Based on the presentation by Emiliano Guevara regarding Amedia’s implementation of the User Needs 2.0 model, I have extracted three core arguments. Below is a detailed elaboration of each argument, expanded to meet the depth and length requirements.
Argument 1: The Necessity of Contextual Adaptation and the “Idea” Over the “Model”
The first core argument presented by Guevara is that the “User Needs” framework is not a rigid, one-size-fits-all mathematical formula, but rather a conceptual starting point that must be rigorously adapted to specific cultural, linguistic, and organizational contexts. Amedia’s experience demonstrates that successfully deploying an AI-driven editorial strategy requires stripping a model down to its philosophical core—that audiences consume news for reasons beyond simple information transfer—and rebuilding it to fit the specific reality of the publisher. For Amedia, this reality is defined by the Norwegian language, a decentralized structure of over 100 local newspapers, and a strict separation between the technological publisher and the independent editors.
To understand this argument, one must first look at how Guevara redefines the User Needs 2.0 model (originally developed by Dmitry Shishkin and SmartOcto). He explicitly states, “I don’t see the model, I see an idea.” This distinction is crucial. If treated as a rigid model, the framework becomes a compliance exercise where journalists are forced to tick boxes that may not fit their reporting. If treated as an idea, it becomes a flexible lens for understanding audience psychology. The core idea is that the “Update Me” function—the simple transfer of facts—is only one of many reasons a human being interacts with media. They also seek to be educated, inspired, diverted, or connected. Amedia’s first hurdle was acknowledging that while the BBC or SmartOcto might have defined these categories in English for a global or broad national audience, those definitions do not automatically map onto the hyper-local ecosystem of Norwegian journalism.
The adaptation process at Amedia was not merely a translation of keywords from English to Norwegian; it was a semantic and cultural localization. A prime example provided in the presentation is the transformation of the “Connect Me” category. In the standard User Needs 2.0 model, “Connect Me” often implies connecting the reader to broader trends, ideas, or complex social issues—linking the individual to the “outside” world. However, Amedia’s portfolio consists primarily of local newspapers serving small communities (e.g., a village with a Coca-Cola Christmas truck). In this context, “Connect Me” felt abstract to the journalists. The analysis team, therefore, redefined it. It became “Make me belong.” This is a profound shift in perspective. For a local newspaper, the user need is not about connecting to the abstract world, but about reinforcing the reader’s identity as a member of their specific village or town. The “Give me the fact that you belong to something” definition turns the category into a tool for social cohesion, which is the lifeblood of local media. Without this adaptation, the AI model would have been trained on a concept alien to the content actually being produced, leading to poor predictions and low adoption by newsrooms.
Furthermore, this adaptation process involved significant friction and debate, which Guevara highlights as a feature, not a bug, of the implementation. He notes a “heated discussion” regarding the number of categories. The SmartOcto model proposes eight micro-categories (two for each of the four macro-categories: Know, Do, Understand, Feel). Amedia’s content analysts initially argued that they only saw evidence of seven categories in their production. This creates a dilemma: does one force the data to fit the theoretical model (eight categories), or does one alter the model to fit the observed reality (seven categories)? Amedia ultimately decided to stick with the eight-category structure for the sake of standardization and potential future utility, but the existence of the debate highlights the argument that these models are not divine writ. They are heuristics. The eventual compromise involved writing extensive guidelines to interpret these categories specifically for Norwegian local news. This suggests that the “code” of the AI is less important than the “code” of the annotation guidelines—the human consensus on what a story actually is.
The argument for adaptation extends to the organizational structure of Amedia itself. Guevara emphasizes that Amedia is a publisher based on an “economy of scale,” owning shares in over 110 independent publications. Crucially, “publishers cannot instruct or overrule the editor on editorial issues.” This legal and ethical divide means the AI team cannot simply mandate the use of User Needs 2.0. They cannot walk into a newsroom and say, “You are writing too many Update stories; stop it.” They can only provide tools, insights, and inspiration. This constraint necessitates a model that is persuasive rather than coercive. If the model had not been adapted to the specific language and “flavor” of Norwegian local news (e.g., categorizing a story about a local healthcare worker as “Inspire Me” rather than just “Update Me”), the editors would likely have rejected the insights as irrelevant. The AI had to demonstrate that it “understood” local journalism.
Finally, this argument touches on the temporal nature of these models. Guevara alludes to the fact that models must evolve because reality changes. He shares an anecdote about a previous classification model during the COVID-19 pandemic. The system had learned that “Corona” was a beer brand associated with summer and leisure. When the virus hit, the model began classifying global health crisis articles as “Entertainment/Free Time.” This failure underscores that a model is a snapshot of a specific worldview at a specific time. Adaptation is not a one-time setup task; it is a continuous process. The User Needs model relies on how humans interpret stories now. As society changes, what constitutes “Inspiration” or “perspective” might shift. Therefore, the core argument is that implementing User Needs 2.0 is not a software installation; it is a continuous cultural and linguistic translation project. The technology (BERT, PyTorch) is secondary to the intellectual labor of defining what these categories mean for this specific group of people, in this specific language, at this specific moment in history. Without this granular adaptation, the model remains an abstract academic concept with no traction in the gritty reality of daily news production.
Argument 2: The “Cold Start” Challenge and the Trade-offs of AI Implementation
The second core argument derived from Guevara’s presentation is a technical and strategic one: implementing advanced NLP (Natural Language Processing) in a specialized domain requires navigating the “Cold Start” problem—the total absence of training data—and making calculated trade-offs between data volume, model complexity, and classification granularity. Guevara demonstrates that high-performing AI solutions for media do not require millions of dollars or massive proprietary datasets; they require smart resource allocation, leveraging open-source advancements (like Transformer models), and making difficult decisions about data simplification to achieve actionable results.
The primary obstacle Amedia faced was that “there is basically no data available.” While English language models have vast resources, and generic topic classifiers (Sports vs. Politics) are common, there was no pre-existing dataset for “Norwegian User Needs 2.0.” This places the team in a position where they cannot simply fine-tune a model immediately; they must first become data creators. Guevara details the “very boring work” of the annotation campaign, where human beings manually classified over 8,000 articles. This highlights a critical argument in modern AI: despite the hype around generative AI and automation, the foundation of specific business applications is still human labor. The quality of the AI is entirely dependent on the quality of the “Ground Truth” established by these human annotators. The team had to create a balanced dataset, ensuring that rare categories like “Understand” were not drowned out by the overwhelming volume of “Update” stories. This manual curation (taking two months with four people) effectively is the intellectual property. The code is standard; the annotated data is the gold.
Following the data creation, Guevara introduces the strategic choice of technology: the use of “NorBERT,” a Norwegian-specific BERT (Bidirectional Encoder Representations from Transformers) model, fine-tuned via PyTorch on Google Cloud’s Vertex AI. The argument here is about the relationship between data size and model architecture. In the “pre-transformer” era of machine learning (using CNNs or standard word embeddings), one needed massive amounts of data to train a model from scratch to understand language nuances. Amedia did not have the time or budget to annotate 100,000 articles. Therefore, they utilized the “Transfer Learning” paradigm. By using a Large Language Model (LLM) that already “knows” Norwegian grammar, syntax, and semantics (NorBERT), they only needed a smaller, high-quality dataset (the ~8,000 articles) to teach it the specific task of User Needs classification. This implies that for modern media organizations, the barrier to entry for AI is significantly lower than in the past. One does not need Big Tech resources; one needs a smart application of open-source tools. The cost of training—cited as roughly $20 per run—democratizes this capability. It argues that the limitation is no longer computing power or cost, but organizational will and literacy.
However, this implementation required a controversial simplification: the decision to use “Single-Label” classification. In reality, a complex piece of journalism often serves multiple needs simultaneously. A long-form article might start with an update, provide deep context (Educate), and end with a call to action (Help). A multi-label classification system would reflect this reality. Yet, Amedia chose to force the model to pick one dominant category per article. Guevara argues that this was necessary to reduce noise. Multi-label systems are notoriously difficult to train on small datasets because the “correct” answer is ambiguous even to humans. If humans cannot agree on the secondary and tertiary labels, the machine will learn nothing but confusion. By simplifying the target to a single label, they increased the model’s stability and performance (achieving F1 scores well above the random baseline).
This leads to a nuanced sub-argument about interpretability versus precision. While the training was single-label, Guevara reveals in the Q&A and final slides that the output can still be interpreted flexibly. The model outputs a probability distribution (logits) across all categories. Even if the system tags an article as “Do Connect” (67% probability), it might also assign a 28% probability to “Feel Inspire.” This allows the organization to have it both ways: a rigorous, simplified training regimen that produces a robust model, and a nuanced, probabilistic output that reflects the complexity of journalism. This approach—simplifying the input to master the training, then interpreting the output with nuance—is a masterclass in practical AI deployment.
Furthermore, the presentation highlights the limitations of the “No Update” category bias in machine learning. In their first iteration, the dataset was dominated by standard news updates. Neural networks are “lazy” optimizers; they quickly learn that predicting the majority class (Update) is the safest way to get a high accuracy score, even if the model learns nothing about the other categories. This required Amedia to actively balance the dataset, underscoring the argument that “raw” real-world data is often toxic to training. You cannot just feed a model your archives; you must engineer the data to teach the distinctions you care about.
Finally, the deployment architecture (Google Vertex AI endpoints) allows for scalability and near real-time processing. This moves the project from an academic experiment to a production utility. The argument here is that AI in media is only useful if it fits into the “flow.” By creating an endpoint that accepts text and returns a prediction in seconds, they pave the way for future integrations where journalists get feedback while writing. The technical journey described by Guevara—from lack of data, through manual annotation, to transformer-based fine-tuning and cloud deployment—serves as a blueprint. It argues that the “secret sauce” is not a proprietary algorithm, but the disciplined process of adapting general tools to specific, curated data.
Argument 3: The “Production vs. Engagement” Mismatch and the Future of Personalized Journalism
The third and perhaps most impactful argument concerns the business intelligence derived from the AI: there is a fundamental misalignment between the type of content newsrooms instinctively produce and the type of content audiences actually value and engage with. Through the lens of User Needs 2.0, Amedia identified a massive oversupply of “Know/Update” journalism and an undersupply of “Understand,” “Feel,” and “Inspire” content, despite the latter categories driving significantly higher engagement. This insight shifts the conversation from “how do we use AI” to “why are we doing journalism this way?”
Guevara presents qualitative and quantitative evidence to support this. Visualizations of front pages from various Amedia newspapers show a sea of “No Update” tags. This is the default mode of the journalist: something happened, so I will write what happened. It is the “transfer of information” model. However, when analyzing the performance of these articles (specifically in the sports domain, though applicable generally), a stark contrast emerges. The volume of “Update” articles is enormous, but the average reads per article are low. Conversely, the “Feel Inspire” category—stories that might profile a local athlete, tell a comeback story, or connect a sporting event to community identity—is low in production volume but high in average readership.
This argument exposes a legacy inefficiency in newsroom culture. Journalists define productivity by the number of stories filed, and the easiest story to file is a factual update. Writing an inspiring profile or an educational deep-dive takes more time and cognitive effort. However, the data argues that this “efficiency” is an illusion. Producing ten “Update” stories that few people read is less valuable than producing two “Inspire” stories that the community loves. The AI model acts as a mirror, reflecting this imbalance back to the newsroom in a way that intuition cannot. It quantifies the “gut feeling” that readers want more than just facts—they want connection and understanding.
The implications of this argument extend into the concept of “Actionability.” Guevara is careful to note that they are currently in the observation phase, but the roadmap is clear. The goal is to move from descriptive analytics (telling editors “you write 80% updates”) to prescriptive or nudge-based workflows. The disparity between volume and clicks suggests that newsrooms could actually work less (produce fewer articles) but achieve more (higher total engagement) by shifting the mix of user needs. For example, instead of five articles simply stating match results (commoditized information available everywhere), a local paper could write one match report and one deep-dive analysis (“Understand”) or player profile (“Inspire”).
However, Guevara complicates this argument by introducing the potential for AI-driven personalization, which he dubs his “letter to Santa Claus.” If the disconnect is that users have different needs, why force a single front page on everyone? He envisions a future where the User Needs tags are used to curate the user experience dynamically. If the system knows a specific reader ignores “Update” sports stories but clicks on every “Do Connect” nature story, the front page should morph to satisfy that specific need profile. This takes the argument beyond editorial correction (telling journalists to write differently) to algorithmic curation (changing how content is distributed).
Furthermore, the presentation suggests a future where Generative AI (GenAI) bridges the gap between supply and demand. Guevara proposes that if a journalist writes a “No Update” story, GenAI could be used to rewrite or augment that story into a “Feel Inspire” or “Understand Context” version. This is a radical proposition. It suggests that the “angle” of a story is no longer a fixed attribute of the journalist’s work, but a fluid variable that can be manipulated by AI to fit the reader’s preference. This would solve the production bottleneck; the journalist provides the facts, and the AI helps mold the narrative frame to maximize engagement across different user segments.
Finally, this argument highlights the limitations of current metrics. The team is looking at “clicks,” but Guevara admits that true success should be measured by “time spent” and deeper engagement. A “No Update” story is designed to be short; an “Educate” story is designed to be long. Comparing them purely on volume is tricky. Yet, the trend is strong enough to warrant attention. The overarching argument is that newsrooms are currently flying blind, driven by habit and tradition. The implementation of User Needs 2.0 is not just about tagging; it is about breaking the “commodity news” cycle. By identifying that the audience craves meaning (Understand) and emotion (Feel) more than they crave raw data (Update), Amedia is arguing for a fundamental restructuring of the local news value proposition—shifting from being a ticker of events to being a builder of community and understanding. The AI is merely the tool that makes this structural inefficiency visible and actionable.