Using Artificial Intelligence to Write Scientific Press Releases: Is AI About to Take Over Our Jobs?

Is artificial intelligence already capable of writing press releases for scientific articles or is it still too early to tell?

When we posed the question of whether artificial intelligence (AI) could write a scientific press release, it led us to pose a series of other questions. What would happen if we asked an AI bot to write a press release for a scientific paper we ourselves had not even read? Would the authors of the study realise what had been done? And what would happen if we sent the press release created with AI to the media? Would anyone notice? There was only one way to find out.

The following is the story of a small experiment. The underlying issue is a source of fairly widespread concern, which comes down to a single question: Is AI going to reduce our workload or is it going to take away our jobs? Let’s see!

The scientific article we chose was one by researchers working in the ISGlobal programme on the lifelong health impacts of environmental factors. The article had been accepted for publication by a scientific journal but had not yet been assigned a publication date, which meant that we had time to change course if the experiment failed and we needed to return to the traditional method of writing the press release ourselves. A quick look at the abstract revealed that the press release for this study was not going to be “one of the easy ones”. Let’s see what ChatGPT thought.

We used the free version of ChatGPT, 3.5, which is based on a dataset of knowledge collected before September 2021. In any case, since the article had not been published, the only way to proceed was to paste the text into the AI model.

Note: In this exchange, ChatGPT answers yes to the author’s question about whether it can write press releases. In response to the question of whether it can access scientific articles on the web, it says it cannot because the cutoff date for its dataset was September 2021 and it is unable to access new data online.

So, I pasted in the abstract and in return was immediately hit by a cold shower of reality. It doesn’t matter that the press release it produced had a few inaccuracies. It doesn’t matter that it was missing some of the details found only in the full text of the article. This was only version 3.5 of ChatGPT! And it did the job! ChatGPT was able to find the news angle and produced a press release for the general public and an appropriate headline, which was precisely what I had asked it to do. The creators still have to improve a number of aspects of how the tool works, but it is abundantly clear that this is only a matter of time. AI is not a tool for the future. It is already here. And what comes after this point is what in basketball is called “garbage time”: the minutes left in a game after the outcome has already been decided. Even so, let’s take a look at some of the complexities associated with this process, of which there are many.

How we got ChatGPT to produce the desired press release

The first snag we encountered was that, at least in version 3.5, ChatGPT is impatient. In my first instruction (see below) I explained that I would paste in the abstract first and then the rest of the text and, from its response, the bot appeared to have understood the instruction. However, as soon as I pasted in the abstract, it started to produce the press release without waiting for the rest of the text.

In my next instruction I tried to redirect the bot by telling it that I was going to provide more information.

Second surprise—there is a character limit and I can only paste in the introduction and methodology sections. Since ChatGPT is unaware of this constraint, it immediately created a second press release, this time in Spanish. Although I gave it instructions in Spanish, the first press release it produced was in English, the language of the article.

It was at this point that the problems started: no matter how often I instructed the bot to wait for the complete text, my friend Chat just could not wait to demonstrate its skills. On a couple of occasions it even forgot what we are doing and I had to remind it that I wanted a press release and not a text summary. If I forgot to specify what language I wanted, it alternated between English and Spanish.

Note: In this exchange the author reminds ChatGPT that he wants a press release. The bot responds appropriately. He then adds additional text, asking the bot to take it into account. The bot responds appropriately and adds a comment “this text is quite technical and detailed making it important to adapt it and make it more accessible … to the general public.”

I ended up with as many as six different versions of the press release. Although I had asked the bot to include a paragraph on the methodology used, its description left much to be desired. To remedy this shortcoming, I read one of the summaries it provided to understand the methodology and wrote a paragraph myself that I liked better.

Note: The author requests a description of the methodology for a general audience. The bot responds appropriately

Great. Now, which of the six versions should I choose? Each one has at least one better written paragraph than any of the others, so the only way I can see to proceed is to make a collage of the press releases provided by ChatGPT. Since I don’t want to spend more time doing this than it would have taken me to write the release myself in the first place, I try to do it as quickly as possible, remaining faithful to the self-imposed condition of producing the press release without reading the text of the article. When the cut and paste operation is complete, I send the resulting text to Parisa Montazeri and Martine Vrijheid, first and last authors of the study, respectively, asking them to review it thoroughly, but without giving them details about the role of my colleague Chat.

The researchers review

The document came back with a good few modifications in tracked changes. However, there were no significant errors since the authors’ edits were mainly aimed at adding precision and polishing some details, something we are used to seeing when we write press releases ourselves. Interestingly, one of the authors’ comments referred to a redundant sentence, an error not attributable to ChatGPT but rather to the human who intervened by cutting and pasting from the different versions of the press release. I polished the text and sent the final version to the authors together with a short questionnaire.

• In general, what do you think of the quality of this press release?
• Was the factual content correct?
• Was it well written?
• Does it meet the professional standards expected from a research institution like ours?
• Did you notice anything unusual about this press release?

In their responses, the researchers said that they had not noticed anything out of the ordinary until they received the questionnaire, that the quality of the text was “good” or “very good”, and that the data and wording were correct. They also said that the release met professional standards, although both added that they had to “correct details”. Parisa Montazeri mentioned that she intervened to emphasise one of the methodological novelties of the research (something that ChatGPT had done in at least one of its six versions of the press release). In other words, ChatGPT passed the test with flying colours.

Ethical considerations on the use of AI by communication departments

This experiment raises a lot of questions and we only have space here to mention a few:

• Is it ethical to use ChatGPT without stating this in the press release, even though this type of communication is normally produced by humans anonymously with no byline?
• Can we afford the luxury of using AI tools given their enormous contribution to CO₂ emissions?
• What does ChatGPT do with the information we give it? Could the information be leaked? Are we breaking an embargo if we provide the text of an unpublished article to an AI bot?
• If we hand over our writing tasks to an artificial intelligence, what will happen to our own natural intelligence?
• If we start outsourcing the tasks that used to exercise our brains, will we lose our skills? Will we lose our writer’s skill and finesse and, therefore, our journalistic judgment when correcting ChatGPT? Will we lose our creativity and ability to innovate?
• To what extent is ChatGPT capable of proposing something new that transcends the logic of the information on which it is based?

A few thoughts out loud (not conclusions)

Many of the difficulties we encountered in this experiment were due to ChatGPT’s character limit (which also exists in version 4). This would not appear to be an insurmountable barrier.

Despite the inconvenience of ending up with six versions, the total time invested in producing the press release was much less than it would have taken if I had followed the usual process.

Artificial intelligence has been compared to the printing press, to electricity and even to the atomic bomb. Another apt analogy might be the mechanisation of production that led to the industrial revolution. Those of us who work in communications should already be thinking about other ways we can generate added value, since it is already possible today to automate a large part of our work.

Today—and I don’t know how long this will continue to be the case—I believe that any well-trained communications professional can produce a higher quality press release than ChatGPT (but there was also a time when the great chess champions could beat the machines).

More research is needed, obviously. In fact, some other experiments we did after the one described here have not produced such good results. But that’s a story for another day.

And what happened to the press release?

The release, once properly edited and corrected, was sent out to a number of media outlets and was published on our website (if anyone is curious you can read it here). The information was picked up by a good few Spanish and international media, including CNN. If anyone noticed that the basic text was produced by ChatGPT, they have not said anything. 

Would we do it again?

As part of our work, we have to educate ourselves on developments in AI and the tools available to us. However, we do not intend to delegate the communication of scientific results to AI. For that task, we have a magnificent communications team who are not only highly intelligent but also stand out because of their great humanity.