Everyone is talking about ChatGPT – we are, indeed – for several reasons, including timing, marketing, hype, and so on: this is the very first time a generative machine based on deep learning has gone viral.

For these reasons, this is not only a guide about ChatGPT for your job but also for approaching the whole world of the start-upper, incumbent, and brand-new generative software based on deep learning.

We are now focusing on text generation, even if we can find generative tools in any field right now (music, images, data, voices). And we are assuming this is just the beginning.

What is ChatGPT?

You can ask the tool itself if you want.

Here is the answer I received: “ChatGPT is a conversational AI model developed by OpenAI. It uses deep learning techniques to generate human-like responses to text inputs, allowing it to have conversations on a variety of topics”.

So, you log in to the service, write something and receive something else as an answer. Whatever you will ask, you will receive a reply.

What kind of problems related to text-generative machines do we have?

We need to understand the problems to start this guide.

First, let’s clear the field of any misunderstandings: even if it has an interface like a chat, simulating someone typing, ChatGPT isn’t really conversing with you and doesn’t have the faintest idea of what it’s doing.

Moreover, as Luciano Floridi and Massimo Chiriatti wrote in their essay for Minds and Machines, any interpretation of GPT-3, the model on which chatGPT is based, “as the beginning of the emergence of a general form of artificial intelligence is merely uninformed science fiction”. Or marketing, I would add.

Second: we must be aware that these technologies pose political, ethical, moral, and economic problems. They do not write the truth – whatever the truth is – and they do not in any possible way replace the method of journalistic verification. Machines cannot understand the status of truth, either moral or factual.

They are systems trained to write starting from dozens and dozens of terabytes of texts and data produced by others (perhaps even by you or by me: who knows?) or by other machines. The training poses a severe problem with sources: who can guarantee that these machines are trained with proper facts put in appropriate contexts? Are these machines transparent? And what about copyright? For example, Sarah Andersen, Kelly McKernan, and Karla Ortiz have launched a lawsuit against Stability AI and Midjourney for their text-prompts-to-images generator.

We need to know that, like other technologies, they could lead to reducing jobs and to a profound transformation of our societies: we should guide those transformations and not be overwhelmed by them.

It should also be known that, combined with other generative software (such as those that generate photos or images, or audio, for example), they create a perfect ecosystem for speeding up the production of any content, authentic or fake. In a short time, it will be challenging to rely on audio or video recordings to have a source to cite that we can trust, even if they seem to be different and taken from separate and independent sources.

Moreover, these machines have biases that depend on the sources with which they are trained. If I feed a generative machine with a series of sources written by white, wealthy men, the machine will necessarily reflect their points of view.

Third: we can’t trust these machines. We can’t use them as a learning tool, we can’t use them to produce verified content. To avoid mistakes (even in the way we cover this topic), we can read and use this guide to pitfalls in AI journalism and remember that, according to Princeton computer science professor Arvind Narayanan, ChatGPT is a bullshit generator. And in fact, ChatGPT always answers you (or specifically refuses to satisfy your prompt if the topic is a sensitive one), sometimes fabricating an answer to satisfy your question.

They also can be used as very cheap fake content creators, of course.

Fourth: even if it would be useful to know, the battle to understand if a text has been written by a machine or by a human being is already lost: check the classifier launched by OpenAI and see how many limitations it has. Moreover, what about text partially created with these tools and partially edited? It would be much more interesting and compelling to have tools to spot AI generated images or videos, for reasons connected to the spreading of disinformation.

If we know all this and if, in general, we approach the topic with a sufficient critical spirit, we can start with the guide.

Surely there will be those who will use this software to quickly produce a lot of crappy content. But this, unfortunately, was done even before, underpaying human beings, for example, to make traffic thanks to SEO techniques.

We want to talk about something other than shortcuts and tricks here.

We want to talk about how to use ChatGPT and other generative software to use them as our assistants. We can use them to free our time and focus on the core of journalism.

Learn to ask. The art of prompt hacking

Some machines rely on algorithms similar to what, for simplicity, we can call “artificial intelligence” that we undergo (for example, recommendation algorithms, unless you know how to train them).

Some machines are based on machine learning: we work with them by giving commands (for example, tools like Grammarly or ProWritingAid, Google Translator, and Pinpoint).

There are machines based on deep learning, like ChatGPT, that we can ask questions using natural language.

Technically, these questions are called prompts.

The more you specialise in writing prompts, the more these machines will help you save time.

One way I suggest learning how to ask is undoubtedly to start “playing” with ChatGPT or other software, exploring their limits. You will find, for example, that it makes logical errors and that it’s easy to “convince” of something that is not true. However, upgrade after upgrade, the tool gets better. And when those connected to Google’s massive database will arrive (Sparrow), these types of errors will be minimised or reduced to zero.

Playing with the machine to explore its limit is not to write articles mocking the tool – boring! – but to remind us that the method of verification is still up to us human beings. And to have a hacking approach, like taking toys apart to see what they look like inside and what we can do with them.

Learn what to ask, and how

Exploring limits is an excellent way also to understand what to ask and what kind of jobs you can partially delegate to ChatGPT or similar machines. ChatGPT is not so good at creating text from scratch – you can still do better if you have enough time and are well paid – but it becomes powerful if you use it as your personal assistant in your workflow and for things that you already know.

Identify entities – ChatGPT is good at finding entities. If you want to find all the emails in a text, all the names and surnames, all the locations, all the five-letters words, or whatever can be defined by a machine-understandable rule, ChatGPT can do that for you. Just copy-and-paste the text and ask: “List all the emails/name/location/chemical elements… in this text”.

Summary and abstracts – ChatGPT is very good at summarising points. Insert a long-form article and ask the machine, for example, “Provide me a summary of this text in 5 bullet points, focusing on the five W, Who, What, Where, When, and Why”. And then copy-and-paste the text you need to be summarised.

This operation could seem scary. But it’s just because we, as human beings, associate the summary with understanding the text, because that’s what we must necessarily do to summarise.

The machine does not understand, but it’s good at calculating the frequency of words and at producing the next part of a sentence based on probabilistic maths. So, use it for summarising. And then spend some time editing and verifying the summary.

Please note that I’ve been very specific with my prompt. It’s like giving the proper brief to a junior assistant.

Different versions of the same text for headlines, social posts, and meta-descriptions – You can use ChatGPT to write or rewrite different versions of a sentence or text. You can use it for writing meta-descriptions, headlines, different versions of a social post, and even changing your tone of voice, adapting it to the context.

As you can see (forgive the machine for the maths mistake), the result is quite disappointing and very basic, even if someone could find this precisely what they were waiting for. I’d never use these texts for my social media.

Why has this happened? Because I needed to be more specific in the prompt. The machine does not know what I am talking about in my article. So, I copied and pasted the article you are reading into ChatGPT and asked the machine: “Create a post for Linkedin from this article, summarising in 4 bullet points with long sentences the following article, and use emoji instead of bullet points”.

This is the result.

Then, I asked the machine: “Create also a Twitter thread from the same article”.

And finally, I asked: “Now create a three-paragraph emoji post for Facebook about the same article, making fun of the fact that a machine can already do the equivalent of an underpaid journalism job well”.

As you can see, I went too far asking to make fun of something: that’s not fun at all, but it’s helpful to show the trial-and-error process.

The following prompt to correct my mistake – remember: the machine does what I am asking – could be like this: “Let’s start again from my article. Don’t make fun, but use a casual, informal tone of voice to summarise it for a Facebook post with a couple of emojis and to engage a journalistic audience to read the whole article”.

Once you have found a series of prompts that work for you, you have to save them and use them for your workflow.

Of course, you must verify that the content fits your needs.

Mental map or summary of a topic – Asking ChatGPT to create a mental map of a particular topic is helpful to see whether you are missing something and to organise your thoughts when you start working on something.

The prompt could be something like: “Create a mental map of a comprehensive, advanced guide about…”. For example, Italy has a huge journalistic interest in anarchism these days.

Then, you can ask the machine to go deeper on any single point. For example: “Let’s take in particular point IV A and create a mental map of that point.”

The result could be a good starting point for personal or professional research, that you can integrate with your knowledge. Of course, the more you know about a topic, the more you can use this machine as an assistant.

Am I missing something? – You can text ChatGPT with some context, and then ask the machine if you are missing something. For example, I asked ChatGPT if I was missing something from this article.

These are the gaps in my article, according to ChatGPT.

Of course, you can agree or disagree with the suggestion: again, the cooperation between the human being and the machine makes the difference.

And if you find something missing, you can always ask the machine to provide you with some ideas to fill the gap.

Like I did about suggestions of examples and case studies of how ChatGPT are currently being used by journalists.

I don’t agree at all with point 1 and I’d neve recommend it. I’d never use these tools to create articles quickly from scratch. They are still forced to answer you whatever you ask them, and you can’t trust them as a reliable source.

Points 4 and 5 are interesting but far from the aim of this article.

Point 3 is something that I can add, experimenting a little bit.

Sentiment analysis – Yes, you can do some sentiment analysis with ChatGPT. For example, with a prompt like this, I reworked starting from this “Cheat Sheet”:

“I like pizza, positive

I don’t like pizza, negative

Sometimes, I like pizza sometimes I don’t, neutral

Analyse the sentiment of this sentence: While the movie was good, I sometimes thought it was a bit dry.”

Let’s say you want to analyse the replies to a tweet of someone trendy. If you provide examples to ChatGPT with a precise classification of what you consider positive, negative, or neutral (you can add different degrees of sentiments or change the kind of classification) and then copy-and-paste a series of tweets on a topic, for example, you can ask the machine to provide you with a sentiment analysis.

Remember that ChatGPT will be forced to assign a degree of sentiment to any sentence, starting from your classification.

You can also ask ChatGPT to create a table, adding this to your prompt:
“Provide me the analysis results in a table with two columns. The first column is called ‘Sentence’, and the second column is called ‘Sentiment’”.

You will have a table ready to copy into a sheet.

Please notice that I’m always redundant when I write prompts.

Code and formulas – And something else is missing: you can use ChatGPT to write code. For example, I used it to create a quick javascript for the Chrome Console that extracts any URL into a web page.
You can use ChatGPT for excel formulas, too, asking the tool the function you need.

A workflow with audio or video recordings – I integrated ChatGPT in my workflow when I deal with audio or video recordings, in this way. I put the file into Pinpoint to get a transcription. Then, I copy-and-paste the transcription into ChatGPT and ask the machine to summarise it.

Again, this is an incredible saving time operation, but you need to double-check what you are using for an article.

The upshot

This is just the beginning of these tools, and that’s the reason why it’s the right moment to play with them.

Apart from methodological, political, ethical and economic considerations, from a technical point of view, here are the takeaways:

we can’t trust ChatGPT (or any machine);
the machine is like an assistant;
the machine does not know what it is doing;
identify the points in your workflow where you can use it;
use the machine to save time
the better you ask the machine, the better the answers;
be very specific with your requests, rephrase them if they don’t work;
give the machine context, even train it with whole chunks of content;
save prompts that work for you;
combine prompts and test new ideas;
you can change the tone of voice, you can change the type of text (formal, informal, detailed, simplified);
you can also change the kind of personas you are asking the machine to impersonate, or the kind of personas you are asking the machine to address
combine various prompts and techniques;
always check everything before posting;
it’s an early stage of this technology: before paying tons of money once they’ll become for paying users only, it’s probably wiser to wait

journalism

Latest in journalism