Trust. The issue of trust really encompasses both of the above. Trust is the currency of journalism. Any deployment of new tech tools must in no way leave room for people to question the integrity of a publication. Having said that, we’ve found that readers are generally happy to embrace robot written content – as long as the information is valuable to them, and clearly labelled.
If a publisher asked “What does generative AI mean for our business?”, we’d like to ask back: “What do you want it to mean? The AI is not in control, you are.”
We would advise publishers to keep focussing on delivering solid, valuable journalism and use generative AI tools where they are helpful in this mission. Charlie Beckett, director of the JournalisAI project at LSE expressed it perfectly in a podcast recently, saying that these tools cannot ask critical questions or work out the next step in investigating a story, but that they can be a support to journalists in doing this work. “But I think it’s even more interesting how it puts a kind of demand on those journalists, saying ok – you’ve got to be better than the machine – you can’t just do routine, formulaic journalism anymore, because the software can do that.”
We’re only at the beginning of exploring how generative AI can support the business of journalism. Trying out ChatGPT is easy – working large language models into robust and useful processes within a publishing business will be considerably harder. It will be crucial to keep a razor sharp focus on the use you’re trying to extract from the tech and not get sidetracked by its inherent capabilities.
At United Robots, we’re testing a number of possible uses for large language models, including prompting them to turn text into structured data (our “raw material”), also attempted elsewhere. It’s early days, there are lots of opportunities, and the measurable use and value we can derive from this tech is what will ultimately determine how we deploy it.
Good journalism is about people – those who produce it and those who consume it. It’s about the unique work and voices of great reporters, something that can’t be replaced by ChatGPT. It’s about meeting the needs and expectations of readers, in a way that differentiates yours from other publications. Large language models are not able to work out what your unique product should be.
AI can help improve our work processes, but it cannot produce journalism.
Let’s not have an identity crisis.
Rules based AI vs generative AI
> Rules based AI. In terms of generating automated fact based text, you can only really gain efficiencies by using rules based AI (at least at this point in time). With the United Robots platform (left in illustration) – and our model where we sell content rather than deploy tech – the all-important human(s) in the loop are involved before the text generation. The raw material is structured data from reliable sources.
• Our team programmes the robot
• Word, clause and sentence alternatives are also written by humans. As well as translated if we’re building for a new market.
• When a new publisher is onboarded, this is also where they are involved. Our language team helps editors train the robot to the editorial guidelines and needs of their particular newsroom.
Once all that is done, the humans’ work is done and the robot automatically writes and distributes texts to whatever end points the publisher needs.
With the human in the loop before text generation you can create volumes of texts with guaranteed accuracy – as long as the data is correct. The content is scheduled, and publication predictable and can be automatic.
> Generative AI. By comparison, to the right in the illustration is what generative AI looks like for editorial quality text generation. The human in the loop needs to be involved in two sets of processes.
• Someone needs to create the prompts to drive the text generation, as well as optimise prompts to refine the outcome.
• Even after that, in order to guarantee accuracy and editorial quality, you also need to have someone check the facts, sources, conclusion etc in any AI generated text.
At this point it's hard see efficiency gains with generative AI for automated text generation.