Twitter/X: The Ultimate Forensic Unofficial Biography

Chapter 13: 19 Countries

I thank my readers for their recommendations and my team for preparing this compendium.

This is my new book that I believe is timely on the basis of recent events. It would not have been realised so quickly if it were not for the work my team did.

This is a long article; Chapter 13 from my new book Twitter/X

The researcher’s apartment in Berlin was small and cold in February, and the heating had been arguing with itself since November, producing warmth in intermittent bursts sufficient to sustain work but not to forget that it was winter.

She had three screens open and a fourth dataset loading, and the coffee she had made forty minutes ago was sitting untouched at her elbow going through the stages of cooling that she had learned, in three years of this work, to measure by touch without looking, to use as a rough clock for how long she had been inside a particular set of numbers without coming up for air.

The data was from Brazil. She had been looking at the weeks leading up to the October election, the amplification figures for specific account categories, the spread of specific content types, and the relationship between algorithmic distribution and the geographic location of the accounts receiving it. The pattern was the same. The same pattern she had found in the data from Türkiye. The same pattern she had found in the data from Hungary, the Philippines, and the eight other countries she had completed before Brazil.

She picked up the coffee. Cold. She drank it anyway. She opened a new document and began to write, slowly and carefully and with the precision that the data required and the moment, she understood, warranted.

Outside the window, Berlin was going about its business in the grey February light.

In the document, the numbers told a story: nineteen countries had lived without knowing they had been written about.

The data does not have a nationality.

It does not arrive with a flag, a political affiliation, or a stated preference for any particular outcome. It arrives as numbers, as the accumulated record of what the platform amplified and what it suppressed and who it made more visible and who it made less visible, measured across time and across geographies and across the political contexts of the countries where the platform operated, and the numbers are what they are regardless of the interpretation that any particular analyst brings to them or the conclusion that any particular interested party prefers they support.

The numbers, studied by the researchers who had access to the methods available after the closure of the academic data programme, in the countries where the researchers were watching, and the political situation was legible enough to allow the comparison between platform behaviour and political outcome that the study of algorithmic influence requires, told a story.

The story was not simple. It was not the story of a platform programmed to produce a specific political outcome in a specific election in a specific country. It was more diffuse than that, more structural, more consistent across contexts that were otherwise very different from each other, consistent in the way that structural phenomena are consistent, the way that a river runs downhill regardless of the specific geography it is running through, because the force producing the movement is not local but systemic.

The systemic force in the countries the researchers were observing was amplification.

The amplification finding had first emerged in research conducted before the acquisition, in 2021 studies of the platform’s recommendation algorithm that examined the political content of the tweets the algorithm amplified relative to the political content of the tweets users had followed and chosen to receive.

The study, conducted by Twitter’s own researchers in collaboration with external academics, examined the algorithmic amplification of political content in seven countries, Canada, France, Germany, Japan, Spain, the United Kingdom, and the United States, and found that in six of the seven countries the algorithm amplified content from right-leaning political accounts more than content from left-leaning political accounts, and that the amplification was not marginal but substantial, that the content of right-leaning accounts was reaching audiences significantly larger than the content of left-leaning accounts whose organic follower base was equivalent.

The study was published by Twitter’s own research team. It was published with the caveat that the researchers did not fully understand why the algorithm produced this outcome, that the outcome was not the result of deliberate design, that the recommendation system optimised for engagement and the engagement patterns of the platform’s users produced, through processes the researchers could describe in their outputs but not fully explain in their mechanisms, a consistent skew in the direction of right-leaning political content.

The platform’s own researchers documented, in a peer-reviewed study published by the platform itself, that its algorithm was amplifying one side of the political spectrum more than the other.

The study was noted by the journalists and the researchers who followed this work closely. It was not widely read. It did not produce the policy changes that its findings implied were warranted. The algorithm continued to operate. The engagement continued to be measured. The amplification continued to occur.

That was before the acquisition.

After the acquisition, the researchers who were studying amplification patterns worked with the data available to them, the public API, and the methods for sampling and analysing public content that the pricing changes had not yet fully foreclosed, and found not a reversal of the pre-acquisition pattern but an acceleration of it.

The acceleration was documented across multiple studies, by multiple research groups, in multiple countries, using methods sufficiently varied that the convergence of their findings carried more weight than any single study could alone. When researchers working independently, in different institutions, with different methodologies, looking at different countries, arrive at findings that point in the same direction, the direction itself becomes the finding.

The direction was consistent.

Right-leaning political content, specifically content from right-leaning political accounts that had been reinstated following the acquisition, was receiving amplification at levels that the organic reach of those accounts, based on their follower counts and the historical engagement patterns of accounts of comparable size, would not have predicted. Left-leaning political content, specifically content from accounts associated with fact-checking, independent research, and journalism critical of the platform or its owner, was receiving distribution at levels below what the organic reach of those accounts would have predicted.

The gap between what organic reach would predict and what algorithmic amplification produced is the measure of the algorithm’s effect. In a platform that processes hundreds of millions of pieces of content each day, small percentage differences in amplification produce, at scale, very large differences in the audiences that specific content reaches and the audiences that other content does not reach, differences large enough to be meaningful in the specific political contexts where the content is operating and where the difference between reaching an audience of one hundred thousand and an audience of one million is the difference between a message that circulates within a community and a message that shapes a national conversation.

Brazil had an election in October 2022, three days after the acquisition closed.

Luiz Inácio Lula da Silva defeated the incumbent Jair Bolsonaro by a margin of less than two percentage points, the closest Brazilian presidential election in decades, in a campaign that had been conducted in an information environment shaped substantially by the social media platforms on which both candidates and their supporters were active and on which, the researchers who studied the campaign documented, the amplification dynamics were consistent with the pattern the studies were beginning to describe.

Bolsonaro’s supporters did not accept the result. In the weeks following the election and into the new year, the claims that the election had been stolen, that the electronic voting system had been compromised, and that the result did not reflect the genuine will of the Brazilian people circulated on the platform with the velocity of content that the amplification mechanisms were treating with the generosity that the pre-acquisition research had documented for right-leaning political content.

On the eighth of January 2023, two years to the day after the events at the United States Capitol, a crowd of Bolsonaro supporters stormed the presidential palace, the National Congress, and the Supreme Court in Brasília.

The symmetry of the dates was noted. The similarity of the dynamics, the claims of a stolen election, the social media amplification of those claims, and the movement of a crowd toward the institutions of democratic governance were noted. The researchers who had been studying the information environment of the Brazilian election published their findings, which were noted, and the platform on which the amplification had occurred continued to operate.

The European Union had been watching.

The Digital Services Act, which came into force for the largest platforms in August 2023, imposed obligations on platforms that met the threshold for designation as Very Large Online Platforms, obligations that included the requirement to assess and mitigate the systemic risks that their services posed to democratic processes, public security, and fundamental rights, and to make available to vetted researchers the data necessary to audit their compliance with those requirements.

Twitter, designated as a Very Large Online Platform, was subject to the Act.

The European Commission opened a formal investigation into Twitter in December 2023, citing concerns about the platform’s compliance with the Act’s requirements on risk assessment, content moderation, transparency, and researcher data access. The investigation was ongoing. It was being watched by the governments of the member states and by the research community and by the advocacy organisations that had been documenting the platform’s behaviour and by the platform’s own legal team, which was staffed differently than it had been before the acquisition, which is to say more thinly than the complexity of its regulatory obligations required.

Musk’s response to the European regulatory process was consistent with his response to the domestic regulatory process and to the research findings and to the advertiser concerns and to the journalist coverage, which was to engage with it combatively on the platform, to describe the regulators in terms that reflected the governing framework of his public communications, which was the framework of free speech under attack from entities that could not tolerate its restoration, and to instruct his legal team to contest the requirements in ways that produced extended timelines and continuing uncertainty about the platform’s compliance status.

The European Commissioner responsible for the Digital Services Act sent Musk a letter. Musk published it on the platform. The publication of the letter was itself a communication to the audience that was watching, a signal about the posture the platform’s owner intended to maintain toward the regulatory process, which was the posture of someone who considered the regulatory process an obstacle rather than an obligation.

India had fifty-seven million Twitter users at the time of the acquisition, one of the largest user bases outside the United States, a population of users operating in a political context in which the relationship between social media platforms and the government of Narendra Modi had been the subject of sustained documentation by the researchers and the journalists who covered platform politics in South Asia.

The Indian government had, in the years before the acquisition, sent Twitter an increasing volume of content removal requests, requests to remove tweets and suspend accounts that the government had identified as inconsistent with Indian law, requests whose volume and character the researchers who studied government pressure on social media platforms found instructive about the government’s relationship to the concept of free speech on the platforms it was requesting to censor.

Twitter, under the previous management, had complied with some of these requests and resisted others, had published its compliance data in transparency reports, and had, in specific cases involving accounts documenting the treatment of protesters and journalists, declined to comply in ways that had produced legal consequences for the platform’s local employees.

After the acquisition, the transparency reporting on government requests changed. The detail available to the researchers who had previously used that reporting to study the relationship between platform behaviour and government pressure decreased. The compliance data that had enabled the comparison between what governments requested and what the platform provided became less granular, less frequent, and less useful for independent assessment.

What remained was the data on account behaviour, and the data on account behaviour, studied by the researchers who were watching India with the attention that the size of the user base and the significance of the political context warranted, described amplification patterns consistent with what was being documented in other countries, patterns in which the content aligned with the governing political tendency received distribution at levels the organic reach of those accounts would not have predicted.

India had elections in 2024. The elections were the largest democratic exercise in human history, with 969 million eligible voters, a process conducted over six weeks across a country whose diversity of language, geography and political culture makes the administration of a national election an achievement of extraordinary logistical complexity. The Bharatiya Janata Party of Narendra Modi won a majority, though a smaller majority than the pre-election polling had predicted, in an information environment that the researchers who studied it were still, months after the result, in the process of fully describing.

The nineteen countries in which the researchers found patterns consistent with the amplification of right-leaning and government-aligned political content and the suppression of opposition, fact-checking, and critical journalistic content were not geographically contiguous or politically homogeneous. They spanned continents, political systems, and cultural contexts that had almost nothing in common except their presence on a platform whose algorithm operated according to the same mathematical instructions, regardless of the national context in which those instructions produced their outputs.

The countries were Brazil, India, the United States, the United Kingdom, France, Germany, Italy, Spain, Argentina, Mexico, Turkey, Hungary, Poland, Israel, South Africa, Nigeria, Indonesia, Australia and the Philippines.

In each, the researchers found evidence of the same pattern. The pattern was not identical in each country. The specific accounts amplified and suppressed, the specific content categories favoured and disfavoured, and the specific relationship between platform behaviour and political outcome varied with the local context in ways that reflected the complexity of the political situations in which the platform operated.

But the direction was the same.

In Hungary, where Viktor Orbán had been constructing the architecture of illiberal democracy for fifteen years, where the independent media had been systematically captured or marginalised, where the opposition operated in an information environment that the researchers who studied it described as the most comprehensively hostile to democratic competition in the European Union, the platform’s amplification patterns favoured the governing party’s content and suppressed the content of the journalists and the researchers and the opposition figures who were documenting what the governing party was doing.

In Turkey, where Recep Tayyip Erdoğan had been consolidating power through a combination of electoral success, institutional capture, and the systematic marginalisation of critical voices, the pattern was consistent with what was being documented in Hungary, consistent in the direction of its effects if not in the specific mechanisms that produced them.

In the Philippines, where Ferdinand Marcos Jr had won the presidency in 2022 in a campaign that the researchers who studied it described as the most sophisticated deployment of social media disinformation that Southeast Asian politics had yet produced, the platform’s amplification of the content associated with the Marcos campaign and the suppression of the fact-checking accounts that had spent years documenting the historical revisionism at the centre of that campaign was documented in sufficient detail that the relationship between platform behaviour and electoral outcome was, in the assessment of the researchers who had studied both, not a coincidence.

The word coincidence deserves attention here, because it is the word that the defence of the platform’s behaviour in all of these countries has relied upon, and because the examination of that reliance produces a finding that the word is insufficient to carry.

A coincidence requires an absence of connection between the things that have coincided. Two things happen at the same time, in the same place, without a causal relationship between them. The word coincidence applied to the amplification patterns across nineteen countries would require that the algorithm produced consistent outputs in consistent directions across consistent political contexts without a common cause, that the mathematical instructions operating in Brazil produced the same directional bias as the mathematical instructions operating in Hungary and Turkey and the Philippines and the other sixteen countries not because the mathematical instructions were the same but because the political contexts, despite their enormous differences, happened to independently produce equivalent outcomes from the same algorithm.

That argument is available. It has been made. It requires, for its acceptance, a degree of credulity about the neutrality of algorithmic systems that the research literature on algorithmic amplification, accumulated over a decade by researchers who approached the question with the scepticism that the scientific method requires, does not support.

The algorithm has no opinion.

But the people who wrote the algorithm had objectives. And the people who modified the algorithm after the acquisition had a different set of objectives. And the outputs of the algorithm in the nineteen countries the researchers were monitoring were consistent with those objectives in ways that the word coincidence was not designed to describe.

Elon Musk, in the months following the acquisition, had become something that the platform’s previous owners had never been, which was a political actor, a person whose endorsements and criticism and amplification and suppression of specific politicians and specific movements and specific pieces of political content were tracked by the political researchers and the political journalists and the political operatives in countries around the world with the same attention that they tracked the endorsements and the resources of any other significant actor in the political environments they were studying.

He endorsed politicians. He endorsed them on the platform to an audience whose size had been supplemented by algorithmic amplification beyond what any organic follower count could explain. He criticised politicians. He criticised them on the platform, directing his audience’s attention to them with the specific vocabulary of a communicator who, from years of experience, understood what framing produced the engagement that the platform’s architecture would then amplify into the broader information environment.

In Germany, in the weeks before the federal election of February 2025, he hosted a conversation on the platform with Alice Weidel, the leader of the Alternative für Deutschland (AfD), the far-right party whose positions on immigration and European integration and the cultural identity of Germany had placed it at the edge of what the German political mainstream considered acceptable democratic discourse. The conversation reached an audience of tens of millions. The endorsement of Weidel and of the AfD that accompanied it was made explicit. The AfD came second in the election with its highest ever share of the vote.

In the United Kingdom, he involved himself in the debate about the grooming gangs inquiry with a consistency and an intensity that the British journalists who covered the story documented in detail, amplifying specific narratives and specific actors in the debate in ways that the researchers who studied online political communication found consistent with the amplification patterns they had been documenting elsewhere.

In the United States, he had spent two hundred and seventy seven million dollars supporting the campaign of Donald Trump in the 2024 presidential election, had been given a position in the incoming administration overseeing government efficiency, and was using the platform whose algorithm he controlled to communicate with an audience whose size the algorithm ensured, to an audience that was receiving his content whether or not they had chosen to follow him.

The platform’s owner was a political actor. The platform’s algorithm amplified the owner’s content. The owner was using the platform’s algorithm to amplify the political content of politicians he supported. The politicians he supported were, in the countries where the researchers were watching, the politicians whose success the amplification patterns in those countries were consistent with producing.

These are not opinions. They are the documented outputs of a system that had been built to be neutral and had become, through a sequence of decisions each of which was describable in the record, something other than neutral.

The people who had built the thing in the warehouse, who had stayed up through the nights sustained by the conviction that what they were building would make the world more connected, more equal, more free, those people were not, in the main, still at the company.

Some of them had left voluntarily, in the years before the acquisition, when the distance between the founding vision and the operational reality had become too large to bridge with the language of ongoing commitment to the original purpose. Some of them had been let go in the waves of reduction that had followed the acquisition. Some of them had stayed until they could no longer stay, until the specific moment arrived when the thing that had happened to what they had built became too legible to continue working within.

They were watching from outside now. Watching the nineteen countries. Watching the amplification data. Watching the political outcomes in the places where the platform’s behaviour was most legible in its relationship to those outcomes.

Some of them spoke to the journalists documenting the period the way people speak to journalists when the alternative is silence, and silence has, in these circumstances, become a form of complicity. They spoke carefully, aware of non-disclosure agreements and of the fact that the institutions they had left had the resources to make speech costly, but they spoke.

What they said, in the various forms people use to say things they have been holding in for a long time, was not complicated.

They said that what had been built was not what they had intended to build.

They said that the distance between what they had intended and what existed now was not a distance that could be measured in any single decision or any single year but in the accumulated weight of the choices that had been made, one at a time, in the conference rooms with the themed names, in the advertiser meetings, in the moderation decisions, in the funding rounds, in the acquisition, in the reinstatements, in the algorithmic modifications, in the lawsuits, in the firings.

They said that the bird had been real.

And they said that what had replaced it was not.

Aldo Grech is the author of HOW: Elections Are Won in the Digital Age, The Great Populism Hustle, and Kleptocracy, among others. He writes on democratic erosion, elite capture, and the mechanics of political manipulation.

”The future is embedded in the choice”. **Books and [private advisory](http://www.aldogrech.com/)**.