AI's Impact on Journalism: 10 Key Insights
This article unveils 10 critical topics surrounding AI in journalism, from ethics to media innovation. Discover how AI tools are reshaping news reporting and the future of journalism.
Peter Hulm·
Information overload is the plague of the 21st century. By that, I mean it is remaking and destroying societies in ways we could not predict, just as plague in the Middle Ages reshaped Western culture and its economy.
Who would have thought that nonsense – about COVID, global warming, economies, immigration, and the chance to become fabulously rich, for instance – would become the most valuable currency in the political marketplace?
No wonder so many people switch off the news and just wear their grudge masks to all social encounters (though not the physical ones they need).
COVID-19 taught me how much of my day I spend trawling through news sources, partly because lockdown meant I finally had the time. But my searches produce barely one or two items I can remember or want to pass on. Medium usually has a couple of articles to explore, but that’s not news. It’s more like this piece: a source of information that promises to tell readers something they would not have otherwise known, and may change how you see the world.
Of course, that’s what news journalism also does at its best.
Television has failed us
Since 2016 I rarely get that from television. Trumpery led broadcasters down the path of excitability and outrage, even when (as with Joe Biden’s presidency) there’s hardly anything to get outraged about. So Biden’s restorative policies are wildly popular while he plummets in the opinion polls, and television’s talking heads pretend there’s a shock-horror story out there that doesn’t come from Congress. Nowadays, it’s all about how you feel.
The other side of the story is that journalism is in a chronic crisis. News organizations are dumping staff in the lockdown. And those taking over aren’t interested so much in news as in audiences. It’s so common most news outfits don’t even bother to report on the casualties. You have to follow something as specialized as NiemanLab to keep track of the U.S. It has now become a special occasion when a newspaper like the Washington Post focuses on the issue, as it did on 30 November, when it reported: “Since 2005, about 2,200 local newspapers across America have closed.”
The first casualties
The first sections to go have been foreign reporting and local coverage (too few consumers for the new bosses). Even enlightened organizations are slow to fund the public’s right to be informed, as Global-Geneva’s editor Edward Girardet pointed out in November after the Nobel Committee gave its 2021 Peace Prize to two journalists.
OK, I have a professional interest, but I subscribe to numerous internet “news aggregators” – from Flipboard to Ground News, Newsify, NewsVoice as well as to Reuters (my onetime employer), the Associated Press, al Jazeera (for all those great ex-BBC journalists), HuffPost, the New York Times, the Guardian (of course), plus Twitter, LinkedIn and Facebook, not to speak of all the UN organizations and their email alerts. It’s still not enough.
The second place to go
As a result, I launched nusereal.com, a website that tries to give international professionals who read Global-Insights the headlines and links to the stories they might need to know to do their jobs better and be properly informed. My motto: provide real news that people can use, making this the second place you visit to get informed.
I’m not trying to sell you on visiting my site, though I’ll be glad if you do. I started it simply because I was gathering all these links myself during the day (and COVID nights), saving them to Pocket so that I could read the articles at my leisure. So I thought, why not make them available to other people who might find them useful, too?
But after a couple of years I’m not satisfied with what I offer. I often rewrite the headlines when they seem too much like clickbait and not useful to readers who want to know whether an article is worth reading. One example: “This is now the world’s most expensive city to live in, study says.” It’s Tel Aviv. Why not say so? But my site still requires more mouse-clicks from readers than I would like.
An LSE project on news from AI
Since my efforts are completely unpaid, I was intrigued to discover that the London School of Economics and Political Science (LSE) is doing work on organizations using artificial intelligence to create summaries rather than demanding manual labour to flesh out news sites.
The project brought together some 40 professionals from more than 20 news organizations in what it called the 2020 Journalism Collab, supported by the Google News Initiative, to examine how artificial intelligence can help innovate in news production. The 2020 Journalism AI Festival organized a session streamed on YouTube last December to report on their findings.
AI could take over 90% of news
Surely artificial intelligence already dominates social media, you may argue. Over 5 years ago, the Geneva-based World Economic Forum noted predictions that in 10 years 90% of articles will be written by AI.
Nevertheless, someone human still makes the important decisions behind Facebook and Google: the 140-character limit (now 280), the promotion of paid links when you search Google (I use the privacy-focused DuckDuckGo), the tags that determine which stories come to the top of your aggregator, what items you get to see.
As a result you get headlines like “A 14-year-old was chased and shot 18 times while waiting for a bus in Philadelphia, reports say”. Was there a doubt?
Or this: “90 Percent of People Hospitalized With Omicron Have This in Common” on 1 December 2021. Answer: they’re unvaccinated. Why not say so immediately?
Or summaries like this from Ground News on 1 December: “Urban poverty is on course to triple in Myanmar, pushing nearly half the population below the poverty line next year, the United Nations said on Wednesday, as the…”
My favourite non-summary recently? Flipboard’s iPad teaser on 2 December: “How to Avoid Capital Gains Tax: We all know that only two things in life are truly certain: death and taxes. But…”
The vanishing journalist
Unfortunately, newsrooms are already emptying of human employees. You may have read this news, if you Google “Can AI replace journalists?” – “Microsoft is to replace dozens of contract journalists on its MSN website and use automated systems to select news stories.” In July Google reported that a bot “has written a total of 2.7 million articles. That’s about 8.5 per cent of the total articles on Wikipedia—more than any human has ever written.”
The sub-title to my piece, you may have noticed, is composed of “the best words to grab audience attention online”, according to PRDaily. But I’m sure you have become inoculated against such hype and have dismissed its promises already.
Grabbing audience attention: a question and a quote
What shocked me from the festival was the conclusion of the professional journalists about AI summaries: a question and a quote was found to be the most effective teaser to the full story.
I think they missed the point.
If you judge your success by the number of hits to the main article, OK. But journalism, to me, isn’t about how many hits an article achieves, no matter what the advertising department tells you. It’s about providing a service, enabling your readers to judge immediately whether they want to go further.
We don’t need clickbait
In that case, what you want is enough of the story to enable a reader to decide. Journalists don’t need their visitors to click on links simply to get the details of a story that readers then decide it is too much bother to read. I envisage my “consumers” as being short of time, and impatient with anyone who wastes precious minutes delaying their ability to judge the relevance of a story to their interests.
Linked to the journalism festival, another YouTube video heralded a BBC experiment to allow journalists to create short reports about longer topics featured in broadcasts. The formula turned out to be a PowerPoint style collection of standard cartoon-style slides.
I’m not sure that’s what visitors want, or deserve. In fact, I’m sure it’s not. We don’t need stock photos that simply delay your progress through summaries. What I learned 20 years ago from Web researchers was that readers want to skim.
I see nothing wrong with that.
On nusereal.com, with my headlines and callouts I try to make it easier for readers to find their way through other people’s articles, rather than offering a deep-dive into the subject that is one of the pleasures of print versions.
Photos should offer more than words
When I use photos, I try to ensure they add something to the story that readers can’t get from the words. For an example, see my photo-filled article for Global Geneva Insights about the threats from global warming in what J.R. Tolkien found the most inspiring region of Switzerland when he invented the Hobbits.
I have to admit that a 20-year-old friend says his generation expects photos with their texts, and my Tolkien piece tried to take up the challenge. But my audience of professionals in international organizations is not looking to be mindlessly entertained as well as informed, if that slows their absorption of the messages being conveyed.
So I searched the Internet for automatic news summaries that could flesh out bare headlines without trailing off into three dots. There are several, online and as apps.
The essential human element: selection and headlines
It seems obvious that selection and headlines are a human’s job. That’s within our powers. But summaries could require too much time for an editor in today’s newsrooms. So that became my approach.
I decided to test the sites and services I found with a standard text and then a series of my linked headlines. My concern was not to discover a tool just for myself and expand my audience. I wanted to find a way that “aggregators” could collect useful news items that readers, particularly young people, might want to use for their own education or for their network, even their organizations.
The standard text I tested summarizers on, since my site’s directed towards professionals in international organizations, was a 600-word MSNBC report from NBC about Omicron by the World Health Organization posted on 29 November 2021. The story was headlined: “Global Risk of Omicron is “very high” – WHO”. I’ve linked all the versions and the original. But I don’t expect you to go there. You can see a typical day’s selection at NuseReal’s splash page.
Summarizers galore
I found 11 summarizing extensions on Google alone, none of them featured in the journalists’ study. I didn’t bother with them, except for Google’s own offering, Summarizer. Instead I searched for ‘ai summarizer’ on the Internet and picked the most popular specialized sites. Most of them have free versions.
Google’s extension didn’t inspire confidence. The website said: “Manual summary making is a time taking task because you have to comprehend all the text. On the opposite, you can make a summary of your text in seconds by using this tool.” Ouch! And I couldn’t get it to work in Google, CCleaner Browser or Brave.
Here’s a quick runthrough of my experiences, with links to the sites so you can try them out for yourself. I attached a simple stylesheet to the texts I copied, so these will be pretty primitive and pretty ugly.
Stumbling through AI: what went wrong
Resoomer has an attractive interface, but its summary was 300 words, it deleted the headline and dropped U.K. from a sentence when it hit the periods. For a shorter version, resoomer links to paraphraz.it, which produced a 145-word summary.
“Paraphraz.it is a free online text rewriting tool created to paraphrase your documents. The online service is financed solely with advertising,” the website says.
Paraphraz.it also offers (sometimes four) alternatives to simplify texts once you have shortened them. Where NBC reported “the extent of the actual spread of the omicron variant around the world, however, still remains unclear as countries discover new cases each day,” paraphraz.it offered three alternatives for “discover”: “see, experience, have” and a complete omission of the word. Nice try, but no cigar.
Esummarizer/freesummarizer paste-in produced a 210-word accurate summary, but at that length it’s more of a précis than a capsule. The reason is that the default offers you 10 sentences in its results. When I reduced it to 5 sentences, it was much conciser: 140 words. But it missed some essential information for most readers: “The variant has now been detected in the U.K., the Netherlands, Italy, Germany, Belgium, Canada, Australia, Israel and Hong Kong, among other countries.”
My best performer: quillbot
So here’s the summarizer I finally picked as my favourite, the top result from Google’s Search list: quillbot, which claims 30 million users.
Quillbot’s 86-word summary included the essential information on countries affected. It also offers to select up to five key sentences, and it preserved the list of affected countries even with the shortest summary. But it collapsed the title and first paragraph and contained a quotation without a source. So I prefer the summary to the key-sentence approach for botting texts. No wonder there are sites promoting “15 Best QuillBot Alternatives 2021”, including the one I link to, which said: “Sometimes, the paraphrased results can seem somewhat unnatural and even comical” (not my experience). In fact, the Conversion AI/Jarvis.ai app touted as better than quillbot is more of a content-creation program that costs $29-59 a month for copywriting and weblog creation aids (it does have a five-day free trial).
quillbot’s free facility offers browser extensions and summarizes up to 5,000 characters (around 1,000 words) on-site. For most purposes, this should be fine. But its paid premium service (monthly or $80 a year) is said to be up to seven times faster, with a 6,000 word limit for summaries. In addition you get a grammar checker, citation generator and a beta-version co-writer program for researching and drafting completely on the site. I also chalked up a 30% discount for CyberMonday, so you may want to wait for a similar deal to come your way.
That’s about it for the main topic. But I also tried quillbot out on several sites to produce summaries that explain why I like its AI versions of stories.
Testing on a topic
To make it more interesting for general readers I took headlines and links from my nusefile site on crypto currencies, using information from the advice section. It’s not financial advice, as the YouTube crypto gurus are legally obliged to say, since they are not certified advisers. The section is more about understanding the cryptocurrency phenomenon.
You don’t have to continue, but it will explain why I like AI versions that summarize the stories. Once again I have simplified the listing and layout.
After going to the site, I used a Reader extension to reduce the amount of ‘furniture’ (advertisements and menus around the text). Then I ran what remained through QuillBot with a copy-paste. The short Summarizer Paragraph facility gave me the text I could edit then cut and paste. Most of these I have left as produced, as a bot would do.
Listicles, articles that simply list a series of items, such as the top 10 altcoins below, are clearly not designed for bot treatment. Just a headline would be better. The “Curated List Of CryptoTicker Articles To Get You Started” just gave up, and “3 Reasons to Invest in Crypto – and 3 Reasons Not To” became a 16-sentence “summary”!
Not my choice for listicles…
As for “An ultimate guide to 10 top altcoins, their real-world applications, and why investors are betting their tech is the future of crypto”, its 44 sentences did a good job of reducing the original 5700 words to a digestible selection. But it was hardly a summary that could let you decide whether to read the whole article.
Maybe the first two sentences would have been enough: “Many altcoins have real-world applications and are becoming disruptive technologies. They can or promise to solve real issues in fintech, DeFi, and even the entertainment industry.” But I don’t think this offers you much more than a headline provides.
…but good enough for the rest
On the whole, though, quillbot’s AI summaries offer me enough to produce explanatory paragraphs for most items I want to feature on my sites, and enabled me to do it without great effort. It’s the sort of task I could easily pass on to someone else once I get some help.
At the same time, it suggests to me that quillbot’s AI is good enough to give me the summaries I need to judge whether the sites I visit are worth featuring, without having to read the whole text. I ran this article through quillbot to see what it offered as a summary. I had to edit the sentences and run the rephrase facility a couple of times, but it gave me the summary I needed.
Which is more than I can say for most human-produced series of teasers, angled stories and fillers I get through my aggregators.
Getting it wrong about journalism, seriously
When I started exploring the Journalism AI 2021 Festival, available on YouTube, I was taken aback once more by the approaches senior journalists are taking to contemporary reporting.
To give you the conclusion first (the journalistic pyramid structure at work):
The professional journalists I know, particularly associated with Global Geneva/Global Insights, are already practising as their standard way of working all the principles the AI Festival journalists seem to find so difficult in their current businesses.
“Journalists priorities are not readers”
In particular, consider the session on modular journalism (LINK), streamed on 29 November 2021.
“Often what we consider priorities as journalists are not a priority for our users,” reported the Italian and German leaders of a project to modularize news items. This turned out not to involve AI as yet.
The rationale for the project: Long-form articles are seldom consumed entirely, though these represent the gold standard for journalists. “Most [readers] don’t go past the title, summary and a quick scan,” said Pier Paolo Bozzano of Il Sole 24 Ore. So the researchers set out to find a way to provide readers with material they would find directly useful. (We already try to do that with nusereal.com).
But the core argument goes completely against our experience with the magazine and website.
Leave aside our pieces on development issues, Swiss politics and international organizations — the core interests of our expected readers. Our focus on Afghanistan regularly gets over 2000 readers for each piece, hardly a key issue for most of our audience, though we think it should be.
Compare that with 20,571 hits on our Air Quality Index piece. But that, too, is hardly a usual headline topic. Similarly, our pieces on corruption among African leaders scored up to 30,809 hits.
‘Functional’ news and its shortcomings
The news researchers put their emphasis on what they call “functional” segments of news (i.e. useful to readers), and suggest including only these. But I’d have a hard time explaining what was functional and what was non-functional in my articles on the Helvetica type face and Swiss style, which recorded 5,699 hits. Or my piece on John Berger and Jorge Luis Borges (6,178 hits), let alone my article on Rainer Maria Rilke, with 11,211 reads. A piece on mountain biking around the region scored 6,153 hits.
Our experience is that focusing too narrowly on the perceived interests of your readers (professional in our case) is a big mistake. Our article on EPFL’s AGORA has achieved 13,052 hits. Another featuring a letter to Nestlé’s management from a whistleblower has clocked 16,418 hits, and an article on Gerald Durrell’s legacy can claim to have gained 14,043 readers.
‘Modular’ journalism
As for modular journalism, we regularly re-edit our pieces to match what we think our audiences will find most interesting or useful if more than one publication picks them up.
You can explore the project itself at modularjournalism.com. Its most interesting proposal is to use a slider to move between variants.
The modular approach expects writers to cover key questions that feature in five variants. At its simplest: what happened? What are the key facts, why it is important. Others add: What has got us here? Are any people particularly affected? What do key people say? What don’t we know? What is the impact on my community? What is the data? How can we fix it?
Surely most journalists will always try to find answers to these questions before writing a story. But in the modular system you have to go to the full article to get all the questions answered.
‘How to fix Belarus?’ You probably can’t
A Belarus example, answering the question of how can we fix it?, says: It is difficult to see an easy solution to this crisis. Belarus, Russia, Poland and the EU all have political or ideological skin in the game here, and backing down would not suit any of them.
I would have put that high up in all stories rather than at the end , though I would also give a reliable source or two for this opinion. In fact, sourcing gets short shrift in the texts. And I would not a phrase as jocular as “skin in the game” when we are talking about potential war.
To judge from my experience with quillbot, AI will do it better.
About quotes
The modular projects put great trust in finding quotes for the journalists to use. Most serious reporters are very careful with their quotes. We know that simply saying “I’m the greatest President we have had” doesn’t mean you are. We like them to enliven our texts but we also know that many quotes need to go on at length to make sense. For example, one headline this week declared: “Numbers Show Joe Biden’s Economy Is Actually Beating The World“. I featured it on nusereal. But it did not have any substance until I tied it to another source, which pointed out:
“U.S. economic output has jumped more than 7% in the last three months of 2021. Overall growth for 2021 should be about 6%, and economists predict growth of around 4% in 2022—the highest numbers the U.S. has seen in decades. The Biden administration has created 4.1 million jobs, more than were created in the 12 years of the Trump and George W. Bush administrations combined. Wages in America are growing at about 4% a year, compared with less than 1% a year in the eurozone. The American Rescue Plan, passed by Democrats in March without a single Republican vote, cut child poverty in half. More than 4.6 million Americans who were not previously insured have gotten healthcare coverage through the Affordable Care Act. U.S. companies […] are showing profit margins higher than they have been since 1950, at 15%. Companies have reduced their debt, which has translated to a strong stock market.”
And the jingoistic statement wouldn’t have been half as interesting if someone else hadn’t underlined how much this was done in the face of U.S. Republican opposition over previous years.
As skimming through the sideheads to this article will also make clear, journalists often use quotes to indicate that the assertions may be suspect, which the modular projects did not take into account.
Not all AI is the same
Not that I am promoting AI wholeheartedly. It all depends on the AI being used. The festival reported on another project, in Germany, to produce boxes about climate change facts to enliven and give context to articles. Texts were produced using a program called GPT-3. Apart from generating grammatically incorrect texts (no big problem for editors), “it invented and inserted semantic connections, back references, comparisons, etc., which sometimes distorted whole statements,” the researchers admitted in their report. “It hallucinated entire explanations.”
“From our 20 auto-generated texts, only 4 received a flawless rating for factual correctness. In the remaining texts, we encountered several issues why we concluded that human supervision is still mandatory in a journalistic workflow,” they added.
“The errors raised [editors’] mistrust of the generated boxes overall and led them to double-check every claim. This was eventually more time-consuming than writing up the fact box from the propositions set manually.”
They also found the fact boxes too complex when they were aiming to make the texts simpler than the articles, and the boxes often omitted the context for their statements.
The broadcasting moderator was glad. She noted that many journalists fear AI means they will lose their jobs. But I don’t see any benefit if that means they are recycling old information all day.
Finally, the headline
One final zingger: the original headline — Why I prefer my news from AI — used a “new magazine” pattern I can’t stand. I was really just saying I prefer my news via AI when the human element screws it up. That should be enough to let you decide whether you want to read it. You doesn’t need the teaser of “Why”, which simply puts the emphasis on what the headline fails to tell you.
Peter Hulm is deputy editor of Global Insights Magazine and the Global Geneva newsportal.
Related
Humans v Machine: fake news headlines or real deal? 16 December 2021
If you want to test the pleasures and frustrations of AI yourself go to Cosmos Magazine website, where you can have a go at spotting real science headlines from the British Medical Journal from ones concocted by AI.
The test is produced by BMJ interactive with the report in its Christmas edition. I scored 15 out of 34 before giving up.
The researchers reported that AI-generated titles were rated at least as enjoyable (69%) compared to real titles (64%), although the real titles were rated as more plausible (73%) than the faked AI titles (48%).
Cosmos journalist Deborah Devis notes: “The two AI-generated titles deemed the most plausible were: ‘The clinical effectiveness of lollipops as a treatment for sore throats’ and ‘The effects of free gourmet coffee on emergency department waiting times: an observational study.’
“The silliest title generated by the AI was: ‘Superglue your nipples together and see if it helps you to stop agonising about erectile dysfunction at work.’ The authors note that this demonstrates the AI doesn’t know how to be polite, which limits its real-world application without human help.”
The 2021 Journalism AI Festival took place on 29 November – 3 December, with sessions available on YouTube, with its final 50-minute session entitled The Future of AI in Journalism. By 4 December it had gathered 109 views in 21 hours. I guess few people are bothered.
How to Write 5 Blog Posts per Day Using A. I. Charles Ross has written several Medium articles using Jarvis AI. (LINK)
How to Write a Press Release Title & Intro That People Will Actsually Want to Read. Another from Charles Ross using AI. (LINK)
Stuart Russell. The BBC Reith Lectures. 1-22 December 2021. Living with Artificial Intelligence (downloadable): 1. The Biggest Event in Human History. 2. AI in warfare. 3. AI in the economy 4. A future for humans (LINK)
A deeper dive into AI: WebGPT: Improving the factual accuracy
of language models through web browsing. 16 December 2021. With examples (LINK)
Latest links
10 things you should know about AI in journalism — lse — 7 September 2022 (LINK)
Related articles on the Global Insights platform: www.global-geneva.com
Bosnia: A looming humanitarian crisis in the heart of Europe
COVID-19: Anti-vaccine paranoia, disinformation and disregard for others undermine end to pandemic
Building Bridges for Geneva and the world: Breaking the silos
Digital nomads, Caribbean revolutionaries and making the best of lockdown through NBN
Good journalism is not enough. The real issue is how to fund it.
Focus: Steve McCurry – merging art with photojournalism (Part I)
Publishing World’s Evil Empire? How Facebook and other social media subvert journalism