In the last year Artificial Intelligence (AI) has leapt into public attention, with the launch of tools like Chat GPT and Google’s Bard. There has been much speculation about what it means, how it works and what it might mean for our society. One key issue is how AI may be used in news, politics and opinion forming. Our readers are entitled to know how we are responding.
What is generative AI?
We are all used to search engines like Google. These work by trawling through the internet to find the words you have entered, and they return links to the webpages which make most use of those words.
Generative AI is different. It works by gathering vast quantities of data, and applying huge computer power to create text, images and sounds in response to any prompt. Whereas Google finds existing material, generative AI creates something new which is like existing material.
Large language models
In the world of journalism and news, a key form of AI is the Large Language Model (LLM) – Chat GPT and Google Bard are both examples. Both are “trained” on large bodies of existing text, and then seek patterns in that material from which to create new text. If you ask Bard to explain this, it will tell you that,
LLMs are trained on massive amounts of text data, which allows them to learn the patterns and relationships between words. This knowledge allows them to generate new text that is grammatically correct, coherent, and even creative.
These are not patterns that humans have prescribed, but patterns that the AI itself has detected in the data. And those patterns may be undetectable to humans.
So, an LLM can answer questions, summarise and translate text, and generate original material. At the mundane level, it can answer complex questions. For example, asked what Keir Starmer thinks about the EU, Bard produces, in seconds, a reasonable summary of his public position, based on what has been reported, illustrated by a set of quotes from speeches over the last three years.
Versatile and creative …
Bard itself takes a wildly optimistic view of the issues. Asked to generate a poem about LLMs, it produces.
In realms of words, where knowledge lies, Large language models, ever rise. With minds of vastness, swift and keen, They weave through language, unseen. From depths of data, they gather lore, Unraveling secrets, evermore. In every query, they find their art, A symphony of words, set apart. Through codes they dance, with logic's grace, Unraveling mysteries, secrets to chase. With every prompt, their minds ignite, Illuminating pathways, casting darkness to flight. Oh, large language models, wise and grand, Your potential boundless, in this digital strand. May your words inspire, uplift, and guide, As we journey together, side by side.
Apart from the rather weak understanding of scansion, this could have been written by a human: but it wasn’t.
… Or perhaps dangerous?
This is evidently a powerful, but dangerous, tool. Generative AI creates things which look like reality, but do not necessarily exist. Concerns have been expressed that when the data is inadequate, LLMs can invent material. Asked to produce an academic paper on a given topic, an LLM will produce a decent result. However, if there are gaps, it may fill them with plausible inventions, like references to non-existent articles, perhaps attributed to a real researcher in the field.
Similarly, in the wider field, generative AI can create pictures or events which never happened. It may be positive for people to be able to use generative AI to make an online speech in their own voices, in languages they do not speak. But, on the other hand, it can also generate a speech, delivered in the voice of a politician, who had never even made it.
What do people think?
Understandably, people are anxious about these developments, and the pollsters Savanta have recently carried out an opinion poll, asking the public what they think about it. They found significant differences in how people viewed different ways of using AI in journalism. As the table shows, they were most worried about AI writing opinion pieces, and least worried about using it to generate things like quizzes.
Gender and age matter. Older people and women are more suspicious of most uses, notably for fact checking, where 51% of men, but only 41% of women, think this is acceptable.
For some years, public trust in the media has been declining. AI is likely to increase that anxiety, and Savanta found that 52% of people said they would trust a news outlet less if they used generative AI without explaining how it was being used.
Control and regulation
So, around half of all people are worried about some aspect of AI use. But while 78% of respondents expect government to ensure that AI is used responsibly by the media, there is less clarity about how that could be achieved. Thirty percent believe that businesses and news organisations should be required to disclose how they are using AI. A third of people were worried about the implications for peoples’ jobs, and wanted regulation over accuracy and ethical issues.
But, given the speed at which AI is evolving, control and regulation will be difficult, as new uses and challenges are constantly emerging. Twenty percent of people want government to ban all use of AI until it is better understood. But, given how widespread, and global, its use already is, it is difficult to see how that could be achieved.
Full disclosure: how we are using AI
Like most news organisations, at East Anglia Bylines we are experimenting (cautiously) with LLMs. Our broad approach is consistent with the views which Savanta found. There are valuable uses which we are exploring, but with careful checking of its products for accuracy and honesty.
Firstly, we are using it as a research tool. We have been using Google for years. But the quality of a Google search depends on careful choice of keywords, and once the search is completed, one has to trawl through the links to look for reliable and relevant answers. By contrast, a careful question to Bard will produce, not a list of websites, but a coherent answer, drawing on the most relevant sources. It can also provide leads to investigate. The answers need checking, but for a small team of (unpaid) citizen journalists the time saving matters.
We also use it for summarising. Some of our stories are based on, or use, Government reports and academic papers. These can be fairly impenetrable, but AI can quickly simplify the language, generate a summary, and suggest headlines and tweets to promote the article. Bard will produce three different summaries of a long report to a specified length. We can select the best, check its accuracy, edit for style, and perhaps add material.
Its editing capacity also has potential for a citizen journalism site like ours. We are keen to involve a wide range of people writing for us. Many have important things to say, but not much experience of writing for publication. LLMs can rewrite a rough draft into an appropriate style, and the result can help people to improve their own writing. In the same way, we sometimes publish stories based on press releases, which can be shaped to fit our style and formats using Chat GPT. When a story has been rewritten in this way, we always indicate this.
We aim to be a trustworthy source of news and information. We are using AI, and will continue to develop our use, carefully, and we hope with due regard to ethical and accuracy issues. If you think we are not achieving that, please tell us!