The impact of artificial intelligence on journalism and other areas of communication

Notes from the course “How to use ChatGPT and other generative AI tools in your newsroom”, taught by the Knight Center for Journalism in the Americas.

PARTE I

Regardless of a communicator's area of ​​specialization, their role will always involve information management and content creation as fundamental tasks. Therefore, at the heart of the discussion on Artificial Intelligence (AI) is the topic of content creation, and especially journalism and news production. For this reason, I will be publishing two blog posts on the implementation of ChatGPT and other Generative AI tools, based on my participation in the course: “How to Use ChatGPT and Other Generative AI Tools in Your Newsroom,” offered by the Knight Center for Journalism in the Americas between September and October 2023.

During the first module of the course, the conceptual foundations of AI were reviewed, beginning with a definition that the module's expert instructor, Aimee Rinehart, Senior Product Manager for Artificial Intelligence at the Associated Press, described as simple: “Artificial Intelligence (AI) is a subfield of computer science.” And I love the next idea she added to her definition: “The word intelligence implies that the computer somehow has free will or can reason, but it cannot.” Let this introduction serve to dispel some of the fears of those who think that AI could replace human intelligence because, in fact, it must first be trained with information generated by humans.

While ChatGPT is the most popular system, and the one everyone seems to be talking about, there are other generative AI systems, such as Google's Bard and Anthrophics' Claudde, which, according to Rinehart, sometimes offer better results. But as she herself points out, ChatGPT is the first and already has more than 100 million users. This figure seems to suggest that there is certainly a great need for tools like these in certain professional fields. However, the question of how secure the tool is in guaranteeing the quality of the work it performs seems to have arrived as part of the same package as its advantages, so this is an active discussion that has not yet yielded conclusive answers. But there are many implementations underway, and even results from its application, which can already be consulted as a reference on the subject. Meanwhile, it falls to us as humanity to continue witnessing the discoveries that occur simultaneously with the passage of our lives, and this seems to generate considerable unease, especially because we inherited many other discoveries that we only came across, used, and took for granted.

We are now at the heart of the changes that are brewing, both technologically and in other areas that directly affect how we operate and that could have an impact on the job market. And, although many questions remain unanswered, it is certain that work, in some fields more than others, will not be the same after AI. This, however, I want to reiterate, does not mean that human intelligence will be displaced, but rather that, as has happened with other technological advances, both will complement each other. At least that is what the implementations carried out so far indicate. And it is even possible, experts anticipate, that data created by humans will become even more valuable.

But how exactly does generative AI work, and what problems are associated with it? The foundation of generative AI, like ChatGPT, lies in what are called Linguistic Models (LLMs), defined as “large swaths of information extracted from the internet to train machines…”. But the question is, what are the machines trained on, and how do they obtain answers to specific questions? In journalism, where verifying sources is practically a fundamental tenet of the discipline's code of ethics, this poses a serious problem. In academia, where sources must not only be verified but also acknowledged, it also presents a problem of considerable magnitude.

From this perspective, it can be concluded that, in both cases, the concern about transparency is legitimate, and unfortunately, there is no clear answer to date. Furthermore, experts have conducted experiments using a technique called Sparse-Quantized Representation (SpQR), which have revealed how LLMs "erode and collapse when focused on other AI-generated content," going so far as to describe it as "a photocopy of a printed photo."

Y mientras la discusión sobre estas cuestiones continuará en las próximas semanas, me ha maravillado de este curso del Centro Knight para el Periodismo en las Américas, el análisis de los grandes desafíos que presenta la IA para la humanidad debido a la forma en que esta opera. El siguiente es un recuento de esos desafíos:

– Sesgo en la propiedad intelectual debido, principalmente, a que las fuentes de las que se nutren ya tienen esos sesgos que pueden ser de género, étnicos o de otro tipo.
– Alto impacto ambiental, porque la IA consume grandísimas cantidades de energía, dejando una huella ecológica que se suma al que ya es un muy deteriorado medioambiente.
– Disparidad de riquezas. Se dice que: “dos tercios del trabajo realizado con IA se ha realizado en solo 15 ciudades de Estados Unidos”.

To the above, let's also add the potential hegemony that can be exerted by deciding where these systems will be available and where they won't. If you live in Latin America, you've probably already heard, for example, that the Claude chatbot tool isn't available in the region. Regarding the different operations that can be performed with AI, a distinction is made between: Process automation, low-risk AI, and high-risk AI. To apply this to the field of journalism, the course offered the following examples:

– Process automation: enviar simultáneamente una historia al sitio web y a las Redes Sociales.
– Low-risk AI: crear historias a partir de fuentes de datos consistentes como reportes policiales y otros tipos de información estructurada.
– High-risk AI: escribir nuevas historias utilizando fuentes de datos estructurados y no estructurados.

To better illustrate the implementation of AI in the field of journalism, I share the following quote from one of the readings that was part of the required bibliography for the course, and which contains an example of generative AI implementation by Scott Brodbeck, founder of Local News Now, on his site ARLnow:

Quote translated from English:
“Earlier last week, the audio summary automation failed and didn’t pass the day’s stories to the model. So, when the AI ​​was asked for an ARLnow summary, in the absence of stories, it just made things up. And it sounded pretty convincing!”

And, precisely in the field of journalism, this can lead us to ask the question suggested by Oxford Internet Institute PhD candidate Felix Simon, and brought to the discussion table by instructor Rinehart: How much control could AI have over the news? But while we are discovering the answer, she also offered several suggestions on how to best use chatbots and leverage them in journalism: these tools like ChatGPT, she said, “have two components: language and knowledge. Use their linguistic capabilities. Don't ask it to write a story for you, but input your story, making sure it's not a sensitive topic; and ask it to write headlines, summarize it, and turn it into a broadcast script. Nothing should be published directly without human oversight.”

The first module concluded with the suggestion that “we must understand the technology because only then can we anticipate how these systems can be integrated into workflows and prevent the technology itself from dictating this.” Furthermore, only with knowledge of how it operates can we demand transparent and ethical regulatory frameworks in all aspects.

During the second module of the course, which featured Sil Hamilton as an expert instructor, who is a researcher on Artificial Intelligence (AI) at Hacks/Hackers, “a network of journalists who reflect on and analyze the future of news through talks, hackathons and conferences”, the evolution of language models was addressed, from the one proposed by the mathematician Claude Shannon in 1951, to those currently used by the new Artificial Intelligence tools.

The development of Artificial Intelligence as a technology

In general terms, as Hamilton explained, “language models offer a probability distribution of a certain vocabulary,” which is achieved through what is called an artificial neuron. These neurons, together with others, form networks of artificial neurons that can be trained to learn to model sequences of words, eventually developing language models. This is the point at which these systems are considered to have “become intelligent” (Artificial Intelligence).

The instructor explained, however, that it took scientists so long to figure this out because “they thought it was easier to tell the machine what to say than to train it to learn to speak,” in what was called Symbolic Artificial Intelligence. Specifically, this means “that instead of telling the AI ​​to learn what a dog is, with examples, they told it that a dog is an animal. But to tell it what an animal is, they had to tell it what life is, and then that a dog is life.” Following this logic, the world would have to be described to the AI ​​word for word, which is impossible. This conclusion led to what became known as the “AI winter” in the 1980s.

Given the human spirit of constant inquiry, researchers in this field believed that neural networks could learn language and predict each subsequent word in a linguistic sequence, a feat achieved in the 2000s. More recently, particularly in the last five years, language models have been so refined that they are considered to have become "intelligent." This was further enhanced by the development of diffusion models, which can produce a complex object in a single step, proving particularly useful for images. A key advantage of diffusion models is their ability to work effectively in conjunction with language models. Additionally, the development of multimodal models, which combine language and diffusion models, has also begun.

Fundamentals of text generation

Since text generation is based on predicting words within a given linguistic sequence, one of the questions to answer is: What is the probability of that sequence appearing in nature? Language models are designed to predict this probability; that is, they are trained to provide the next mathematically most probable word in that sequence. And this process is repeated over and over. It involves providing the model with "sufficient computational power, enough artificial neurons, and examples of words extracted from the internet."

Ultimately, it's about applying logic to our existing knowledge of the world to produce words that, when combined in a sequence, appear coherent. Therefore, in a way, it becomes a global model, and because its predictive capacity has been deemed "so good," it might seem as though these systems have become intelligent. However, it's simply a system returning its collective knowledge about things to the world, without revealing its source. This is how, for example, the currently popular ChatGPT and other generative artificial intelligence systems work.

During this module of the course, however, it was emphasized that “it is important to remember that although it may seem that Artificial Intelligence systems are performing mathematical operations or writing essays, they are actually always performing the same operation.” That is, predicting the next word in a linguistic sequence or the number in a mathematical sequence. But what happens when the model is unsure of which words should come next in a given sequence? This is what is known in computer science as “hallucinations,” which is when the system appears to be lying, providing incorrect or false information.

Within this module, the example of a landmark legal case was cited, involving a lawyer who requested precedents from ChatGPT for a case he was working on. However, when the judge reviewed the list provided, he determined that all the precedents had been fabricated. The system, however, “doesn’t consciously lie… but these delusions will be difficult to correct and are certainly a problem, because the model will never know everything about the world, but will always do its best to continue with the given sequences,” the instructor explained. To put it more clearly, the instructor said, “It’s like a human being exaggerating their knowledge in an area where they have less confidence.”

Ultimately, hallucinations are what make generative artificial intelligence systems questionable for matters requiring high precision. At the heart of the discussion, for example, are news reports, where accuracy and fact-checking are essential principles. Addressing issues like these, a technology called Retrieval Augmented Generation (RAG) has been developed. In the journalistic field, this is equivalent to having the model mimic the work of a journalist, researching before writing an article. The model is given a set of data, which could be the verified archives of a news outlet on a specific topic, and it reviews these archives and produces an article in which hallucinations are less likely to occur.

Open discussions on generative Artificial Intelligence

– Given that the systems have been trained on large amounts of pre-existing information, whose information was used to train them without attribution? And another question: when an Artificial Intelligence system is fed new information and asked to perform an action, such as transcribing, translating, or improving the text, how does it reuse that information to continue training and generating more content for other users?

We will address these and other issues in Part II of this work, which we will do in a future blog post.