OpenAI’s new AI conversational bot, ChatGPT, has quickly gained popularity, accumulating one million users within five days of its release. The language model, which is based on GPT-3, has the ability to generate text and engage in conversations. In this text, the authors examine the potential impact of this technology on media research.
In just five short days, the new AI conversational bot from OpenAI, ChatGPT, amassed a staggering one million users. This spin-off of GPT-3 has caused quite a stir on social media, with users posting their results and discussing the technology. ChatGPT has the potential to revolutionise how we communicate, with its ability to generate text in response to prompts and engage in conversation. The language model has been trained on an extensive dataset of over 300 billion words, including books, online texts, Wikipedia articles, and more. However, it should be noted that ChatGPT is unable to browse the Internet and is therefore limited to the knowledge it had prior to its training in 2021. This means it may not be aware of current events, such as Russia’s attack on Ukraine or Argentina’s win in the World Cup. It’s also worth noting that while ChatGPT excels in producing impressive text in English, its performance may not be as strong in smaller languages such as Finnish, Swedish, Danish, or Norwegian. For example, it may not have the same level of proficiency in producing rhymes in these languages.
The performance of ChatGPT in English has been nothing short of impressive, raising concerns about its potential impact on various industries. In higher education, for example, some have speculated that the technology could transform the way research and essays are assigned to students. With ChatGPT able to retrieve and summarise information in response to prompts, the need for students to write may become obsolete, and detecting student fraud may become more difficult. The technology also poses a threat to the dominance of Google search. In Finland and Norway, teachers have raised the alarm about the potential for ChatGPT to be used for cheating, despite some positive learning outcomes. Others think it should be embraced as a supportive tool in teaching.
However, not everyone is convinced of ChatGPT’s capabilities. Some have criticised the technology, calling it a “meme machine” that produces nonsensical texts, or “hallucinations”. The problem of hallucinations in complex language models like ChatGPT is a serious one, as demonstrated by the quick demise of Meta’s large language model for science, Galactica. After only three days of operation, Meta took Galactica offline due to its tendency to produce biased and incorrect gibberish. This issue is reminiscent of the controversy surrounding Microsoft’s chatbot, Tay, which was trained by Twitter users to become racist in a matter of hours. One notable difference with ChatGPT is its ability to admit when it is wrong or lacking access to fresh data on current events, a step towards the development of “explainable AI”, a pressing need in the creation of applications like automated decision systems. ChatGPT is happily correcting itself when fed with appropriate information, for instance, by other generative AI.
The impressive performance of ChatGPT can be attributed to its use of a feedback process called Reinforcement Learning (RL), in which the technology learns by interacting with humans. In this case, people employed by OpenAI conversed with ChatGPT and rated its responses in order to teach the bot which responses were appropriate and which were not. Based on this feedback, ChatGPT was able to improve its output. It’s worth noting that the lack of human feedback in other languages may be a factor in ChatGPT’s weaker performance in languages other than English.
As the hype surrounding generative AI begins to dissipate, it’s worth considering what the future may hold for this technology and its various applications. In particular, the ability of AI to produce decent articles without supervision or pre-planning raises questions about its potential impact on the media and journalism industries. While generative AI has the potential to transform workflows and products in creative fields where facts and timeliness are less critical, it remains to be seen how it will be integrated into fields such as news and journalism.
In 2019, a group of Nordic researchers came together to explore the concept of Augmented Journalism, or the use of advanced technologies in journalism. These workshops, funded by the Nordic research councils, were held both online and in person in Bergen, Stockholm, and Helsinki throughout 2022. The pandemic necessitated the initial workshops to be held online, but restrictions eased enough for in-person meetings to be held later on.
Our discussions on Augmented Journalism had already begun prior to the launch of the first version of GPT-3 by OpenAI in June 2020 and the subsequent release of Dall-E, an AI system capable of creating images and art from natural language descriptions, in January 2021. This autumn, Dall-E was made available to the general public, and several other synthetic media tools such as Craiyon, Midjourney, and Stable Diffusion have also emerged. Meta has even showcased Make-a-Video, which can generate five-second videos based on text prompts. While the output of these AI-generated videos is currently easy to identify as such, it remains to be seen how these and other synthetic media tools will be integrated into journalism and other fields.
While it may be amusing to play around with ChatGPT, the potential for its capabilities to be utilised for nefarious purposes is a serious concern. Could we be entering a new era of disinformation with the advent of ChatGPT? Casey Newton has outlined some of the ways in which text generators like ChatGPT could be used, including the creation of spam, the design of sophisticated phishing attacks, and the writing of artificial op-eds meant to spread hoaxes or further information operations. These are just a few examples of the potential dangers of such technology, and it will be important to consider the ethical implications as these tools continue to develop.
While ChatGPT and other tools that can produce text in a convincingly natural language are impressive feats of engineering, they also pose a risk to society by eroding trust. As Wael Abd-Almageed, a professor at the University of Southern California, warns, “once the line between truth and fake is eroded, everything will become fake. We will not be able to believe anything”. The potential for these tools to negatively impact trust in information and communication is a pressing issue that must be addressed as they continue to advance.
Janet Haven, the executive director of Data & Society, echoes these concerns, stating that ChatGPT and similar tools “will be used in adversarial ways that are intended to undermine trust in information environments, pushing people away from public discourse to increasingly homogenous communities”. As AI technology continues to advance, it will be important to consider the potential consequences for trust in society and the risks of such tools being used in harmful ways.
This presents a unique opportunity for media researchers to delve into the emerging field of advanced forms of disinformation and the potential interdisciplinary approaches that may be necessary to address it. For instance, media scholars will need to work closely with computer scientists in order to gain a thorough understanding of the technology at play. One example of this collaboration can be found in the Nordic Observatory for Digital Media and Information Disorder (NORDIS), a consortium of media and journalism scholars, political scientists, and computer scientists who work alongside fact-checkers from Denmark, Norway, Sweden, and Finland. NORDIS is one of the national hubs of The European Digital Media Observatory (EDMO), which covers all EU countries.
In addition to exploring the potential impacts and challenges posed by generative AI, media researchers can also play a role in helping publishers understand how to use this technology in a responsible manner.
ChatGPT offers researchers new tools for conducting media research, including the ability to systematically categorise news and journalism output through the use of taxonomies, tagging, and other classification methods. It can be used to produce stimuli for qualitative and quantitative research on audience responses to different formats, styles, levels of complexity, personalisation, and customisation. Additionally, ChatGPT can provide access to procedural knowledge through its ability to respond to “chain-of-thought” or “step-by-step” prompts. There is also the potential for more advanced language models – perhaps future versions of ChatGPT – to reveal new knowledge through more advanced “explain why…” prompts. Additionally, researchers can use ChatGPT’s text generation abilities to communicate research results using simpler, more accessible articles, research summaries, multi-article summaries, and other forms of publishing. Furthermore, the model may be useful in analysing and summarising large numbers of text documents, for example, in enabling researchers to quickly analyse selected documents for literature reviews.
Beyond providing useful tools for researchers, the arrival of ChatGPT also highlights urgent new areas for research in media. For example, there is a growing need to explicitly clarify the essential contributions of human journalists to the field of journalism – the human component of “augmented journalism”. Furthermore, with the likely rise of disinformation and misinformation produced using language models, research on these topics and their effect on the public’s trust in media takes on new urgency. Additionally, the emergence of ChatGPT highlights the potential for entirely new journalistic workflows and news products, requiring further research on newsroom processes in which human reporters are augmented with cognitive machines.
The final Augmented Journalism workshop took place in Helsinki in December 2022, just a few days after the launch of ChatGPT. The news sparked a lengthy debate among workshop participants, many of whom agreed that we are entering a new age of ubiquitous synthetic media with long-term implications that are difficult to predict. The conversation surrounding ChatGPT and its capabilities highlights the need for a new concept to understand the complex relationship between advanced technologies and journalism: one that moves beyond technological determinism, solutionism, and defeatism to consider how human capabilities, values, and emotions – particularly empathy – will become increasingly important. How can AI and other computational tools be used to support and “augment” these human qualities and strengthen journalism as a democratic institution? In developing this concept, we draw inspiration from the idea of the augmentation of human intellect as outlined in Doug Engelbart’s seminal report from 1962.
During the final workshop, participants couldn’t resist the urge to play with ChatGPT, and they submitted the prompt “Define ‘augmented journalism’“. ChatGPT’s response was “Augmented journalism is a term used to describe the use of technology, such as artificial intelligence, to assist journalists in their work“, which is both accurate and succinct but also quite generic. A useful definition of what it means to truly “augment” journalism will clearly still be up to us. ChatGPT continually surprises its users with its capabilities, and it appears to be only a matter of time before it is routinely used within augmented journalism workflows to help journalists produce publishable content.
What might such content look like? Well, you’ve been reading an example of it in this article. The authors produced a draft text and some supplementary notes, and then turned to ChatGPT to polish those into an accessible and readable article, which was then lightly edited and supplemented with a small amount of hand-written text. This was indeed a joint effort between people, who provided the original elements, and machines, which produced many of the stylistic elements.
Illustration is made by image from Cotton Bro via Pexels and screenshot from ChatGPT.