ANALYSIS Who decides what is true on the internet: the promise of algorithmic filters and the limits of the "invisible editor"

The pre-launch of the eYou application, a European-designed social platform that promises to analyze every post through four models of artificial intelligence and classify the content as true, partially true, or false, is not just a product announcement. It is a symptom of a moment in which technical solutions are called upon to solve a structural political problem: who has the authority to decide what information is legitimate in the digital public space. Data shows that, in the last six months, the expression "fake news" has generated over 13,200 mentions in the international press, with a negative tone in 31% of cases and opinionated in 66% of appearances, distributed across the Web, Facebook, YouTube, and TikTok. The term circulates intensely and constantly, reaching peaks of over 200 mentions per day, in parallel with major political events in the United States, Brazil, Romania, or the geopolitical space generated by conflicts in Ukraine and the Middle East. In such a context, any system that assumes the role of an automatic arbiter of truth deserves rigorous analysis, not just presentation.

How does an automatic information verification system actually work?

At a technical level, an "algorithmic fact-checking" system involves several steps: identifying verifiable claims from a text, comparing them with databases and sources considered authoritative, calculating trust scores, and aggregating the results into a final label. The process unfolds automatically, in real-time, at a large scale, which fundamentally differentiates it from classic human verification. The journalist or human fact-checker chooses what to verify, explains the process, and presents it publicly for debate. The algorithm applies a uniform rule, without explaining and without having doubts. This difference is not minor. It transforms the platform from a simple distribution channel of messages into an invisible editorial actor. The verdict does not appear as a fact-checking article, with transparent methodology and sources, but as a standardized label placed directly alongside the content. The user sees the conclusion, not the reasoning. This compression of the editorial process is, at the same time, the efficiency of the system and its underlying vulnerability, because it removes from the equation precisely the element that makes verification legitimate: traceability and public accountability.

The pressure for automation: where it comes from and what fuels it

In this context, the advantages of automating verification are evident at an operational level. Speed is the first solid argument: in an information crisis, contested elections, armed conflict, natural disaster, the window in which false information can dominate the public narrative is measured in minutes. A system that instantly marks suspicious content can limit initial distribution before a journalist has time to write a correction. Scaling is the second argument: at the volume of content generated daily on digital platforms, no newsroom and no consortium of fact-checkers can cover more than a tiny fraction of the information flow. Automation is the only way to extend verification at an industrial scale without proportionally multiplying costs. The third argument is formal consistency: the algorithm applies the same criteria to all messages, eliminating individual variations generated by fatigue, subjectivism, or cognitive overload.

Where the logic of algorithmic neutrality breaks down

Each of these advantages comes with a specific cost. Speed also means irreversibility: a label applied incorrectly can be technically retracted, but its social effect, discrediting an author, limiting the distribution of a message, remains. Scaling means that every system error reproduces at the same scale as the advantage it promises. Formal consistency masks a deeper risk: the algorithm applies uniformly, but does not apply correctly if the rules with which it was trained are themselves distorted. Algorithmic bias is, in this sense, the central structural risk. Artificial intelligence models are not neutral: they are trained on data, and the data reflect the imbalances of the world from which they come. If the dominant sources in the training set belong to a certain geopolitical or cultural space, the system will tend to implicitly validate the narratives from that space and penalize positions coming from underrepresented areas. The geographical distribution of sources that dominate the discourse about "fake news," massively concentrated in Anglo-Saxon media and, secondarily, in Brazilian and Eastern European media, suggests that a model trained on this corpus would implicitly reproduce an already existing editorial hierarchy, rather than create a new neutrality. Classification errors add another level of risk. Political language is often ironic, metaphorical, deliberately ambiguous, or hyperbolic. A system that semantically reduces a text to verifiable claims constantly misses the real meaning of messages that do not operate in a literal register. Claims that are partially correct, which include a real factual core and a manipulated context, are precisely the most difficult to detect automatically and, at the same time, the most effective as tools of disinformation. Labeling them as "partially true" can paradoxically function as an indirect validation. Opacity completes the picture. Without public access to the evaluation methodology, to the reference sources used, and to the procedures for contesting a verdict, the user faces a technical oracle that they can accept or reject, but cannot examine. This is a form of editorial power without assumed editorial responsibility, a combination that, in any other public institution, would raise serious questions.

Algorithmic centralization versus the distribution of authority

There is a fundamental distinction between centralized and distributed models of managing disinformation, and it does not stem from technology, but from the political philosophy of information. The centralized model, in which a single platform or system applies a uniform verdict to all messages, has the advantage of efficiency and coherence. However, it has the disadvantage of concentrating power: whoever controls the algorithm indirectly controls which narratives are credible and which narratives are marked as suspect. In a media ecosystem where the discourse about "fake news" is itself a political weapon used from conservative American press to judicial institutions in Brazil, from Karoline Leavitt to Emmanuel Macron, the concentration of labeling power in a single system raises real democratic issues, regardless of the good intentions of those who build it. The distributed model functions differently. Verification is carried out by multiple independent actors: specialized newsrooms, fact-checking consortia, academic communities, public institutions. The conclusions are visible as distinct editorial products, with explained methodology and the possibility of public contradiction. Disagreement among verifiers is not an error, but a faithful reflection of the complexity of reality. The model is slower, more expensive, and more difficult to scale, but it produces a type of authority that is more legitimate, because it is more transparent and contestable. Between these extremes, a hybrid approach has also emerged, in which algorithms function as tools for prioritization and detection, without producing the final verdict. They identify potentially problematic content, map coordinated distribution networks, and highlight clusters of suspicious messages, after which human analysts intervene for substantive evaluation. The system gains speed and coverage without giving up contextualized judgment. It is, so far, the model that best balances efficiency with responsibility, although it remains vulnerable to the quality of human selection and institutional pressures.

Are we reducing disinformation or just moving it to a less visible level?

The distribution of the tone of public discourse regarding "fake news" in the last six months, 66% opinionated, 31% negative, only 3% positive reflects a chronic frustration with the phenomenon and, at the same time, a heightened expectation for solutions. Technical AI-type solutions respond exactly to this expectation, and their legitimacy derives, to a large extent, from the perceived urgency of the problem. It is precisely this urgency that deserves critical examination: the need to do something quickly does not guarantee that what is done is correct or sufficient. Automated labeling systems can reduce the visibility of certain types of content, can signal to users that a claim is contested, and can limit the speed of dissemination of obviously false messages. But they cannot rebuild trust between the public and the information-producing institutions, cannot eliminate the polarization that makes part of the public interpret any correction as a form of censorship, and cannot replace individual information culture. Algorithmic labels do not change beliefs; they can, at most, introduce minimal friction in the distribution of content. Moreover, there is a systemic risk that is not addressed in the current debate: if a limited number of algorithmic systems come to play the role of de facto standard in assessing the veracity of information on the internet, their combined power becomes comparable to that of any editorial monopoly, with the difference that it is less visible and harder to contest. The concentration of discourse about "fake news" around a few dominant sources and platforms, which data confirms, may be a precursor to such algorithmic centralization on a global scale. The working conclusion is that automated disinformation detection systems are useful tools, but insufficient and potentially dangerous if used in isolation. Their real efficiency depends on the transparency of the methodology, on integration with independent human verification structures, on robust contestation mechanisms, and on a clear separation between filtering erroneous factual content and influencing the editorial orientations of the platform. Without these conditions, the fight against disinformation risks producing not more clarity, but a new layer of opacity – harder to identify and, therefore, more difficult to correct than the problem it claims to solve.

*****Synthesis made with the help of a data monitoring stream provided by the media monitoring platform NewsVibe Romania. The analysis, data, and images presented have been enhanced with the help of Machine Learning and Artificial Intelligence tools.