Sources is a GitHub repository (a piece of code) that takes a news event and returns sources, categorized and ranked, representing a range of diverse viewpoints.

Giorgio Catalani (find him on LinkedIn) started his career as a journalist and slowly moved closer to technology. He works as a sr product manager for a major publishing group; everyday, he tries to make life easier for editors and journalists. Big internet nerd, movie buff and wannabe cook. This is his first article for Unbubble Hub.

Suppose the European Central Bank (ECB) announced this week that it was holding interest rates steady; a decision met, as these always are, with disagreement about whether it was too cautious, too late, or exactly right.

Suddenly, the news is everywhere. A reader, wanting to understand what is going on and seeking the broadest possible perspective, turns to a news aggregator, a search engine, or an AI assistant. Ten results come back. They arrive from ten different outlets, across four political leanings, in three languages.

If those ten results are wire reports - e.g Reuters, Ansa, Ap - they will be strikingly homogeneous in stance: the rate, the vote, the statement, the market reaction, and a quote from Christine Lagarde. As a set, they may do a reasonable job covering the key verifiable facts, with only subtle differences in what gets foregrounded. If, instead, the ten results are op-eds, they will look gloriously diverse in stance: “the ECB is choking the recovery,” “the ECB is fighting the last war,” “the ECB is right to wait.” Yet, they will be redundant on fact, because op-ed writers all argue from the exact same narrow set of available data.

Foto di Evie S. su Unsplash

The News Ecosystem

A breaking news item and an opinion column both shape public perception, but through fundamentally different mechanisms. A breaking news item shapes perception by selecting which facts to present and which to omit; an opinion column shapes it by arguing what those facts mean, who is favored by the decision, and what the repercussions might be.

The worst disservice we can do to ourselves as readers is to compare these outputs as equals. It is fairly easy to label them as “subjective vs. objective” or “biased vs. unbiased,” but that misses the point: they are completely different products engineered for entirely different needs.

Journalism studies has known this for decades. Work on Western newspapers tracing back to the 1960s highlights this distinction across national systems, proving that the news item, the interpretive report, and the commentary are distinct genres governed by different norms and yielding different effects.1

What is strange is that almost none of this nuance has crossed over into the systems currently deciding what gets read. News aggregators, search engines, LLMs, and “compare coverage” features treat an article as an article, and a link as a link. They adjust for source reputation, recency, and language, but the actual genre of the article remains invisible to them. A reader looking for the broadest possible perspective on a fact is gambling, hoping an AI will solve a problem it was fundamentally not built to solve.

This is the exact problem we are trying to solve with Sources.

What Publishers Already Know

Publishers already encode these distinctions into their infrastructure. They assign different genres to different desks and templates. They mark bylines differently: a political editor writes analysis, a columnist writes opinion, a fact-checking unit handles verifications. Most importantly, this architecture is visible in the metadata, often right in the URL: /news/, /analysis/, /opinion/, /fact-check/, /guide/.

Publishers do not do this to help algorithms; they do it because they are speaking to two different audiences simultaneously:

• The Reader: Scans titles and bylines to gauge intent.

• The Platform: (Google, Apple News, AI search) Reads URLs and structured metadata.

A URL containing /opinion/ is a direct message sent to the algorithm: This is an argument, not a report. Weight it accordingly. A URL with /fact-check/ signals: This is a verification, not a claim.

These signals are practically free. They are already produced by the publisher and baked into the page source. Yet, the machines deciding what to show you mostly ignore them.

There is, however, a real complication. Research shows that journalistic cultures vary wildly. Italian newspapers, for example, blend opinion into news at significantly higher rates than their British or German counterparts, while Spanish and French outlets sit somewhere in the middle.2 This is not a bug to design around; it is a feature of different journalistic norms regarding where the line between reporting and interpretation sits. Any system attempting to classify articles by URL and title will naturally work better in some languages than others. This requires careful calibration, not assumption. But it is still a better starting point than pretending the signal doesn’t exist.

A Seven-Type Framework to Define the Ecosystem

To operationalize this, I am proposing a first-pass taxonomy: seven article types, separated by their core intent, and identifiable (albeit imperfectly) by URL and metadata signals.3

To make this concrete, let’s watch our hypothetical week of ECB coverage refract through each type:

1. The Breaking News Item: A concise, fact-focused report produced under time pressure. The lead answers who, what, where, and when. Sources are cited as authorities, not interlocutors. No thesis is advanced.

• Example: “ECB holds rates at 3.5%, cites persistent core inflation.”

• Signals: /news/, /world/, /economy/; wire origin; staff/newsroom byline.

2. The News Analysis: Goes beyond current facts to speculate on significance, outcomes, and motives. It remains journalism, making empirical claims rather than normative judgments. It answers why at length.

• Example: “Why Lagarde signalled caution despite cooling headline inflation.”

• Signals: /analysis/, /news-analysis/; senior correspondent or bureau chief byline.

3. The Opinion or Editorial: Explicitly exercises normative or evaluative judgment. The author takes a position and argues for it, often using first-person and evaluative language (”should”, “must”, “dangerous”). The purpose is to persuade.

• Example: “The ECB is choking the recovery—and Italy will pay first.”

• Signals: /opinion/, /commentary/, /editorial/, /op-ed/.

4. The Explainer: Pedagogical content that provides historical, institutional, or structural context. It answers “what is…” or “how does X work.” Often evergreen.

• Example: “How does the ECB actually decide where to set rates?”

• Signals: /explainer/, /guide/, /background/; didactic formatting (FAQs, lists).

5. The Interview or Testimony: Dedicates extended space to a single voice (e.g., Q&A, profile, long monologue). The subject is the focal point of the piece.

• Example: A Q&A with a former central banker on the rate decision.

• Signals: /interviews/, /voices/, /profiles/, /long-reads/.

6. The Fact-Check: Evaluates the truth of a specific public claim. Structured logically: claim → evidence → verdict.

• Example: “Did inflation really drop as much as the government is claiming?”

• Signals: /fact-check/, /verifica/; dedicated fact-checking outlets (Full Fact, Reuters Fact Check).

7. The Wire Republication: Reproduces a wire service report with minimal to no editing. Identical across multiple outlets.

• Example: The exact same Reuters story run verbatim across six Italian dailies.

• Signals: Explicit wire credit; identical headlines across properties. (Note: This category is the most fluid; it may function better as a “provenance tag” applied to Type 1 rather than a standalone genre).

What is Next for “Sources”?

Our goal is to operationalize the concept of “perspective”: to define it with enough precision that a system can act on it. If our mission is to provide the broadest possible perspective on a given fact, this framework might play a substantial role.

That being said, there are some open questions that we still want to address:

• Does the framework classify accurately enough to use? We will run the taxonomy against current Sources output for a representative news event and compare the distribution against hand-labeled ground truth. The ultimate stress test will be the Italian media ecosystem, where the blend of reportage and interpretation will require heavy calibration.

• Can the framework define a useful subset of sources? Our current ranker selects the top ten sources based on relevance and five existing diversity dimensions. If we add genre type as a pre-filter, does it produce a subset that is materially better? Specifically: is it better balanced between fact-supply and argument-supply?

• Does this actually enhance perspective? Adding type might make the system’s output measurably more diverse, or it might just add an invisible layer of structure that readers ignore.

Ultimately, we believe that you cannot fix the blindness of algorithmic retrieval just by feeding the machine more links. You have to teach the machine what kind of text it is looking at. We don’t have all the answers yet, and this first version is a starting point, not a solution. We’ll publish what we learn here as it comes, including the parts that break.

Sources is a GitHub repository (a piece of code) that takes a news event and returns sources, categorized and ranked, representing a range of diverse viewpoints.

David La Barbera (find him on LinkedIn) is a researcher at the University of Udine, with a PhD in Artificial Intelligence and a focus on misinformation and information retrieval. With Unbubble Hub he works on what he sees as one of the central problems of our information environment: the slow, structural way polarization and disinformation shape what people get to read, often without anyone noticing. Sources is part of that effort. This is his first article for Unbubble Hub.

Consider this scenario. It’s spring 2026, you live in Europe, and you’re thinking about buying an electric car. You read something about the Chinese ones. They’re cheap and the reviews are decent. But you heard something about EU tariffs, and before spending twenty or thirty thousand euros, you want to understand what’s going on: why is the EU trying to make these cars more expensive?

So you do what anyone would do: you ask a search engine. You read the first article. Then maybe the second. They’re pieces from reputable outlets explaining the EU’s reasoning, the Chinese government’s response, and some analysts’ views. In fifteen minutes you have an opinion: tariffs protect European jobs from subsidized Chinese competition. China objects, talks are ongoing. You close the tab.

This is what most searches look like: users typically read one or two of the top results, end without any click at all: the user finds an answer in the snippet at the top of the page, or simply moves on.

But here’s the catch: there are at least seven or eight serious positions on this issue, held by people who have thought about it carefully and arrived at different conclusions. And you probably just formed your opinion based on one or two of them.

What you didn’t read

Let me name three perspectives on the matter, just to make this concrete.

The first one you almost certainly encountered: European carmakers can’t fairly compete with Chinese manufacturers backed by massive state subsidies. Without tariffs, you lose factories, jobs, industrial know-how that took decades to build.

There’s a second perspective: the energy transition is urgent, Chinese EVs are good and cheap, tariffs make electric cars more expensive which slows electrification and means more emissions for longer. The right frame, from this angle, isn’t “Europe vs. China” but “how fast can we electrify the European fleet.” This perspective tends to be nearly invisible. Not censored, just not foregrounded by the editorial logic of business and political coverage.

The same can be told about a third perspective, at least for queries in any Italian, French, or Spanish language, even though it’s been central to the German debate. The fact is that the European auto industry is not a monolith and carmakers like BMW, Mercedes, and Volkswagen have decades of investment in China, to the point that they now sell more cars in China than in Europe. For them, EU tariffs are a threat: China retaliates and their exports get hit, and some of the targeted vehicles are partly their own. This is why a significant part of the European industry actively lobbied against the tariffs.

Did you know these three perspectives? I did not, before writing this piece. And there are others I’m not naming here, cause this isn’t a piece about EV tariffs: it’s a piece about what happens when you try to inform yourself about questions like this one.

Why this kind of question is different

But what kind of question is this one, exactly?

For some questions, the facts alone settle the answer, and the job of a search engine is just to surface those facts accurately. Ten results saying the same thing means the system did its job.

Take “when did the French Revolution begin?” There is a date, July 14, 1789. Expecting “multiple perspectives” on that would be nonsensical. Or take “is human activity causing the Earth to warm?” Here too, there is a scientific consensus built over decades of evidence. This doesn’t mean objections are pointless (scientific claims, unlike calendar dates, are sharpened by scrutiny) but scrutiny is not the same as balance. Treating climate denial as “the other side” is not diversity. It’s misinformation dressed up as fairness.

But the EV tariff question is in a different category entirely. There are shared facts here too: Chinese EVs cost less, the EU has imposed tariffs, and so on. Nobody contests these. What’s contested is what to make of them. Different people, looking at the same facts, might reach different conclusions because they weigh different considerations: jobs now versus climate acceleration, industrial sovereignty versus consumer welfare, short-term protection versus long-term competitiveness. The disagreement isn’t a gap in the facts. It’s that the same facts admit more than one legitimate reading.

We can call these substantively contested questions. Their defining feature is that reasonable people, looking at all the available evidence, can and do legitimately arrive at different positions. The disagreement isn’t a bug to be fixed by collecting more data. It’s a structural feature of the question.

This is worth pausing on, because there’s a subtle point here that’s easy to miss. Saying that more data won’t resolve the disagreement is not the same as saying that nothing is missing from your view. Something usually is missing, but what’s missing isn’t facts. It’s exposure to the range of substantive positions through which those facts get interpreted. You can have all the data on Chinese subsidies, European jobs, and emission targets, and still hold a partial view if you’ve only encountered one or two of the positions that read those facts. The gap isn’t in the evidence, it’s in the conversation.

This matters because questions like this are at the core of democratic life. Immigration policy. Energy strategy. Taxation. Urban planning. Foreign policy. The governance of technology. These aren’t questions we’re going to solve by collecting more data. They’re questions we answer, and re-answer, by working through the competing positions that real people hold for real reasons.

Why retrieval systems work this way

It’s worth a brief word on why this happens, because the mechanism is not malicious and not fixable in the obvious ways. Any retrieval system such as search engine, recommendation feed, AI assistant, has to make three kinds of choices. First, it chooses what to index in the first place (some content is harder to collect, some is behind paywalls, some is in languages the index handles poorly). Second, it chooses how to rank what it has indexed (and ranking algorithms tend to reward already-popular sources). And third, it has to choose how to interpret your query. None of these choices is avoidable. Showing you everything is not an option; everything is of course too much. So someone has to decide what to put in front of you, and someone has, algorithmically. But every choice produces a boundary, and the boundary is not a flaw of the system, it is a structural property of any system that selects a finite subset from an effectively infinite informational reality.

This is worth distinguishing from the filter bubble argument that dominated the 2010s: the idea that algorithmic personalization shows each user a world tailored to their past preferences. That problem exists. But the problem I’m describing sits upstream of personalization. Even for a user the system knows nothing about, with no browsing history, the result set is structurally truncated. Personalization can shift the boundary; it doesn’t create it. The consequence is subtle. If I form an opinion based on systematically partial input that opinion isn’t necessarily wrong. The things I read may all be true. The facts I cite may all be accurate. But my opinion is incomplete in a way I cannot assess from within it. I don’t have the tools to know what I’m missing, and I can’t even suspect it, because the missing dimension doesn’t present itself to me as “missing”, it simply isn’t in my space of considered alternatives.

This is the specific harm: not error, but blindness through non-exposure. I don’t believe the false. I just don’t know part of the true.

Sources: what we’re building

At Unbubble Hub we’re working on tools that try to address the problem I just described: to make this boundary, at least for the questions where it matters, partially visible. The first of these tools is called Sources, and a first version is already online.

The goal is straightforward to state: given a query on a substantively contested topic, Sources tries to maximize the diversity of substantive perspectives covered, with the smallest set of sources possible. The underlying assumption is that on this kind of question, diversity itself is a proxy for quality. A small set of sources that spans the range of substantive positions is more useful, for someone trying to form an opinion, than a large set that repeats the same one or two positions in different voices. Sources expands the retrieval to deliberately surface positions that default ranking does not foreground, annotates what was found, and reports transparently which axes of diversity it tried to cover.

We don’t know yet how well this works. The first version is a starting point, not a solution. We’ll publish what we learn here as it comes, including, and especially, the parts that don’t work. If you have critiques, corrections, or want to collaborate, we’re listening and very open to feedback!

What we still don’t know how to do

The hard part of this work though, isn’t technical: it’s conceptual.

One might ask: who decides which perspectives count as substantive? Any answer is a normative choice, and different perspectives might not carry equal weight in public discourse. Any system that treats all perspectives as equivalent is making a claim it should be forced to defend. We haven’t solved this. We’re not sure it can be fully solved by an algorithm. What we believe is that the criteria a system uses have to be declared openly and made contestable, not buried inside the ranking.

One might also ask: how do we even identify the actors in a given debate? In some cases the actors are obvious while in others they emerge over time. A retrieval system that decides this question silently is making editorial choices without naming them.

Also: how do we formalize “perspective” in a way that is usable computationally without flattening what makes perspectives different in the first place? Reducing a position to a set of tags risks losing the texture that makes it a position rather than a slogan. But not formalizing it at all means the system can’t operate at scale.

Our working principle, across all of these, is that the meta-level has to be made explicit. A system trying to offer position diversity should declare openly which axes of diversity it is trying to cover, by what criteria, and with what known limits. Otherwise it reproduces, at a higher level, the same invisibility the whole exercise was trying to address.

If we don’t manage to make this boundary at least partially visible, for the questions where it actually matters, we’ll keep building solid opinions on foundations we can’t inspect. Not because those opinions are wrong, but because they’re incomplete in ways we can’t see, on questions where incompleteness isn’t a technical detail. It’s the substance of the problem.

That’s what Sources is trying to change, one (contested) question at a time.

References

Fishkin, R. (2024). 2024 Zero-Click Search Study. SparkToro, on Datos/Semrush panel data. sparktoro.com

Pan, B., Hembrooke, H., Joachims, T., Lorigo, L., Gay, G., & Granka, L. (2007). In Google We Trust: Users’ Decisions on Rank, Position, and Relevance. Journal of Computer-Mediated Communication, 12(3), 801–823.

Pandey, S., Roy, S., Olston, C., Cho, J., & Chakrabarti, S. (2005). Shuffling a Stacked Deck: The Case for Partially Randomized Ranking of Search Engine Results. Proceedings of VLDB.

Pariser, E. (2011). The Filter Bubble: What the Internet Is Hiding from You. Penguin Press.

Further reading

If you want to dig into the EV tariff debate yourself, here are some entry points spanning different positions. They are not exhaustive, and they intentionally include perspectives that disagree with each other.

Pro-tariff / industrial protection. European Commission (2024). Provisional countervailing duties on imports of battery electric vehicles from China. Available at ec.europa.eu. The Commission’s own framing of the case. https://trade.ec.europa.eu/access-to-markets/en/news/eu-commission-imposes-countervailing-duties-imports-battery-electric-vehicles-bevs-china

Skeptical of the tariffs (free-trade and consumer perspective). Dadush, U. (2024). The European Commission’s duties on Chinese electric vehicles are a mistake. Bruegel.

https://www.bruegel.org/first-glance/european-commissions-duties-chinese-electric-vehicles-are-mistake

Climate-pragmatic angle. Mazzocco, I. (2023). Balancing Act: Managing European Dependencies on China for Climate Technologies. CSIS Brief.

https://www.csis.org/analysis/balancing-act-managing-european-dependencies-china-climate-technologies

German industrial perspective. Barkin, N., Kratz, A., & Sebastian, G. (2024). Ain’t No Duty High Enough. Rhodium Group. On why German automakers are exposed to retaliation.

https://rhg.com/research/aint-no-duty-high-enough/

Empirical / contrarian read on the data. Conconi, P., et al. (2026). Don’t swap tariffs for minimum prices on Chinese electric vehicles. CEPR. Argues the “flood of Chinese imports” narrative does not match the trade data.

https://cepr.org/voxeu/columns/dont-swap-tariffs-minimum-prices-chinese-electric-vehicles

For broader context on diversity in news retrieval. Pan et al. (2007), Pariser (2011), and Pandey et al. (2005), cited in the references above.

Methodological note

I used Claude to refine sentences, test phrasings, and pressure-test the structure of the argument.

Unbubble Hub is an Open Research Initiative started in early 2026.

We aim to provide a space for researchers and engineers to come together and collaborate in developing tools to fight social polarization.

We want to help people form their own judgment not by promoting “correct” opinions, but by fostering meaningful, conscious disagreement, encouraging a plurality of perspectives.

In order to do so, we believe that information integrity tools should be open, transparent, and accessible to everyone.

You can find our work on GitHub. Right now we have two working projects:

• /sources, a module that takes a news event as input and returns sources, categorized and ranked, representing a range of diverse viewpoints as output.

• /gdelt-pulse, a pipeline that pulls GDELT's 15-minute data updates, extracts sources and annotations, and builds a searchable database of global news coverage. Designed for media bias and perspective research.

Moreover, we’re working with Ca’ Foscari University to offer small course projects for students, and we’re writing an academic paper that we hope to publish in the next months.

Today we’re starting this Substack to be truly faithful to our mission of building in public! In the next days we’ll start writing about our projects and roadmap more in details.

Subscribe now

Our team currently counts ~20 people between mathematicians, students, professors, engineers, researchers and enthusiasts, from different countries. We’re open to everyone: if you’re interested in joining us or starting a conversation, simply reach out, or (if that’s your' thing) you could even directly open a PR!

Motivation

We are witnessing a collision between two powerful forces shaping public discourse: social media platforms and governments.

Social media platforms have become central infrastructure for public discourse. Their recommendation algorithms, designed to maximize engagement, have demonstrably contributed to polarization and the spread of disinformation. The underlying mechanism is well-documented: algorithms that prioritize engagement metrics (likes, shares, comments) systematically favor content that triggers strong emotional reactions, including outrage, fear, and hostility. This is not a bug but a feature of advertising-driven business models.

The platforms themselves are opaque. Users cannot see why certain content appears in their feeds. Researchers struggle to study algorithmic effects without platform cooperation. Civil society cannot audit systems that shape billions of people’s understanding of the world.

Now, governments are responding. Australia became the first country to enforce a nationwide ban on under-16s accessing social media (December 2025). Spain has announced measures to criminalize “manipulating algorithms to amplify illegal content” and systems to track “how digital platforms fuel division and amplify hate.” France, Denmark, the UK, and others are considering comparable laws. The EU’s Digital Services Act already mandates algorithmic transparency for large platforms.

These regulations respond to real harms, but the regulatory response raises its own questions. Will the governments introduce new black boxes, this time state-controlled rather than corporate-controlled? Who will create the systems to track “how platforms fuel division”?

Generative AI is now layered on top of this infrastructure, and it cuts both ways. On one side, it lowers the cost of producing persuasive content to near zero. But on the other side, the same models make it much cheaper to analyze news coverage, detect narrative patterns, and compare how different outlets frame the same event. Again: whether GenAI ends up deepening the bubble or helping us out of it depends, in large part, on who gets to build the tools, and whether those tools are inspectable.

We believe that having open, transparent technology is a big part of the solution. The technology choices made by the governments in the next few years will shape information infrastructure for decades. If these systems are built as proprietary black boxes, we will have traded one form of opacity for another. If they are built on open foundations, civil society will have tools to audit, critique, and improve them, regardless of who deploys them.

That’s why we’re building Unbubble Hub in public, and that’s why we’d love you to join us!