Last week I delivered a session at the Centre for Investigative Journalism Summer School about using generative AI tools such as ChatGPT and Google Gemini for investigations. In the first of a series of posts from the talk, here are my tips on using those tools for idea generation.

Generative AI tools may not be entirely reliable, but that doesn’t mean that they’re not useful. Journalism, after all, is about more than just gathering information: reporters also need to generate story ideas, identify and approach potential sources, plan ahead, write and edit stories and solve a range of technical challenges. All of these are areas where genAI can help.

Writing prompts for idea brainstorming

Quite often we spend too little time on the idea generation stage of reporting, resulting in a narrow range of ideas that can lack originality.

Spending more time – or accelerating the process with the help of genAI – can ensure you have a wider range of options to choose from.

GenAI tools are useful here because they have ingested the contents of millions of webpages, and are able to effectively predict what sequences of words make ‘a story’ (remember that above all, generative AI is a statistical prediction).

But these tools are also stupid. They don’t know who you are or who you’re writing for.

The first rule of prompt writing, then, is to provide context for the ideas that you’re asking genAI to generate. Here’s an example:

You are an [investigative journalist] for a [national newspaper in the UK] targeted at audiences aged [20-40] in social classes [ABC1]. Generate ideas for an [investigation] related to [housing]

All the parts in square brackets can be changed, but they form part of a general structure which is carefully designed. In other words, we are no longer just querying an AI chatbot — we are practising the skill of prompt design.

Prompt design is the first clue that generative AI is not simply a ‘robot’ which will follow our commands: it is better seen instead as an algorithm which will provide results of wildly varying quality depending on the information that is fed into it (compare, for example, a response to the prompt ‘give me ideas for a story about housing‘).

Like all coding, the adage ‘Garbage In Garbage Out’ (GIGO) is important to remember: the algorithm is only as good as its inputs — from its training data, and from you.

CAREful prompt design

Context is just one element of good prompt design. Kate Moran‘s CAREful structure acts as a mnemonic to remember four:

• Context: describe the situation (for example your role, employer, audience, any events that are important to factor in)
• Ask: request specific outputs, including formats (markdown can be used for formatting) and the steps you want it to take to produce those
• Rules: provide constraints (story type, style, and length or number are just some)
• Examples: demonstrate what you want

Moran also emphasises the importance of iteration: use the first response as a starting point to refine the results, feeding further information into the algorithm through further prompts and interaction.

For example, your second prompt might simply be:

You are an expert on prompt design in newsrooms who has studied the subject and successfully trained many journalists. Give me five ways that I can improve the last prompt.

Quickly understanding systems

Most investigations revolve around a system and its rules: is there a systemic problem that needs a spotlight shining on it? Rules that aren’t being followed or enforced – or which have unintended consequences?

Knowing this, it can be better to ask the genAI chatbot to start from that:

You are an [investigative journalist] working for a [national broadcaster] aimed at an audience aged [20-40]. Identify parts of the [UK school] system that might be suitable for an investigation.

Mapping your target system can help you to identify potential focuses too – and here you can use genAI tools in conjunction with third party tools.

For example, ask ChatGPT to generate code that can be used with the flow chart tool Mermaid. Make your prompt human-centred, with the start and end points specified, like so:

[Write code in Mermaid to] show how a person moves through the [UK criminal justice] system from the point at which [a crime is reported] to the point of [sentencing]

The resulting code can then be pasted at mermaid.live/edit to generate an instant flow chart — and if you want to extend it, iterate with a prompt to ‘generate code that drills down further‘ into a particular step (the custom GPT Diagrams: Show Me is another option. The image below shows results from both approaches).

You should assume that the diagram (like the knowledge behind many initial ideas) is incomplete, incorrect and/or outdated — its key purpose is to get you started more quickly on doing the background research to check the aspect that interests you, to fill in the gaps, and bring it up to date.

Exploring rules and regulations with genAI

Mapping a system might lead you to a particular set of regulations or laws – or you might ask the genAI tool to identify some rules with a prompt like this:

You are a [UK journalist writing for a rural audience] looking for [feature] story ideas relating to [education]. What rules do [UK schools] have to follow in relation to [school uniform]?

Once you already know the rules you want to investigate, you can ask it to help you brainstorm ideas based on those.

Now we are getting more concrete, a good prompt should ‘teach’ the algorithm what types of stories you can tell about rules:

There are four particular types of stories that investigative journalists tell about rules. Those are:

1. Rules having unintended consequences
2. Rules not being followed
3. Rules not being enforced
4. The need for new rules

Generate two story ideas for each of these types

Teaching generative AI about idea generation methods

You can also teach the algorithm specific approaches to idea generation. The 8 angles of data journalism, for example, can be summarised within an initial prompt that begins like this:

You are [an editor] in a [data journalism] team at a [UK news website]. I am going to give you some information about types of story angles that can help you suggest story ideas based on any dataset.

Here are the first four angles. Please just confirm you understand this, and then I will give you some data to generate ideas for.

[Type a summary of the first four angles]

Once the chatbot has confirmed it has understood what you’ve described, you can paste some rows of data at the end of this follow-up prompt:

Here are the first two rows of a dataset on [the gender pay gap at each company in the UK]. Use the four angles I explained to suggest 10 story ideas for this data: [paste 2-3 rows of data]

The same approach can be used with the Iceberg Model from systems thinking, the ‘Five Whys‘ problem-solving method, story-based inquiry or any other method.

Journalism as a creative act – and an editing process

What is important about the techniques detailed above is that they do not rely on generative AI tools being truthful. Just like a conversation with a source or a vague hunch, these tools can be employed as ways of stimulating our imagination, directing our attention to an aspect of a subject that we might not have considered looking into.

Equally important, by adding this extra source of inspiration we can create scenarios where — rather than having to resort to limited options in the absence of anything better, or the time to think of it — we are able to employ a quality control process, choosing the best idea, and not just the only idea.

You are in control of this process: you select and reject the ideas suggested, or develop them further; you pursue avenues of inquiry or abandon them. You provide the seed, and the terms, and do the work that takes that idea through to execution.

In the next in this series of posts, I will look at how generative AI can be used in further stages of the story development process — identifying and approaching sources, and assisting with background research — alongside some of the risks to consider and address along the way.