It's not enough to type a generic command to generate text decent with artificial intelligences… Here because!

For a couple of months now, I noticed a radical change in my editing and proofreading work. If before I only revised texts written by human beings, no I find myself reviewing a good portion of texts written by ChatGPT. At all levels: post per i social, content for sites and blogs, articles for paper magazines, slides and presentations, fairy tales and fables (No novels and essays of a certain length yet… for now).

What's the matter?

The trouble isn't so much with ChatGPT or LLMs (Large Language Models) in general, which can be useful tools for brainstorming and speeding up some parts of the creative process. I talked about it in depth in the article Artificial intelligence for text writing: it will steal our jobs?.

The point is that many people use them in a way foolish.

If you use rough prompts, if you do not reread the texts produced, if output trimming is not applied… the result will be devastating.

What do I mean by “devastating”?

• Vague and repetitive content;

• monotonous linguistic structures, since second grade;

• invented parts;

• calques from English;

• songs that don't make complete sense.

Just to mention the most common critical issues.

Nevertheless, There is a way to use these systems productively! We must not demonize technology, but rather learn to manage it better. In this article we will see how to do it!

Otto third (sensible) to write with AI

In the following points I will refer to ChatGPT because, at the moment, it is the most used LLM globally. However, it is good to remember that he is not the only one, and not even the best performing.

I started studying these technologies before the advent of OpenAI, first with Jasper (ex Jarvis) and then with Sudowrite. They are different tools, but they have many things in common, therefore the eight tips can apply to all LLMs currently in circulation!

1. Recognize hallucinations

In technical jargon, it is said that an AI hallucinates when it makes a mistake, that is, when it reports an incorrect result. Yes, because machines are not infallible, even if many are led to believe it.

The hallucinations of an artificial intelligence, Therefore, occur when an AI system produces incorrect or distorted information, without real basis.

They can be the result of various factors, including:

• training data not representative: if an AI is trained on a limited set of data, it can generate misunderstandings of information that influence its outputs. For example, if you are only shown pictures of black cats, could identify other animals or black objects such as cats (a pon-pon nero, a panther, my aunt Graziella's black hat etc);

Complexity of the task: some tasks require a more advanced level of reasoning than the capabilities of the AI ​​we are using. In these cases, the system may produce results that appear correct, but which are actually based on unverified conjectures.

How to recognize hallucinations?

First of all, we must not take the generated texts as revealed truths: we must always check them. Secondly, we need to ask the AI ​​to report the sources it uses.

2. Find sources

ChatGPT does not cite sources? And us, Instead, we make them say it!

Come? In the simplest way possible: asking him.

In the end (e, possibly, even at the beginning) of our prompts it is good practice to always put "Don't make anything up and cite your sources”.

Magic: our LLM will tell us where it gets the information from.

This method is very effective, but it always works? Unfortunately not, for a few reasons:

• depending on how the prompt is set up, can give us answers like “I'm sorry, but I cannot provide a version of the text citing specific sources since as a language model I do not have direct access to the internet nor the ability to cite specific sources” (spoiler: It is not true. Meaning what, It's true that he doesn't have access to the internet, but you can very well cite the sources);
• the sources are "dated", because prior to 2021;

the sources are fake, because the LLM makes them up on the spot.

There is a way to overcome these problems?

Yes, I suggest four:

• try to adjust your shot, by resetting the prompt and/or regenerating the response. Don't give up: with a few tries, you will achieve the desired result!
• use Bing integrated with ChatGPT, a free resource accessible to all: the sources will be updated!
• new artificial intelligence tools are born every day: don't hesitate to browse around to find the right one for you. A tool that can provide sources that I recently discovered, for example, And Perplexity AI;
• Always check the sources that the LLM tells you to use, because they can also be invented.

3. Use role play

A few weeks ago, I wrote for a science fiction blog an article on the possible colonization of Mars. To find truthful information I needed an effective prompt, that returned correct output, precise and verified. So I wrote:

“Answer as if you were a famous astrophysicist. Answer with scientific rigor, without inventing anything and reporting the sources you use. My question is: what solutions could be implemented to colonize Mars, to survive in its atmosphere and make food last? Answer with scientific rigor, without inventing anything and reporting the sources you use.”

As you can see, I did a roleplay with ChatGPT: “Answer as if you were a famous astrophysicist”. I know this may sound like bullshit, but in machine language it works. This technique can be expanded, developing the role-playing game to make it more specific.

Returning to my example, ChatGPT provided me with a great answer (which I checked later anyway, with an internet search), citing these sources to me:

1. NASA, “Journey to Mars”: https://www.nasa.gov/topics/journeytomars/index.html
2. P. Ziemann et al., “Radiation shielding with polyethylene and concrete – an intercomparison of the calculated dose rates”,Radiation Protection Dosimetry, Volume 146, Issues 1–3, 2011, Pages 239–244.
3. W. F. Schmidt et al., “Martian soil simulant MGS-1: Particle size distribution and mineralogy,” Planetary and Space Science, Volume 73, Issue 1, 2012, Pages 291-300.
4. R. Wheeler et al., “Crop production for advanced life support systems – Observations from the Kennedy Space Center breadboard project,” Advances in Space Research, Volume 38, Issue 6, 2006, Pages 1168-1174.

Once you have the sources, Obviously, I did a further check to verify them.

You may also have noticed that I wrote “Answer with scientific rigor” twice, without inventing anything and reporting the sources you use": indicating it both at the beginning and at the end makes the prompt more effective.

4. Regenerate

What does “regenerate” mean with ChatGPT? Means press the “Regenerate response” button after receiving the first output.

Why do it? Because the initial answer, Often, it's not that great. Many times, the content improves on the second or third regeneration.

Extra makeup: copy all the regenerations into one file and then select the best parts of each regeneration.

As you can see, it is not a simple and immediate process: to be used sensibly, AIs require commitment. But as I always tell my daughters: “If you put in the effort, the results arrive! Otherwise, no Pokemon cards!”.

5. Edit monotone structures

Do you know how I can recognize a text written with ChatGPT from afar? Taking a quick look at the opening lines of the paragraphs. If they are similar to each other, boring and didactic… my “ChatGPT radar” activates immediately.

Let's take a concrete example, using a prompt a lot, Very rare. Let's ask ChatGPT in a generic way: “Write a short article on the importance of healthy eating in everyday life”. Below I report the output, where I will underline in bold the parts that would put even the Jiminy Cricket to sleep on amphetamines (N.B. In this analysis I will limit myself only to the monotony of style, although it is not the only problem with the output produced):

“Healthy eating plays a fundamental role in everyday life, contributing to maintain physical and mental well-being. a diet balanced and nutritious provides our body with the essential nutrients it needs to function properly, promoting good health and preventing many chronic diseases.

First of all, correct nutrition provides our body with a variety of vitamins, minerals and antioxidants that are essential to support vital functions. These nutrients help strengthen the immune system, protecting the body from infections and diseases. Eat fresh fruits and vegetables, whole grains, sources of lean protein and healthy fats can contribute to reduce the risk of diseases such as diabetes, heart disease and some types of cancer.

Furthermore, healthy eating is crucial to maintaining an adequate body weight and preventing obesity. Eating foods nutritious and balanced helps control calorie intake and avoid excessive consumption of high-calorie foods, added sugars and saturated fats. Obesity is associated with a number of health problems, including cardiovascular diseases, type diabetes 2, hypertension and metabolic disorders.

A healthy diet it also has a significant impact on mental and cognitive well-being. Some studies have shown that a balanced diet can help concentration, improve memory and reduce the risk of developing neurodegenerative disorders, come l’Alzheimer. Furthermore, certain nutrients, such as omega-3 fatty acids found in fish, have been associated with a reduction in symptoms of depression and anxiety.

Promote a healthy lifestyle through a balanced diet it also has positive effects on overall quality of life. A healthy diet can help regular sleep, increase energy levels and improve physical appearance. Healthy skin, shiny hair and strong nails I'm alone some of the aesthetic benefits that derive from adequate nutrition.

It is important to underline that healthy eating does not it's just about what we eat, but also how we eat. Pay attention to portion sizes, chew slowly and enjoy your food in a quiet environment can help better digestion and absorption of nutrients.

In conclusion, nutrition very much it is essential for a life healthy and balanced. Provide the body with the right nutrients through a diet balanced and varied helps prevent diseases, maintain an adequate body weight, improve mental health and promote general well-being. Choosing healthy foods and making conscious choices about nutrition represents a long-term investment in our health and well-being.".

The unbearable boredom of this text comes from various factors, including:

• didactic writing for a sixth grade essay, with the various “First of all, Furthermore, it also has, but also” and the inevitable final “In conclusion”.;
• the repeated use of the gerund (bro…) in the first paragraph;
• the use of synonyms (or terms very similar to each other) consecutive, which weigh down the text and add nothing to the meaning: “balanced and nutritious”, “nutritious and balanced”, “healthy and balanced”, “balanced and varied”, “mental and cognitive” and so on;
• the repetition of “can contribute, can help, can favor” etc.

Unfortunately, At the prompt level, not much can be done to overcome this simple style (if you reading this have any suggestions, let me know!). We can try the dot role-playing game 3 (“write as if you were a competent journalist famous throughout the world for your accurate style”), we can ask to avoid repeated structures, we can adjust the tone of voice and provide some example sentences, but in any case the modus operandi will tend to always remain the same.

The solution is to identify boring constructions and fix the content yourself, to make it pleasant and dynamic. Must, Therefore, do actual editing of the text… or have it edited by a professional! A human editor is, listen to me.

6. File off the repetitions

In the previous example you will have noticed the disproportionate quantity of repeated words. The same advice we just saw applies: reread carefully and edit the ineffective parts.

7. Calibrate the tone of voice

To make our LLM understand what tone of voice we want from him, it takes commitment and patience. Here are some useful tips for this purpose:

• indicate in the prompt a brief description of the tone you have in mind: formal, informal, friendly, technician, captivating, ironic and so on;
• provide the AI ​​with examples that reflect the desired tone, so that I can adapt the lyrics based on those songs;
• explains the context and purpose of the content;
• describe your target;
• enter in the prompt some keywords or phrases that you want to include in the text.

8. Get rid of Englishisms

In English, “memorable” means both “unforgettable” and “easy to remember”. In Italian, Instead, we have two different words for these two concepts: “memorable” and “memorable”.

Recently, I edited a text that contained the term "memorable" instead of "memorable". Why? Of course, my dear: Why it was written with ChatGPT, which has English as its source language.

Therefore, listen to me: watch out for Anglo-Saxon casts, in the process of rereading and editing!

I hope you found my tips helpful! If you are interested in consultancy or editing for your content written with the help of AI, don't hesitate to write to me!