Modular Journalism 2.0

Modular Journalism 1.0

The problem at hand, then and now, is that the long-form article, the most esteemed of traditional journalism formats, is dead (user data and analytics at hand) to a large portion of our audience. To the point that we can be tasked to play Frankenstein with the corpse, chop it up, turn it around, and — if only for an exercise in storytelling — see if we can breathe new life into the format by reassembling the pieces in a different way. And, as we are doing that, to make it even more interesting, see what role an artificial intelligence could play in the process.

The approach, three years ago, was of course modular journalism, leitmotiv of that year's collaborative. Our take (like good old jeux sans frontières we were organized in teams and ours was team purple) could be summed up with a question: can the traditional long-form article be deconstructed in functional modules and reassembled in different ways to appeal to different users and audiences that we can't reach today? We were declaredly hunting for unicorns, from Gen Z to flat-earthers, in true collaborative spirit and moonshot optimism.

Luckily, cheap generative AI would not be available until the following year, so we had to conduct our many attempts at dissecting news artifacts using human intelligence and infinite patience. In doing so we learned that [1] it was hard work and [2] the dysfunctional nature of large portions of those long-form articles represented the principal hurdle in isolating modules for the purpose of reusing them.

The term dysfunctional was not chosen to pass a qualitative judgement of the artifact we were working with (although in many cases that argument could be made) but because we could only reliably determine that specific portions of the text were the vector of a journalistic function, while other parts seemed entirely void of journalistic function and, for the purpose of our research, could be safely discarded. This new experiment, riffing on the old one, aims to take a closer look at the discard pile to determine whether, in fact, function existed there as well.

The focus on information needs

User personas and user information needs were at the very core of the collaborative work and were, therefore, the key to deciding whether a text was functional or dysfunctional.

While we brushed up on linguistics to better understand the news discourse and put our feet on a foundational ground, our primary focus remained to point straight to the users, determine what type of journalistic discourse was best suited to meet their needs and see if we could assemble, module by module, square peg by square hole, the very format that would do the trick.

We worked with a real-life array of information needs, compiled with internal research by the news organizations that were part of the initiative, but for the sake of argument, any list of user needs, compiled with whatever methodology, would adhere to a similar logic and match a relevant group of modules.

If a layer of automation were added, we could picture a two-pronged application: on one end, a machine-learning system tasked with determining information needs and sentiment; on the other, a language model leveraging modularity to serve highly personalized content to readers, each with their own preferences about the time, place, interests, and formats they care about. All this would not be happening for the sake of symmetry and science but with the declared intent to drive positive engagement and, with luck, subscriptions. A rare 'if' scenario that might pique the interest of marketing departments at several publishers. Or so we hoped.

In fact, the project ended there. Il Sole 24 Ore graciously hosts the modular journalism manifesto but modularity was not explored further. I know this continues to be an important area of research for Shirish Kulkarni, I am not sure about the other members of team purple. Truth is, modular journalism is not for the faint of heart.

The prospect of reusing and repurposing existing long-form articles — decidedly a square peg and round hole exercise — was quickly put aside during the collab for the reasons described above. Whether generative AI could be used, now that it's more readily available, to bring order to the chaos, that's an argument for debate and an interesting area of research.

Back then, the best way to test a modular journalism algebra was to reach for the low-hanging fruit, leave the long-form articles behind, and instead work with a modular-first approach. We whipped up a modular-first content management system, one that achieved perfect modularity right off the bat and left no room for dysfunction. Journalists were asked to file their stories in a modular grid, directly responding to information needs, that, in our case, were formulated as questions: "Why is this important?" "What are the key facts?" "Who is responsible for this?"

It was immediately obvious how modular-first journalism and ho-hum traditional journalism yielded very different results. One big surprise came when trying to match or place our list of information needs, derived from actual user research, within the categories of the news discourse as they were presented by Teun van Djik. Which, in turn, meant comparing the news report 'as it should be' by the book to what users told us they were looking for.

Van Djik describes a news schema as

"featuring the usual categories that provide the different functions of information in news reports: Summary (Headline and Lead), Main Events, Backgrounds (Context and History). Consequences (Consequent events or actions and Verbal Reactions), and Comments (Evaluation and Prediction). Some of these categories are obligatory (Summary and Main Event), whereas others are optional".

There is a very clear top to bottom progression in his approach:

"The general principle is that of relevance: The most relevant information (from top to bottom) comes first, followed by lower levels, and finally, details of each respective schematic category (from summary, via main events, through backgrounds to comments). Hence, an important verbal reaction may appear before a less important detail of the main event. This relevance structure is intricately linked with news production strategies, the structure of models journalists have of news events, as well as with properties of reading news such as skimming."

See the diagram below with a news schema paired with the information needs taken into consideration during the collaborative project.

It's a very imprecise match, of course, but the result was so lopsided that it's still worth mentioning. Van Djik's pyramid is essentially turned upside down when applying a modular approach. The modules that, according to the news schemata, should have more relevance were in fact less important for the users, and vice versa; modules that we are used to considering accessories were protagonists in the information-need-based stories.

Roughly half of all the information needs we considered related to Background and Context. A smaller, yet still significant, portion of information needs fit into the Comment section, either as voices of the protagonists or as Conclusions drawn from the Verbal Reactions. Compared to Background and Comment, far fewer information needs corresponded to the Main Event of the news report, the focal point of the Episode in van Djik's model. And even within the Episode, almost as a counterweight to the five Ws, we saw that quite a few information needs had to do with the Consequences of the Main Event: "Who pays?" "Who profits?" "Who is most affected in my community?"

Also, one should note, these "secondary" modules not only make up the bulk of the modular stories (which to some extent is true for van Dijk's as well), but are presented in prominent positions in the news artifact. Even Headline and Lead hardly make the cut, as they become redundant in terms of information needs. If they were maintained in the showcase of modular articles, it was mostly for the need to have something to 'click on' in the UX to get users to the stories and allow the primitive CMS logic to file them and generate URLs.

The role of Headline and Lead is in a way in contrast with an information-needs-based approach because we precisely don't want to give users, to put it in van Djik's terms, "the intended semantic macrostructure of the text." We instead would want the users to engage in "a bottom-up process," helping them to surface the information they seek with modularity and being transparent about the function of each portion of the text.

An inevitable segue is that a true CMS built for modularity should handle modules and relations among modules instead of articles, which cease to be unique and unequivocal documents dated in time and become algorithmic. This is particularly interesting because both search engines, with the integration of generative AI, and social media, through memefication, are progressively driving away from an article-first approach. Frontend applications should perhaps also rethink the way URLs are structured and used. In exchange for the extra work, a news product made up of algorithmic articles is much more difficult to be cloned, aggregated, summarized, or replaced with AI-generated snippets. To get a unique experience, that is, users must necessarily go to the source (and possibly subscribe).

So why don't we do it?

Again, the information needs in our pool were a narrow spectrum, far from statistical relevance and merely a hint. It's inevitable, however, to ask ourselves why we write stories the way we write them and don't instead follow a data-driven approach.

It's fair to point out that the user information needs, as we refer to them in modularity, should not be confused with the extraordinary User Needs model developed by Dmitry Shishkin. By user information needs, we refer to atomic, hypothetical needs users may have, and that can be expressed with a simple question that can be answered in a paragraph. In fact, the modules API has an endpoint to explore the possible synergies between modules and Shishkin's core needs. As shown in this very preliminary doodle.

So, if user information needs — the modular journalism kind — were, at least in part, the deciding factor in determining how a news report is constructed, then the resulting news reports would look significantly different from what they are today. Did sabermetrics (obsessive advanced statistics in sports analysis) ruin baseball or perfect it? Neither? Both?

A data-driven approach is applied to how we organize content in our news products, to how we market content, to how we price it and to whom, and to how we promote it on social media, so why would it not be applied to how we write stories in the first place?

Upside down stories

As I said, one of the most interesting aspects of looking at news artifacts through a modular lens and with information needs in mind is how far apart they are from the stories we happen to read every day. And yet how they make sense and are, frequently, an enjoyable read. The few we have read or written, of course.

A modular-first approach requires a certain amount of flexibility, and journalists need to buy into the process. Which happened only in part during our experimentation. We gave writers two possible ways to do it.

They could write using a form that largely mimicked van Dijk's news schemata, a traditional approach yet tightly confined within a grid of modules. It was not a blank page, but it allowed writing more or less as usual, with some structural constraints. A Summary, made of Headline and Lead, and the body of the story, composed of Situations and Comments, each with their matrix of expected and orderly development of submodules. Van Dijk was, above all, a placeholder for the rigorous development of the news discourse; it did not, in itself, carry a clear correspondence with information needs, which were our main focus. So we asked those who preferred to use this more traditional approach to manually link each module to one of the information needs the chunk of text corresponded to. So, instead of writing an answer to a specific question, they were writing their answers and then scrolling the list of questions to match the one that came closest.

A second approach was to use an alternate form which assumed certain information needs as core building blocks and asked writers to compose their story simply by answering to them. "What happened?" "Why is this important?" "What do key people say?" This approach had absolutely no strings attached to the past, and nothing was taken for granted. Does a story even need a headline? Or a summary? If it does not, leave them out. Is there anything wrong with starting a story from the end instead of starting from the beginning? Go for it and we shall see!

The experiment was certainly interesting, and some of the stories produced were a compelling example and an encouraging prompt to continue researching the methods. In some cases, although the assignment was to work within the scope and the tools of the project (i.e., the custom CMS built for modularity and the list of compiled information needs), the stories were written in Microsoft Word or Google Docs beforehand and then rejigged into the modules grid at a later time. The 'rejigging' was done by the authors themselves, who had to shorten, revise, and rewrite. In some cases, the preexisting content was just shoved into the new matrix, and inevitably, the significance of a clear correspondence between modules and information needs was in part lost. This is sometimes the feeling I get reading some reports, where modularity is used as a UX template that does not match the news discourse's modularity.

The rabbit hole

What we did not attempt to do so far, though, is to grab an existing long-form and convert it to a modular approach, manually or using AI. I have never fully made peace with the 'discarded pile'. So I went back to the first few weeks of the collaborative research and looked at the stories we had examined and discarded and tried to infer which functions, if any, the modules they contained were corresponding to. In most cases, where I spotted function, I could not match it to any of the information needs in our working list.

So I widened the field. I started to read random stories from a wide range of sources in search of patterns. I had a hunch that information-needs-based modularity was particularly lacking in news reports with a declared political agenda. I therefore found myself dissecting for modules the coverage of the events of January 6 by certain American media as well as articles promoting climate change denialism or other controversial topics where fact-based journalism takes the back seat and the news discourse is driven by an entirely opposite "semantic macrostructure of the text."

The dysfunction manifested in different ways: in the form of fragmented and incomplete answers to information needs, or as fluff that seemed to have no function whatsoever, or as a jumbled mix of different things, or, in some cases, as an apparent intent to do the exact opposite of matching information needs — a deliberate intent to not answer basic questions. One could say: deliberately confuse or misinform.

I had to assume that the authors of those stories intended to write them in the way they did. So I tried to define a different entity in the modular API, an anti-information-need of sorts. For lack of a better definition, I named this new entity "user effect," as it resonates with the effect in front-end development frameworks — a trigger in the code that produces a shift in behavior on the user's side. User effects could be described as 'bad habits' of modularity: portions of the text that not only don't serve a purpose for the users, but appear to be crafted with the intent to disinform and manipulate or do so inadvertently for lack of journalistic standards and malpractice.

My suspicion — it's only a suspicion — is that a lack of information needs, regardless of modularity, in a news artifact, and even without a significant presence of user effects, is a sign of adherence to unsatisfactory and unhelpful traditional modes of journalism.

User effects should not typically be found in ethical journalism, but some of them are quite frequently used by mainstream sources in many countries, particularly when it comes to sourcing transparency, mixing opinion with facts, neglecting context and verification, and relying on false narratives.

Adding the user effect perspective to the picture, we have a new tool to flag potentially unethical content by its role within the news discourse and without expressing a qualitative judgement. A perfect example would be this piece titled "Autism, the genetic code deciphered. Breakthrough in early diagnosis and treatment", published with great relevance in the newspaper La Repubblica on September 2, 2024. The article lacks crucial elements we would expect to find in ethical and responsible journalism, particularly when dealing with a sensitive topic like autism and potential scientific breakthroughs related to it. Let's break it down:

• There is no reference to peer review or any reference to discussion on similar studies that might highlight problematic aspects of machine learning applied to autism research.
• The article does not include comments from other scientists or experts in the autism field to provide context and a critical assessment of the research.
• The article uses very assertive and definitive language, without any hint of caution or uncertainty. In the lede, it uses expressions like "breakthrough" or "deciphered the genetic code."

We could approach the article from a journalism standards standpoint and say that this is a standard case of misinformation, even if the research turns out to be valid in the future. We could say that the lack of rigor and critical context can contribute to the spread of dangerous narratives, from false hopes to misunderstandings, to prejudices. We could say that inaccurate science journalism and sensationalism can undermine public trust in science and the research process. Or we could, more aseptically, flag the lede of the article with the user effect: "I will report findings from scientific papers or research as established facts, without providing caveats or acknowledging the broader context and the peer-review process," and prompt a rewrite based on the information needs.

This is a quick translation of the article's lede on the day it was published (hopefully someone will notice and edit it soon):

Soon, a new tool could put an end to the long waits, months or even years, often needed before an autism diagnosis can be reached. A multi-university group of researchers, co-directed by Gustavo K. Rohde, professor of Engineering at the University of Virginia, has deciphered the genetic code of autism, leading to a breakthrough in diagnosis and, consequently, early treatment. The researchers have developed a new approach, based on a new artificial intelligence, capable of identifying the genetic markers of autism through biological activity in the brain with an accuracy of 89-95%.

There might be an argument for "user effects" aligning with the needs of certain users. Do flat-earther users need fact-based journalism, or do they seek validation of conspiracy theories? They, and similar user personas, consume news not looking for a balanced account of events, but instead for a biased account reiterating specific narratives. This exception doesn't seem to contradict the general assumption; it's just a different perspective on the same logic of facts vs. manipulation.

Unfortunately, most news consumers in many countries experience mildly or grossly substandard journalism on a daily basis. I will refer to a brilliant note I read a few months ago (November 5, 2023) in the weekly column of Luca Sofri, editor-in-chief of Il Post, one of the few Italian outlets that attempts to follow journalistic standards (I apologize for the quick translation).

Those who read newspapers, skim through them, flip through the pages, listen, watch, and follow the work of the newsrooms only in their results, owe their limited knowledge of how they operate mainly to certain American movies or series, which are the main occasion when even the Italian public sees journalists at work [...] And the risk is that they may take as standard practices of the newsrooms what they see in those films: the caution in choosing what and when to publish, the search for verification, the request to find multiple sources, the decision to wait and postpone publication until there are answers and confirmations, the assurance that the quoted phrases can be attributed. But these are practices characteristic of the most authoritative American publications, and rare cases in some other countries of the world. Providing to readers information on the reliability of the news and the sources that report it, for example, is a choice that is not in the Italian journalistic culture, which instead has among its attributes a greater propensity (and ability, sometimes) for suggestive and engaging storytelling. The tools of this different approach are therefore a whole series of formulations about sources that, with a little attention, every reader can learn to recognize: "the rumors say...," "it is whispered that...," "those in the know assure...," "apparently...," "those who have seen it say that...," "they say in his inner circle...."

I have immense respect for Sofri and his newsroom and I am very thankful for the role Il Post has played to promote fact-based journalism in Italy in the last 15 years. However, I disagree that what he's describing is just a "different approach" or that it can have any justification in ethical journalism. Those "formulations" seem to me perfectly suited to be coded as user effects to be flagged and hopefully corrected before they reach users.

The way forward: testing the hypothesis

The next step in this exercise in modular journalism is to test the theory further, examining existing news artifacts and attempting to analyze them, assigning user effects if present, and rewriting them based on information needs and the principles of information-needs-based modular journalism.

The first example is an article about a scientist who denies climate change, published by an Italian newspaper in January 2023.

It's presented in alternate versions, the first one is a translation of the original, with an attempt to match portions of the text to user effects. The remaining versions apply the principles of user-based modularity as defined in the modular journalism 1.0 project.

The most obvious effect of the rewrite is that we keep a total of two sentences from the original, while the rest is discarded, even though the report is an interview.

Out of the box, if Google's Gemini is asked to evaluate whether the first two paragraphs of the article adhere to journalistic standards, it responds that they do not, in the following ways:

Lack of Neutrality: The statement "The cause is not humans, but rather the Sun" is presented as a definitive fact, whereas the scientific consensus attributes global warming primarily to human activity. A more neutral approach would be to present Zharkova's view as one perspective among others or to acknowledge the prevailing scientific consensus.

Lack of Context: The paragraph doesn't provide any context about Zharkova's expertise or whether her views are widely accepted within the scientific community. Providing such context would help readers assess the credibility of her claims.

Sensational Language: The phrase "wants to disprove some common beliefs" suggests a confrontational approach and could be seen as framing Zharkova's views in a way that amplifies their controversial nature.

If ChatGPT 4 is asked the same question, it also agrees that the article does not follow journalistic standards, for the following reasons:

Lack of Context and Balance: The paragraph presents Valentina Zharkova’s views on global warming without providing any immediate context or counterpoints from the scientific consensus on climate change, which overwhelmingly supports human activity as a significant driver of recent global warming. Journalism standards typically require presenting multiple perspectives, especially on scientifically contentious issues, to provide a balanced view.

Attribution and Authority: While the article mentions Zharkova’s credentials and quotes her directly, it does not specify her expertise in climate science. Zharkova is a solar physicist, not a climate scientist, which could be relevant for readers to understand her perspective's scope and limitations. Journalism standards often require clarifying a source's expertise, especially when discussing complex topics like climate change.

Misleading Implications: The paragraph implies that Zharkova’s views are equally credible as the mainstream scientific consensus without explicitly stating that her views are a minority within the scientific community. This can mislead readers into thinking there is more scientific debate about the role of solar activity in current global warming than there actually is.

Lack of Verification: Journalism standards call for verifying claims. The article repeats Zharkova's statements without checking them against established scientific research. Good journalistic practice would include verifying facts and providing context about the prevailing scientific understanding.

I have identified three user effects in the first paragraph of the article:

• I will knowingly omit, distort, or misrepresent details in my reporting to support a narrative.
• I will frame statements or events, giving an interpretation that supports a narrative.
• I will cast opposing arguments as common beliefs, undermining their legitimacy and implying they are unoriginal, uncritical, or even mindless.

If we ask Gemini to rewrite the lead to answer the question "What happened?" and making sure that the three intents are not present, this is what we get:

Valentina Zharkova, a Ukrainian scientist at Northumbria University, offered an alternative perspective on global warming in an interview with La Verità. She suggested the Sun, not human activity, is the primary driver of climate change. Zharkova challenges the widely accepted view that the Earth's temperature has been rising since the Industrial Revolution, proposing instead that the planet has been warming since the end of the Little Ice Age around 1690.

If we ask ChatGPT 4 to do the same, we get the following lead:

Valentina Zharkova, a Ukrainian scientist and professor at Northumbria University in Newcastle, UK, shared her views on global warming in an interview with La Verità. Zharkova, whose research focuses on solar activity, argued that recent climate changes are primarily driven by variations in solar radiation rather than human activities. She challenged the widely accepted view that global temperatures have been rising predominantly due to human-induced greenhouse gas emissions since the Industrial Revolution. Instead, Zharkova cited historical climate patterns, such as the warming trend following the Little Ice Age, as evidence that solar cycles play a significant role in Earth's climate variability. Her position is a minority viewpoint within the broader scientific community, which largely attributes recent global warming to human activities.

While we're here, we could also ask the language models to rewrite the lede of the article about machine learning and autism mentioned above, answering the information need "What happened?" and removing the detected user effect.

This is Gemini's take:

Researchers led by Professor Gustavo K. Rohde at the University of Virginia are exploring a new AI-based approach to identify potential genetic markers of autism through brain activity. Early results suggest a promising accuracy rate, but further research and peer review are needed to validate these findings. If successful, this approach could contribute to earlier diagnosis and intervention for autism.

This is ChatGPT 4's take:

Researchers from multiple universities, co-directed by Gustavo K. Rohde at the University of Virginia, have developed a new AI-based method that may improve the speed and accuracy of autism diagnosis. This approach identifies potential genetic markers associated with autism by analyzing brain activity, achieving an accuracy rate between 89% and 95%. While promising, these findings are still under study and require further validation.

Let's get prompting!

Even though language models add a lot of their own to the piece and this is an unacceptable risk for hallucinations, the results are obviously encouraging. In the new modules API, I will attempt to define prompting for both information needs and user effects.

Further tests will determine whether we can:

• efficiently flag the presence of user effects in a text using AI.
• use the compiled list of user effects as guardrails to generate answers to information needs questions using generative AI.
• use AI to detect information needs and build stories using modules.

A special thanks to Shirish Kulkarni for his invaluable insights and suggestions.