate response.

Many AI tools allow you to upload documents, images and other files to help with context, format, language, style or output, or to provide information for it to read.

refine - modify the response

always check the response (see Hallucination below)

if it's not perfect give AI some clear, new or updated criteria to refine it

Bad prompt:

What is a dog?

unclear aim, no context and no rules - anything could happen!

Unpredictable response:

Canis lupus familiaris is a paedomorphic carnivoran characterized by extreme phenotypic plasticity and specialized interspecific socio-cognitive adaptations. Driven by anthropocentric selection, their behavioral ecology features a truncated predatory sequence and unique genetic hypersociability (specifically via GTF2I variants), facilitating the decoding of human deictic gestures and maintaining a prolonged ontogenetic window for primary socialization.

as a KS1 teacher this is unlikely to be the response you wanted. The prompt was so vague, AI had to do a lot of guessing and it looks like it assumed you’re a zoology professor!

Better prompt:

Write a 50 word paragraph explaining what a dog is. As a key stage 1 teacher I need the language to be appropriate to year 1 pupils. The title is “What is a dog?”

clear aim, explicit context and specific rules

More useful response:

A dog is a furry animal with four legs and a tail. Dogs are part of a family of animals that includes wolves. They are clever pets that love to play, sniff, and bark. People call them “man’s best friend” because they are so friendly and loyal.

a much more predictable and useful response as AI had clear guidelines

remember: if it's not perfect ask AI to refine it with your updated criteria

“Hallucinations”

This is the term used when AI makes something up (which happens more often than you might hope) - and it’s usually done with confidence, making it seem factual. Never assume responses are totally accurate - always double-check.

Clear instructions and constraints (like specifying “if you don’t know, say I don’t know”) can significantly reduce false or made-up information.