I see a lot of claims that generative AI is super-fantastic and will “change everything”; I also see claims that generative AI is either useless or disastrous. For what its worth, my opinion is that its a very useful tech which will have a real impact on productivity and quality of life, but its not going to “change everything” or destroy civilisation. Also impact will take time. This kind of perspective doesnt seem to get much attention in the media, perhaps because it is, well, boring,

So I was very happy to read the report Generative AI in Journalism: The Evolution of Newswork and Ethics in a Generative Information Ecosystem (http://dx.doi.org/10.13140/RG.2.2.31540.05765). It basically surveys how 292 journalists and other news professionals are using LLMs and the challenges they face. It provides a lot of interesting real-world insights, which I suspect are applicable to many other professional uses of LLMs, and it more-or-less aligns with my above-mentioned views.

Incidentally, the lead author of the report, Diakopoulos, also edits a very interesting blog on generative AI in journalism, and wrote a book in 2019 on Automating the News.

I highly recommend the report, and below summarise a few insights from it which I thought were (a) very interesting and (b) applicable to other uses of LLMs in professional contexts.

Biggest use of LLMs is helping to write text

The first point is simply that the biggest use of generative AI in journalism is NLG, ie helping to write texts. Maybe this sounds obvious, but I attended a conference around 8 years ago where some media organisations talked about how they used AI, and there was almost no mention of using NLG to help write stories. Instead the focus was on what the above-mentioned report called Information Gathering and Sensemaking (eg, using text analytics to analyse stuff) and Business Uses (eg, finding potential advertisers). Obviously this kind of thing is still very important, but nice to see that the NLG usage is now the most common!

The report also pointed out that most people are using LLMs to improve existing workflows and products, instead of introducing new ones. Which is not surprising. However the biggest impact of new technologies is often in new services and products. For example, the Web started off in 1990 as a better method to allow researchers to share information (improve existing workflow), but of course in 2024 its having a huge impact in areas (such as e-commerce or social media) which didnt exist in 1990. Similarly I wouldnt be surprised if the biggest impact of LLMs is in enabling new kinds of journalistic activities and products, but these will take time to emerge.

Workflows

In (reputable) journalism, LLMs dont write articles autonomously, instead journalists work with LLMs in “human-in-the-loop” workflows. But what does this actually mean? The report mentions several kinds of workflows which could be called “human-in-loop”, including

• Humans use LLMs as brainstorming tools
• LLMs write draft which human checks and edits
• LLMs check/edit what humans write
• LLMs do support tasks such as transcribing audios

In other words, there are lots of ways in which humans and LLMs can work together in producing text. Which is hardly surprising, but I dont see much discussion in the NLP literature about different types of human-in-loop workflows, and indeed the kind of user-interfaces (and user training) which would best support the different workflows.

The report also points out that LLMs change the skills needed by journalists (eg, prompt engineering) and also how human interact with other humans. For example contributions from freelancers need to be checked to see if they were produced by LLMs.

The above means that journalists and other news professional must change the way they work in order to make best use of LLMs. This is hardly surprising, but change management is difficult and takes time, which is one reason why impact will take time.

How effective are LLMs?

The report does not give quantitative data about the effectiveness of generative AI in actually boosting productivity or otherwise helping journalists; the survey did not specifically ask about this. But it does include free-text comments about this from survey participants, which are mixed; some participants said that trying to use LLMs cost more time than was saved, others said that LLMs really helped them.

I would have loved to see hard data on where LLMs actually helped and where they did not. I appreciate that measuring real-world impact is difficult and can be commercially sensitive (blog), but it would be great if someone did this and published the results!

Ethical issues

The report contains an entire section on ethical issues. The most common ethical concern was that LLMs would be used without human supervision. No details are given, but I appreciate that cost and time pressures may lead organisations (and not just in journalism) to use LLMs without supervision even when this is not appropriate. Which is worrying! If LLMs can only be safely used with “humans in the loop”, we need mechanisms to ensure that this is indeed how they are used.

Not surprisingly, there was also a lot of concern about bias and inaccurate information, and indeed many other ethical issues.

Final thoughts

The report looks only at generative AI in journalism, but I suspect that the above findings apply to most usage of LLMs in professional contexts. They are good at generating texts, but their biggest impact may come from novel applications. There are many human+LLM workflows, and we need to understand which is most effective in principle, and how difficult it is for the relevant professionals to adapt the new workflows. Despite all the hype, we still lack solid data on effectiveness of LLMs in professional contexts, perhaps because this information is commercially sensitive. And one of the major ethical challenges with LLMs may well be stopping cost-cutting organisations from using LLMs in unsafe and inappropriate ways.

None of the above is rocket science, and I suspect many of my readers will have similar opinions about LLMs. I would love to see more work on the above issues, which to me are more interesting and important than debating whether LLMs will revolutionise or destroy society, and indeed evaluations of LLM tech which ignore the above issues.