As ChatGPT has taken the internet by storm, becoming the fastest-growing consumer app in history with 100m users in January, a small swathe of publishers have either announced their intention to use “AI” in their businesses, or had their existing plans scrutinised.
At the end of January, BuzzFeed announced it would use technology from ChatGPT creator OpenAI to “personalise” and “enhance” its content and quizzes, including helping staff to brainstorm ideas.
Last week the UK’s Reach, publisher of the Daily Mirror, Daily Express and dozens of local papers, told the FT that it was reviewing how ChatGPT could “support our journalists for more routine stories like local traffic and weather or to find creative uses for it, outside of our traditional content areas”.
Less positively, it was revealed that the Red Ventures-owned CNET – which runs a range of brands focused on helping consumers make informed purchasing decisions – had found errors in more than half of the articles on personal finance it had been quietly-producing for months using.
The CNET story gets at one of the core challenges around these models. As we said at the start of the year, it’s worth remembering that ChatGPT and its ilk are not the thinking AI of movies, but language-learning models trained on huge volumes of content found on the internet. While they are good at finding information and presenting it in a plausible way, they remain extremely poor at ensuring accuracy, which takes human intervention to correct.
That issue has been underlined by the rapid rollout of the technology to search by OpenAI-investor Microsoft, which uses the same tech driving ChatGPT to power its chatbot Sydney, and Google with its own Bard model. Both have been shown to produce false answers, in the case of Sydney often accompanied by a quite terrifyingly human-style refusal to admit they could ever be wrong. These search-assisting chatbots may come with a disclaimer saying that people should use their own judgment and research. But the danger is that, seemingly as with what happened at CNET, people take the well-crafted but fundamentally flawed outputs as the gospel truth they pretend to be.
What then is the potential business use for for publishers who – at least to some extent – base their brand on accuracy?
Well, there are some areas where it can clearly be used to replace the more mundane tasks currently given to journalists. After all, Bloomberg has been publishing automated short news stories on companies and other financial market information for years. When speed is so important, and the information often well-structured and easily predictable, letting a computer do it is far more efficient. When you need the information in seconds, only a computer will do.
There are other similar areas, such as the weather and traffic information cited by Reach, where these new language models can put similar levels of automation more easily in (mmm) reach.
According to Charlie Beckett – director of the London School of Economics’ Polis media think-tank, which runs a journalism and AI project – publishers will have to make careful judgements about what they can hand almost entirely over to AI, what will need careful oversight, and what will always be the domain of humans.
He says: “I spoke to a German publisher the other day and he had completely revamped using AI. Some things hadn’t worked. They weren’t worth it. They were just too unreliable. You ask humans to do it. Some work really well, but they were always edited at the end. And, I love this, he said horoscopes are the best because you can’t get them wrong. There’s no such thing as truthful horoscopes. So that’s fine.”
Arguably these models may work particularly well for BuzzFeed, because of its DNA. At the peak of its powers, BuzzFeed was doing some of the best news and culture reporting around, but much of its work was at the start, and still is, based on pulling together what other people have been saying on the web. Is there a “truthful” version of the “10 most under-rated sitcoms of the 1990s” that isn’t simply trawling what people have been saying on social media?
Other areas that look promising are on the commercial side of the business. As we wrote in January, the potential for more automation in marketing messaging is very real for subs-based business, and it may well offer new ways to deliver customer service quickly and efficiently.
It could also help generate some forms of tailored advertising copy and sponsored content, where much of the job is re-purposing existing copy to fit new formats or space, rather than synthesising an accurate article.
In an ideal world, it will – as Reach has hinted – help media organisations develop new products, particularly around data. Learning models are very good at finding and structuring information, and tailored versions could make building new data products that source say, real estate information, far easier.
Bespoke versions trained on a publisher’s own content could also help replace the laborious task of pulling together and updating previous content into new forms, something most journalists will have had to trudge through at some point in their lives. Re-working content generated nationally for local audiences is something many regional media groups get humans to do, but it is perfectly conceivable that machine learning models could do this just as well.
As Beckett says: “So much of it is looking at agency copy, looking at what’s out there already, refashioning what your colleague wrote earlier in the morning, you know… or reformatting for different platforms, putting it in the newsletter, and especially as the thing gets more used to you and your particular dataset.”
Will language learning models deliver huge cost savings? The answer is likely not that simple. Beckett compares the impact of digital technology – which saved huge amounts of time particularly in broadcast – and the rise of social media, which may have opened up new ways of sourcing and distributing news. But, if anything, it actually added new costs for media businesses: “Someone’s going to have to do the tagging for all this data. Someone’s going to have to make sure the applications are appropriate to your newsroom. Someone’s going to have to check that it’s not going crazy putting out rubbish.
“On the other hand, it can make big savings. For example, if you are working through great big datasets. If you’ve got a series of platforms that you want to put content on, regularly reformatted for, and indeed it might produce extra stories or extra opportunities.”
A key thing to remember is that not much of this is going to happen quite as quickly as the hype caused by ChatGPT might suggest. Especially if publishers hope to avoid missteps to which audiences will now be far more attuned.
Beckett adds: “One of the things we always say to people is, you know, start small, start slowly, make sure you actually know what you’re talking about. Think about the problem you’re going to solve. Don’t just rush in because you think it’s really trendy.”
Those publishers who use language learning models carefully and cleverly could make efficiency savings, where they can take over mundane human tasks, and also create new products and content. Those that rush in recklessly are just as likely to do lasting damage to their brands. Start learning now.