How much AI is acceptable in journalism?
From fake press releases and content farms to non-existent experts and AI-generated media, the publishing industry’s credibility has been repeatedly tested in recent years. So where does AI find an acceptable place in this landscape?
While the practice of image manipulation predates the existence of image-editing software, it wasn’t that long ago that the existence of an image went some way to validating a claim of something being real. The term “photographic evidence”, after all, is based on this understanding.
An image of a person accompanying a byline was once seen to be sufficient in proving the person exists and that they wrote whatever followed. But in an age of convincing images of non-existent people created in seconds by tools such as Midjourney, a simple headshot no longer passes the test for the more scrupulous reader.
Manipulated images aren’t the only concern when it comes to authenticity, however. For publishers, it’s their entire processes that have come under scrutiny as they seek to diversify their revenue sources and the use of AI tools grows.
Recent scandals
A sense of distrust in traditional media has grown for several reasons in recent years, from the pressure of rolling news coverage and competition with non-traditional sources to Donald Trump’s scathing remarks about any unflattering coverage of his presidency. But this has been amplified by a series of scandals that some would argue could have been avoided, were solid journalistic practices followed.
Several of these scandals have been unearthed by British trade publication Press Gazette. The publication revealed back in January that UK media outlets have quoted over 50 experts who cannot be verified to be real individuals on more than 1,000 occasions.
These “experts” appear to have originated from PR agencies, who have flooded journalists’ inboxes with AI-generated press releases. Doubts were raised when experts quoted within them failed to respond to any follow-up enquiries from journalists – although by the time the practice was uncovered, many news stories that had made use of them had been published.
The aim of this practice appears to be to secure backlinks from respected publications to commercial sites quoted in the press releases, thus improving their SEO and, in turn, their commercial standing. With the rise of publishers now using AI tools to generate significant numbers of news stories on a daily basis, such material can be quickly ingested and churned out at scale, which gives rise to the farcical – and troubling – prospect of a non-existent reporter quoting a non-existent source in what’s otherwise considered to be a trusted publication.
This is a somewhat sophisticated example of how publications have been adversely affected by the use of AI-generated content, but even the use of these tools for more rudimentary tasks shows the importance of human oversight to ensure errors don’t reach readers. Last year, the New York Times reported that Bloomberg had had to correct several AI-generated bullet-point summaries of its news articles after they were found to have summarised the content within them incorrectly.
To some, this may be the inevitable consequence of changing commercial models. As larger organisations look to make cuts, the need to rely on newer tools and processes to streamline content production can lead to oversights in the editorial process.
Similarly, journalists displaced by these restructures face a balancing act. Many previously benefited from a reputable parent brand but have now pivoted to independent publishing through platforms such as Substack, where they must weigh retaining their credibility against producing enough profitable content to make the work sustainable. For some, the temptation to rely on these tools for more than just research will be a natural one.
Lack of consensus
Much of the conversation around these issues stems from there being no consensus on acceptable uses of AI tools, and regulation taking too long to catch up with the pace of technological development. At the same time, people are finding growing utility with AI tools for day-to-day tasks, particularly where fact-checking is less of a concern. Indeed, in many areas, its use appears largely innocuous.
In journalism, while the use of AI tools for reporting may not be commonplace, or in line with the guidelines that are starting to emerge from major publishers, these tools are now widely used for research, transcribing interviews, spellchecking, and so on. And it’s reasonable to assume that many people would be happy with them being used in this way.
A similar division can be seen in audio content. An AI algorithm tasked with creating a playlist of songs within a user’s preferred genre, for example, is unlikely to upset the average music lover, particularly if this exposes them to new artists. But as AI usage becomes more prevalent in radio broadcasting, the trajectory of this is obvious. Several radio stations now feature – in some combination – AI-curated playlists, AI-synthesised presenters reading scripts generated by AI tools, and even musical output that is entirely AI-generated. A commercial radio station that combines all of these, and dispenses with human intervention entirely, does not appear to be far off.
In this specific case, it’s the element of human speech that moves things away from simply selecting music to somewhere more controversial. Just as with headshots accompanying bylines and social media handles, it’s the implication of a real human’s involvement that signals trust and credibility (and, in journalism, the ability of a person with a view of some sort being able to defend it when challenged). On the radio, human involvement in a fleeting piece of content is harder to ascertain than in print. But in the case of the printed word, is it acceptable to use a real person’s byline next to AI-generated content? It’s reportedly already happening at one publisher – at scale.
Perhaps the answer depends on the subject being discussed. Some topics should be obvious no-go areas for AI-generated content. An AI-written eyewitness report from a conflict zone, for example, would be an egregious misstep for any publication. One would expect the same judgment to apply to interviews with prominent figures, although the saga around German magazine Die Aktuelle’s AI-generated interview with Michael Schumacher shows that we can’t always assume this will be the case.
Explainer articles, which may be based on entirely human-generated content, arguably exist in something of a grey area, although perhaps more toward a darker shade of grey, as CNET found out, when it quietly amended 41 error-packed articles of this kind that were written using AI.
But, if an AI agent has sufficiently accurate information to describe, say, a travel destination, driving conditions, or the weather to a level a human might – and assuming we’re not expecting it to imbue this with anything too creative – is this acceptable? Or is it only acceptable if thoroughly checked by a human before publication and bylined in a way that shows both have been responsible? Perhaps joint bylines of this sort is where we may be heading in a bid to increase transparency.
The invisible hand
It’s easy to overlook the fact that while LLMs may be generating their own content, the training data used to inform them will have been selected by humans. The fruits of this frequently go viral on social media, where users claim to feed the same question to different LLMs to compare their responses to detect ideological leanings. It’s an interesting experiment, and one that naturally receives a great deal of interest, given the lack of visibility the average person has over the training data behind these systems.
But, sometimes, these things head down a more extreme road, and remind us of just how much these technologies – which are, of course, used by publishers as well as individuals – are in their infancy. The issue of chatbots providing racist, sexist or otherwise discriminatory answers – and, at a more absurd level, blackmailing engineers who say they will remove it or threatening car owners with keying – have been noted many times.
Quite how frequently the average user is likely to encounter these kinds of situations isn’t clear; headlines are generated by the most flagrant examples, rather than everyday AI discourse. But it does at least underline the importance of the processes humans have in place to deal with broadcasting more problematic language or content. There can be very real financial and reputational consequences when a journalist or broadcaster falls foul of a code of conduct. But what about an AI tool? Will people be as prepared to write off a chatbot they rely on if it’s just a case of releasing new code to ensure it doesn’t happen again?
Is the problem AI? Or a lack of transparency?
Considering the growing incorporation of AI tools within everyday programs we use; the fact that publishers are today reliant on a broad pool of freelancers in addition to editorial staff; and the presence of computational photography in cameras and image-editing software, it follows that there is no easy way to eradicate the use of AI in editorial publishing. A hybrid model that combines the two seems to be the only plausible outcome.
But whether this is a bad thing for journalism depends in large part on how it’s used and the audience’s awareness of it. And as opinions will differ on the first of these, the second is crucial.
It’s this principle of awareness that underpins the Content Credentials pin. Dubbed a “digital nutrition label for content” by Adobe, rather than attempt to grade a piece of media in any way for its veracity, it gives users specific information on what was used in its creation – including AI tools, such as Adobe Firefly – so that they can come to their own conclusions about a piece of content. It also sidesteps the question of who fact-checks the factcheckers, which is commonly leveled at initiatives designed to verify content, rather than simply detail its provenance.
At a time when building trust with audiences is more important than ever, the brands who make the most effort with transparency and the use of AI are likely to prosper. Perhaps it’s less about whether AI is being used and more about the lack of transparency into an organisation’s acceptance of AI tools in its processes that’s the real issue. Major publishers, including Reuters, the Associated Press, New York Times, BBC, and the Guardian, have made their policies publicly available to help foster trust and explain where exactly they draw the line.
Predictably, much has been made about a swing back to authenticity as audiences become accustomed to the polish of AI-generated content. But where it is commercially unviable to do so – more transactional information, for example – we should expect a level of AI to remain.
That might trouble some, although the assumption that a human’s involvement will automatically be preferable or more beneficial than a machine’s becomes harder to defend. The fake expert scandal mentioned at the start of this article shows humans to be imperfect judges when it comes to credibility. Indeed, there’s an irony to consider in that, under the guidance of a human, a capable AI tool tasked with verifying this information could have prevented it from happening at all.