All previous media technologies (printing, film, and broadcasting) have allowed many to speak. The telegraph, telephone and internet allowed us to talk to each other. Now, with artificial intelligence (AI), the machine joins our conversation and changes it in ways we can barely glimpse.
AI has been with us for years; for example, in Google services that translate our words, correct our spelling and finish our sentences. Now, suddenly, AI is the subject of intense attention from the media, legislators, regulators, technologists, investors and researchers, and deserves an entire monograph in this publication. Because right now? Because the latest installment of AI (massive language models –LLM– or generative AI) has been taught to speak. And it sounds like the same thing we do.
It is not surprising that the result of generative AI sounds familiar to us, given that it is the product of software that has mapped the trillions of relationships between the billions of words available online to be able to predict in any sentence the next term that sounds plausible. It gives us back our words, our collective linguistic style, our clichés. And encountering a linguistic doppelgänger is nothing short of disturbing.
The reflexive response to generative AI fits into two categories. The first, common at the birth of any new technology, is moral panic. Teachers worry that they will not be able to distinguish students’ work from that done by machines and fear, in this situation, that an entire generation will never learn to write. The media warns that AI will flood our information ecosystem with disinformation (as if it were not already flooded) and will tilt society towards extremism. However, some recent research shows that most people spend little time on so-called fake news and that the impact of this news is less than feared; Social extremism has deeper roots. There are even AI creators who claim that the new technology represents an existential risk for all of humanity. It’s pure marketing with a sexist bias: guys from the technological field who brag about their amazing power. In some of them, the warnings are the emergence of a false philosophy called long-termism, which values ??future humanity above the present and which seems above all utilitarianism with a pinch of eugenics. Pure nonsense.
The other response to the latest variant of AI is to rush to task it with unnecessary and risky tasks, just because it is new and attractive: a novelty. Some news agencies have used massive language models to write articles, hoping to spare even more journalists, even though it is well known that generative AI does not understand words or deeds. Thus, these agencies have ended up publishing errors, making themselves ridiculous and damaging their reputation. For that reason, I would have advised Microsoft not to attach ChatGPT, the most famous massive language model, to its search engine, Bing, because it is known to return wrong answers.
In June, I attended a federal hearing in New York to witness the case of attorney Steven Schwartz, who had asked ChatGPT to put together a docket for a court filing. When opposing lawyers and then the judge said they couldn’t find the citations, Schwartz returned to ChatGPT, asked them to, and sent the court the text of the missing cases (which turned out to be imaginary). The judge called it nonsense and called Schwartz to the stand, and the lawyer excused himself by saying that he couldn’t believe the computer was lying. Of course, the machine was not lying, because it does not understand what the truth is. It is limited to predicting words to finish sentences. Schwartz was censured and fined, not for using the machine, but for trusting it.
The most sensible response to massive language models would be to spend time studying them and understanding their capabilities and potential uses. In the case of the media, I can imagine a few. For example, massive language models can be useful in helping writers organize work: transcribing interviews, translating texts, and analyzing volumes of research to summarize or outline them, extract key concepts, find patterns, and answer questions.
I wonder if mass language models could help expand the literacy of many people who are intimidated by writing and help them tell (and illustrate) their stories. Discussing this potential with participants in a program for news agency executives, one editor was concerned that the use of AI could homogenize the language of communities, which would lose diversity and nuances in their speech, thereby that the white linguistic hegemony of those who have long had the privilege of publishing would end up being consolidated (in the United States). Mass language models reflect society’s biases in favor of established power. It is one of the main reasons to use them with caution.
In reality, mass models of language reveal how predictable, bland and boring our collective composition has become. If it is so easy for AI to imitate our writing, perhaps that should serve as an incentive to be different, to feed the great variety of voices, perspectives, experiences and discourses that exist in society more than in the mass media.
I fear, instead, that generative AI will be used to produce increasingly banal content. Such is also the fear of Matthew Kirschenbaum, an English professor at the University of Maryland, for whom, when generative AI is trained on AI-generated content, we will fall into a cultural death loop that he calls the textpocalypse. Kirschenbaum has written in The Atlantic: “Today it is easy to imagine an environment in which machines can incite other machines to emit text ad infinitum, to flood the Internet with synthetic text devoid of human intervention or intention: a self-replicating gray plague, but in textual version”.
In a paradoxical way, Kirschenbaum’s fear could lead to my greatest hope: that mass models of language commodify the media concept of content once and for all. In my book, The Gutenberg parenthesis, I write that movable type and print culture enclosed public discourse between the covers of books and transformed conversation and creativity into what we call content. Shortly after Gutenberg’s invention, censors identified a good deal of new and supposedly harmful content as dangerous to control. Then came copyright, enacted in England in 1710 not to protect creators, but to transform their creativity into a marketable good for the benefit of booksellers and publishers. That led us to think of content as a property that can be possessed. In the 19th century, the mechanization and industrialization of printing with steam presses and letterpress machines turned the contents into a product that had to be manufactured for a mass market, like soap. Later came the Internet, which made content more abundant than ever. Writers are now called content creators, journalists are assigned content quotas and no longer work for editor-in-chiefs, but for content managers. And now comes generative AI, which can write ad infinitum and commodify the idea of ??content once and for all and strip it of value.
Hallelujah! I exclaim. “When I hear people talk about content I feel like I’m stuffing a sofa cushion,” said actress Emma Thompson at a recent meeting of the Royal Television Society in Cambridge. The word content, she said, is “just plain rude.”
What will become of our communication now? I hope we can rediscover the true dialogic nature of society. The Gutenberg parenthesis (the theory that inspired the title of my book) comes from three researchers in Denmark: Tom Pettitt, Lars Ole Sauerberg, and Marianne Borch. His idea is that before movable type, information and stories were passed down by word of mouth and modified along the way, with no great sense of ownership or authorship. Or, as Neil Postman said, “culture is conversation.” However, with the printing press, that conversation was fixed on pages, confined between clear beginnings and endings. Our knowledge of the world became linear. “The line, the continuum (this phrase is a good example), became the organizing principle of life,” said Marshall McLuhan. In its beginnings, the printing press continued to be dialogic (Martin Luther conversed with the Pope through his books and his burning) until the mass media put an end to that and turned the text into a one-way proposal, restricted to those who They bought large quantities of ink. Now we are out of the Gutenberg parenthesis, and the Internet allows society to once again have a conversation with itself, since we all have a printing press in our pockets that allows us to talk online. I do not claim that the current public discourse is ideal; far from it, because we are untrained. However, all that online chatter has one benefit: it produces the abundance of language that helps train massive language models to speak like us.
In what situation does this very new technology leave writers and, especially, journalists? Will the machine replace us? No, but it could – should – change the way we think about our place in the world. At the Municipal University of New York (CUNY), I created a degree in Engagement Journalism in which we teach students not to start articles with their own ideas, but rather by listening to communities, understanding their needs and then deciding what benefit they can contribute journalism. I teach students that writing articles is nothing more than a tool at their disposal. In fact, I tell them that they should be careful with the seductive power of telling stories, deciding what deserves to be told, who appears in the story and what can be said. Thanks to the Internet, journalists can engage in dialogue with communities, call them into conversation, report on their exchanges, and help them claim their own stories.
I wonder if, with this, massive language models could transform the forms that news takes. I visited Google recently and saw an early version of their new program, NotebookLM, designed to analyze a folder of all of a writer’s research (interviews, articles, and other documents), summarize it by extracting key ideas, and answer questions related to it. Although NotebookLM uses the abilities of a massive language model, it limits its responses to the set of documents given to it, which reduces the chances of it producing hallucinations (as AI programmers say) and makes it feasible to check the output against data entry.
Google, an engineering company, hired Steven Johnson, a respected author of more than a dozen books, as editorial director of NotebookLM, a rare position in Silicon Valley. Johnson and I have wondered about the uses and impact of that service and whether readers themselves would prefer to interrogate the news rather than just read articles. At a recent journalism conference hosted by Google, editors and technologists proposed that reporters could share full transcripts of interviews and other research so readers could ask questions about the material and find the answers most relevant to their interests.
I wonder, then, if the blank boxes we face every day today (in the chat on our phones, in the Google search engine and now in ChatGPT) will change the main modes of communication in society: less reading of text and more participation in the dialogue. I feel a momentary shudder at the vision of society conversing again and of it being the machine we end up talking to. I remember, immediately afterwards, that one of the main uses we give to our phones is to communicate with other people. I hope that the media (so accustomed to talking and not listening) learns to converse and contribute to our conversations.
A key lesson from The Gutenberg parenthesis is that, over time, technology and technologists will take a backseat, and creators will take the reins. A century and a half after Gutenberg’s Bible (around 1600), a formidable creative avalanche occurred thanks to the printing press: the invention of the modern novel by Cervantes and the essay by Montaigne, the development of a market for the printed plays with Shakespeare and the birth of the newspaper. I don’t know how long it will take for current technologists to fade into the background, but the day will come when what we do with it will be more important than the technology itself. That is why I have dedicated myself to developing an Internet studies program, to incorporate the social sciences and humanities (history, anthropology, ethics, design) into the debate about the future of the Internet, to displace the technicians and promote again scholars and artists.
What are the most important knowledge that students should learn in the age of artificial intelligence? In a podcast about AI that I co-host, Alex Babin, director of a generative AI company, declared that “English majors are taking the world back again.” Knowing how to write instructions to direct generative AI is a lucrative new job skill. And Babin added: “The most attractive programming language on planet Earth right now is English.” I hope the same will soon be said in countless languages. And that could be the greatest impact of massive language models on current communication: allowing us to converse with the machine in our language. It could be the victory of human language over machine language.
Jeff Jarvis is director of the Tow-Knight Center for Business Journalism and holder of the Leonard Tow Chair in Journalism Innovation