Meta in Brazil, Access Bans in Turkey: An Algorithmic History and Technical Anatomy of Elections
Who Holds the Key to the Second Ballot Box?
The Invisible Ballot Box
Meta in Brazil, Access Bans in Turkey: An Algorithmic History and Technical Anatomy of Elections
Dear Young Comrades,
Every election has two ballot boxes. One you can see: a sealed, transparent plastic box placed in a school classroom. The other you cannot see: a piece of software that decides what voters see before they cast their vote, whom they hear, and what comes up when they search for a candidate's name. At the first ballot box stand observers, party representatives and the election board. At the second there is no one. Its key is held by the engineers of a company in California.
Documents leaked from Brazil this week showed, more clearly than ever before, how that second ballot box was tilted in one direction in the middle of an election. In this piece we first explain what happened there. Then we show how this environment was created, through a short history running from Obama to Brexit and from Cambridge Analytica to Romania. Next, through the eyes of a computing worker, we open up the question of "what is technically being done". Finally we return to our own country: what happened in Turkey's past elections, and what should we expect in the next one?
In Brief: What Happened, and What Are We Arguing?
- What happened in Brazil? Ahead of the presidential run-off on 25 October, Folha de S.Paulo published internal messages between Meta employees. The messages show that the company knew Lula's Instagram account really had been restricted while his rival Flávio Bolsonaro was being boosted by its recommendation systems, and that it fixed the problem late. On a single day, the Lula campaign was unable to reach roughly 1.3 million people.
- Was intent proven? No. There is no line in the messages saying "let's interfere". The electoral court also wrote that there was no evidence of intent, but it ordered Meta to treat the candidates equally under threat of a daily fine of 500 thousand reais.
- What is the real issue? Intent is not required. If the fairness of an election has come to depend on how quickly engineers in another country fix which error, the problem lies not in the error itself but in who owns the channel.
- How was this environment built? The revenue model of social media is to classify people's attention according to their behaviour and sell it. A tool sold to advertisers is also sold to campaigners. Since 2008, every election has been a testing ground in which this tool has been sharpened a little further.
- What is technically being done? There are interventions from outside, such as micro-targeting, fake account networks, doctored videos and deepfakes. And there is what the platform can do from inside: throttling reach, tuning search and recommendations, controlling ad delivery. States join the picture with country-based withholding and bandwidth throttling.
- What is the situation in Turkey? Turkey has no platform–state conflict like Brazil's; on the contrary, there is an alliance. The state decides, the platform implements. On the eve of the 2023 election, Twitter shut down accounts at the government's request. In September 2026 alone, at least 1,270 X accounts were blocked.
- What is expected in the next election? The censorship infrastructure is being built before the election, anonymity is being abolished through e-Devlet verification, AI fakery is getting cheaper, and the platforms are staying silent. But the real intervention will most likely take place not on social media but in the courts.
Now let us turn to the details.
Brazil: What Do the Leaked Messages Say?
In Brazil's first round on 4 October, Senator Flávio Bolsonaro, son of the far-right former president Jair Bolsonaro, finished first with a stronger result than expected. Lula, seeking a fourth term, was pushed into a run-off. The second round will be held on 25 October.
On 11 October the Lula campaign announced that the reach of Lula's content on Meta's platforms had been restricted and applied to the electoral court (France 24). Meta rejected the accusation as "completely false". At that point all there was to go on was one campaign's claim and one company's denial.
On 16 October, Folha reporter Patrícia Campos Mello published internal messages between Meta employees. The story quickly spread to the international press (Al Jazeera, ICL Notícias, Revista Fórum). Let us set out what the messages reveal in chronological order:
| Date | What happened? | Meta's explanation / internal messages |
|---|---|---|
| 23 September | Both campaigns' ad accounts were unable to upload ads for about an hour, on grounds of "suspicious activity" | Automated system error |
| 1 October | Some of the Flávio campaign's ads were rejected on the grounds that AI had been used during the restricted period | The campaign said it had not used AI |
| Early October | In Instagram search, Flávio's profile is recommended, while Lula's does not appear even when searched by his full name | "Lula's name is too long" (technical explanation); employees admit Flávio should not have been recommended |
| Early October | Searching WhatsApp for the "Exército do Lula" channel brings up the Flávio supporters' channel "Tropa do Flávio" in the results | Electoral rules ban algorithmic recommendation of candidates |
| 7 October | Some accounts cannot see their view counts | "A global data outage affecting about 3 thousand accounts" (according to internal messages, it lasted 5 hours) |
| 9 October | Lula's official account is unable to boost feed posts as paid content for about 24 hours; by the campaign's estimate, 1.3 million people are not reached | An employee writes that Lula's account is "restricted" compared with Flávio's and that the likelihood of legal action is "high" |
| 9–10 October | The Brazil team asks for the incident to be classified as "SEV2" (a serious incident) | Employees complain that engineers in the US show no sense of urgency |
| 14 October | In a test by Revista Fórum, Meta AI refuses to generate election images for Lula but produces campaign material for Flávio | "An error in the application of safety filters" |
Look at the table carefully: some of the glitches hit both candidates. The Flávio campaign was also unable to advertise for about an hour on 23 September, some of its ads were rejected, the campaign said "our account is working normally", and it acknowledged that Brazilian rules could affect reach. We are not hiding this, because we do not want to be the side that hides evidence. But the biggest and longest impact came in a single direction, in the most critical week of the run-off: a 24-hour advertising block that the company tried to cover with a 5-hour global outage.
What did the court say?
On 15 October, Judge Marco Aurélio Bellizze of the Superior Electoral Court (TSE) ruled as follows:
- Meta may not reduce the distribution of political content, restrict search or remove posts in a way that creates an imbalance between the two candidates.
- The AI tool must apply "the same criteria" to both candidates.
- Automated blocks on ad accounts will go through human review.
- Meta will make up the ad reach not delivered to the Lula campaign by 21 October and document its "technical error" claim.
- 255 suspicious profiles and 1,204 posts will be examined.
- Non-compliance will incur a daily fine of 500 thousand reais (about 98,700 dollars).
In the same ruling, the judge also wrote that there were not sufficient elements to prove that Meta had engaged in deliberate manipulation. The ruling is valid until the end of the run-off and will be heard by the full TSE.
What has been proven, and what has not
Here we need to make an honest distinction, because on social media this story is circulating under the headline "Meta is trying to get Bolsonaro elected".
Proven: The restrictions were real. The most serious one ran in one direction. The company knew it. Brazilian employees raised the alarm, and the fix came late. The recommendation system recommended a candidate in breach of electoral rules.
Not proven: That this was a conscious political decision. There is no admission of interference in the messages.
We will not claim what has not been proven. But we will show that the question of intent is not the real question in this case.
Intent Is Not Required
The liberal framework weighs an incident by asking "was there bad intent?" If bad intent is not proven, the case is closed. The Marxist framework asks a different question: whose property is the structure that produces this result, under whose control is it, and in whose interest does it work? Looking with this question, three things come into view.
An error is political too
In a tight run-off, failing to reach 1.3 million people in a single day is large enough to change the result. A software error is not something that arises spontaneously in nature. Which error works in which direction, how quickly it is noticed, with what priority it is fixed: each of these is a decision. The "technical fault" defence loads responsibility onto the system itself. And the owner of that system is Meta. Whoever owns the system owns its errors too.
The core–periphery hierarchy works inside the company too
Brazilian employees raised the alarm, asked for the incident to be treated as serious and complained that its urgency was not understood. But the authority to decide and intervene lay in California. The fairness of one country's election was left to the prioritisation list of engineers in another. Here imperialism does not appear in the form of a coup or an occupation. It works as a queue of support tickets. The worker in the peripheral country warns; the engineer at the core keeps them waiting in line.
Oversight is possible only through leaks
These problems were not found by an independent auditor; they were exposed by leaked messages. Even the court, in examining 255 profiles and 1,204 posts, has to look at Meta's own records. How the channel between voter and candidate works is reported by the company that runs the channel. At the first ballot box there are observers. The report on the second ballot box is written by the owner of the box.
Searching for intent is the politest way of not asking the question of ownership.
Context: This Did Not Happen in a Vacuum
To say a company has no political preferences would be naive optimism. Meta's, and the big tech monopolies', relationship with Brazil has a history:
- Zuckerberg's January 2025 turn. Announcing that he was scrapping the third-party fact-checking programme in the US, Zuckerberg openly targeted the "secret courts" in Latin America "that can order companies to quietly take things down". This was a message aimed directly at Brazil's Federal Supreme Court. In the same announcement he also said they would start recommending political content again. Yet in February 2024 Meta had announced that it would stop recommending political content by default on Instagram and Threads. How much political content is shown to whom is a setting that a single company changed twice in one year.
- The Trump administration and the Bolsonaro family. Jair Bolsonaro was convicted in September 2025 for refusing to recognise the 2022 election and leading a coup attempt. Citing this trial, the Trump administration imposed a 50 per cent tariff on Brazilian goods in July 2025 and sanctioned Alexandre de Moraes, the judge hearing the case. Today it is Flávio who carries the same family's political legacy.
- Platforms' direct participation in Brazilian politics. In May 2023, Google ran a campaign from its search page against Brazil's platform regulation bill (PL 2630). Musk's X was shut down in Brazil for weeks between August and October 2024 for failing to comply with court orders.
So what we have is a company that has openly shown its political preferences, a judiciary that this company has declared an "enemy", and a US administration in conflict with that judiciary. Even if intent is not proven, the material ground for suspicion is solid.
How Was This Environment Created?
Let us step back now and ask: how was a world built in which a company can influence an election this much?
You are the product
What social media platforms sell is not an app. They give the app away for free. What they sell is your attention: who you are, what you click on, how many seconds you linger on which post, whom you talk to. With this information they sort you into thousands of categories and sell access to these categories to advertisers. In The Revolt of Crystallized Labor we called this "digital enclosure": the fencing off of social interaction, of shared language, of what we say to one another, into private property.
Applying this model to politics is not a deviation but an inevitable outcome. A targeting machine built to sell detergent also works to sell candidates. The line between "legitimate campaigning" and "interference" is drawn not by technical but by political criteria: who is using the same tool, and against whom?
Monopoly, channel, gatekeeper
The second condition is monopolisation. Most of a society's political conversation runs on the infrastructure of a handful of US companies. Meta alone owns Facebook, Instagram and WhatsApp. If three of a candidate's channels for reaching voters in a country are in the hands of the same company, that company is not an intermediary but a gatekeeper. In What Does Mark Zuckerberg Actually Want? we discussed Meta's feed algorithm as "a tool that shapes the user's goals". Brazil is what that looks like on election day.
A short history: Elections as testing grounds
This environment was not built in a day. Let us recall the main milestones in order.
Facebook's voting experiment (2010). During the US midterm elections, Facebook showed 61 million users an "I Voted" button along with photos of friends who had voted. According to the study published in Nature, this single button produced roughly 340 thousand additional votes. The experiment showed that a company, accountable to no one, could change voter turnout in whichever groups it chose.
Obama, 2008–2012: "good" micro-targeting. 2008 was dubbed "the first social media election". Facebook co-founder Chris Hughes was part of the campaign's digital operation. In 2012, the campaign's Facebook app also collected data on the friends of supporters who had authorised the app. None of those friends had consented. In 2018, the campaign's director of media analytics, Carol Davidsen, described this openly: Facebook had been surprised that they pulled the entire social network, but having realised it, did not stop them. The liberal press called this a "data revolution".
Brexit (2016). The Vote Leave campaign channelled most of its budget to AggregateIQ, a Canadian firm linked to SCL, the parent company of Cambridge Analytica. Campaign strategist Dominic Cummings himself described how roughly a billion targeted digital ads were shown in the final days. The Electoral Commission found that the spending limit had been exceeded through a front group, and Vote Leave was fined 61 thousand pounds. A note of correction: the UK Information Commissioner's Office's 2020 report showed that Cambridge Analytica was not as decisive in Brexit as had been thought. The real issue was not a single company but a campaign machine built on money and data, whose oversight remained on paper.
US 2016: Cambridge Analytica. About 270 thousand people used an academic's personality-test app. Through these people's friends, the app collected the Facebook data of 87 million people. The data was put at the service of the Trump campaign. It was the very same door the Obama campaign had walked through in 2012. In 2019 the US Federal Trade Commission fined Facebook 5 billion dollars. Given the size of the company, a figure that can be counted as a cost of doing business.
US 2016: Russian trolls. Russia's Internet Research Agency operated through fake accounts. However, a large-scale study published in Nature Communications in 2023 showed that this content reached a small group of people who were already strongly Republican, and found no measurable effect on attitudes or voting behaviour (Eady et al.). The "Russia elected Trump" narrative, worked over for four years, was a convenient way of not talking about the domestic class collapse that led American society to Trump.
Peripheral countries as laboratories. SCL, Cambridge Analytica's parent company, first tested its methods in peripheral countries. In Trinidad and Tobago (2010), the campaign the company boasted of in its own presentations aimed to steer the young people of one ethnic group away from the polls. It worked on Kenyatta's campaigns in Kenya (2013, 2017) and in Nigeria (2015). In Brazil in 2018, Folha revealed that employers supporting Bolsonaro had bought bulk messaging packages over WhatsApp. Under Brazilian law, this amounted to a banned corporate donation. The pattern is clear: develop the method in peripheral countries where regulation is weak, then sell it back to the core.
The market for paid interference. In 2023 the Forbidden Stories consortium exposed the Israeli "Team Jorge" outfit. The team claimed to have worked on 33 election campaigns, 27 of them "successful". Its methods were bot networks, hacking and the production of fake news. In 2024 the New York Times reported that Israel's Ministry of Diaspora Affairs had commissioned a firm, under a contract of roughly 2 million dollars, to run a fake-account campaign targeting US members of Congress. Interference is now a service industry.
The West's own operations. The selectivity of the "foreign interference" discourse is most visible here. Between 2010 and 2012, the US Agency for International Development (USAID) secretly set up and ran a "Cuban Twitter" called "ZunZuneo" in Cuba. In 2024 Reuters documented that the US military had run a fake-account campaign in the Philippines during the pandemic against the Chinese vaccine. These did not make the headlines the way Russian or Chinese operations did.
Romania 2024: the concept turned into a weapon. In December 2024, Romania's Constitutional Court annulled the first round of the presidential election on the grounds of a TikTok campaign run in favour of the far-right Georgescu. The justification was intelligence documents alleging a Russian-backed operation. Later investigations showed that part of the campaign was a hashtag campaign paid for by a centre-right party, which was subsequently hijacked by others. Two dangers stand side by side in this example. There is real manipulation. On the other hand, the "disinformation" justification can lend legitimacy to the state apparatus annulling the will expressed at the ballot box.
Musk and X: when the owner is the actor. After Musk bought Twitter, it was debated whether his own posts were being boosted by the algorithm. In the 2024 US election he gave Trump direct financial and campaign support. In January 2025, ahead of the German election, he held a live broadcast with AfD leader Weidel. Here there is no intermediary any more. The owner of the channel becomes a political actor.
The point this history arrives at is Brazil. Instead of third parties (campaigns, consultants, states), it is the platform's own algorithm, recommendation system and AI tool that take the stage. The power shown by the "I Voted" experiment in 2010 worked this time in an election, documented and in one direction.
What Is Technically Being Done? An Explanation from a Computing Worker
Newspaper language says "algorithm", "bot", "manipulation" and moves on. Let us open things up a little, because without knowing the mechanism we cannot know what to demand. We divide the tools into three groups: those used from outside, those the platform can use from inside, and those in the hands of the state.
Tools used from outside
Social graph harvesting. An app asks the user for permission, but the permission extends not only to the user but to their friends' data as well. Facebook's application programming interface (API), open until 2014, allowed this. That is how the data of 87 million people was collected with the permission of 270 thousand. In computing terms this is called the "social graph": the map of who knows whom and who talks to whom.
Micro-targeting. The collected data splits voters into small groups: "men over forty, living in rural areas, who react to news about migrants", for example. Each group is sent a separate message. One group is told about economic anxiety, another about fear for its identity. A campaign can say something different to everyone, and no one sees what was said to anyone else.
Lookalike audiences. If you have a list of a thousand supporters, the platform will find you a million people who "look like" them. The campaign does not even need to collect data. The platform does the targeting with its own data.
Dark posts. These are ads that never appear on the campaign's page and land only in the feed of the targeted person. They come up in the Brazil messages too: on the day the Lula campaign was unable to boost feed posts, it could still advertise with dark posts. This is a kind of politics that public debate cannot scrutinise.
A/B testing. Dozens of versions of the same ad (headline, colour, image, wording) are tested on small groups, and the one that gets the most reaction is scaled up. Trump's 2016 digital campaign described running tens of thousands of ad variations a day. The message is no longer written by the campaigner but by click data.
Fake account networks and coordinated behaviour. Hundreds or thousands of accounts that look like real people spread the same hashtag at the same hour, attack the same journalist at the same moment, share the same story with the same comment. The aim is less to persuade than to make a topic look like it is "trending" and to intimidate the targeted voice. Platforms call this "coordinated inauthentic behaviour" and from time to time shut these networks down. They decide for themselves when to shut them down and when not to.
Inflating engagement and views. Raising a video's view count or a post's likes with purchased engagement makes it "popular" in the eyes of the algorithm. And the algorithm shows what it believes to be popular to more people. Fake demand turns into real distribution.
Bulk broadcasting over closed messaging. No one outside can see or verify content circulating on end-to-end encrypted channels such as WhatsApp. As in Brazil in 2018, services that push content to thousands of groups at once can be bought. The lie arrives as a relative's post in the family group and looks trustworthy.
Doctored videos and deepfakes. Fake audio, fake video, footage torn out of context. Generative AI has brought the cost of this close to zero. It has a second effect too: when a genuine recording emerges, it has become easier to say "it was generated by AI". This is called the "liar's dividend". Doubt is not distributed equally; it serves whoever has the strongest resources and media.
What the platform can do from inside
These are the real subject of the Brazil case. None of them is visible from outside.
Throttling reach (shadow banning). A post is not deleted, but it is shown to fewer people. The user does not notice; they only wonder why their engagement has dropped. With classifiers such as "political content", "sensitive content" and "low quality", platforms give billions of posts an invisible mark every day. These marks are explained to no one.
Tuning search and recommendations. Which account comes up when a candidate's name is typed, who is recommended in the "you might also want to follow" box, which word appears in autocomplete: each is a software decision. In Brazil, Lula not appearing even under his full name, Flávio being recommended, and the rival channel coming up in a WhatsApp channel search all happened at this layer. In Turkey in 2023 we also saw a doctored video make its way into Google's suggestions for searches on "Kılıçdaroğlu".
Ad delivery and automated account blocking. How many people a paid ad will reach, at what speed and at what price is decided by the platform's delivery algorithm. An account's ads can be stopped automatically on grounds of "suspicious activity". A one-day stoppage in the last week of an election is a loss that cannot be made good. The TSE's ruling that "automated blocks must go through human review" is aimed at this layer.
Asymmetric refusals by generative AI. If an AI tool embedded in the platform generates images for one candidate and refuses the other, this is presented as a "safety filter" problem. But the result is that one candidate has a free campaign designer and the other does not. Who wrote the filter, and by what criteria, is not disclosed.
Participation buttons. As the 2010 experiment showed, even who is shown the "I Voted" button can affect an election. If the button is shown in one region and not in another, a difference in turnout can be created.
What this layer has in common is this: each of these can change with a single configuration setting, a single model update or a single priority decision, and none of them is public.
What the state has
Country-based withholding (geo-blocking). An account is not deleted, but it is made invisible from a particular country. On X this is called "withheld in Turkey". As we said in From Obscenity to National Security: this is not a court seal but an API call. A database field changes, and eighty-five million people can no longer hear a voice.
Bandwidth throttling. Internet service providers identify traffic to particular platforms using a technique called "deep packet inspection" (DPI) and slow it down. The platform does not shut down, but videos will not open and photos will not upload. It is made dysfunctional without appearing to be banned. As well as being a censorship tool, it is a threat used to bring a non-compliant platform into line.
Mandatory identity verification. Tying social media accounts to the state's identity system abolishes anonymity. Who uses which account is collected in a database. This both creates a chilling effect and concentrates in the state's hands the knowledge of which account belongs to whom.
Summary table
| Tool | Who uses it? | Is it visible? | Who oversees it? |
|---|---|---|---|
| Social graph harvesting, micro-targeting, lookalike audiences | Campaigns, consultancy firms | No | Mostly no one |
| Dark posts, A/B testing | Campaigns | Only to the targeted person | The platform (by its own rules) |
| Fake account networks, engagement inflation | Parties, states, paid teams | Partly | The platform (selectively) |
| WhatsApp bulk broadcasting | Campaigns, employers | No | No one |
| Doctored videos, deepfakes | Everyone, most of all those with resources | Yes, but not that they are fake | Fact-checking networks (late) |
| Reach throttling, search/recommendation tuning | The platform | No | The platform |
| Ad delivery, automated blocking | The platform | No | The platform |
| AI filters | The platform | Partly | The platform |
| Country-based withholding | The state + the platform | Yes | Criminal judgeships of peace (after the fact) |
| Bandwidth throttling | The state | Yes | No one |
| Identity verification | The state | Yes | The state |
Read the last column of the table. Oversight is either absent, or lies with the platform itself, or lies with the state itself. Independent oversight by voters, candidates or the public appears in no row.
Turkey: What Happened in Past Elections?
Now let us come to our own country. Most of these tools have been used in Turkey too. But the mechanism works differently from Brazil.
A documented troll network (2020)
In June 2020, Twitter shut down 7,340 accounts that it said were linked to the AKP Youth Branches. The Stanford Internet Observatory examined the network and published a report: the accounts were amplifying party messages in a coordinated way and targeting the opposition. However, this was the last major takedown a platform carried out in Turkey of its own accord. No similar clean-up followed.
The 2023 election: doctored videos, smears, censorship
The 2023 election was one in which almost all the tools described in this piece were used at the same time (GMF Alliance for Securing Democracy).
- The doctored video. On a rally screen, Erdoğan showed a doctored video in which Kılıçdaroğlu's campaign song had been added to 2022 footage of PKK leader Murat Karayılan. When it emerged that the video was doctored, he said "it doesn't matter whether it is doctored or not". The video made its way into the top Google suggestions for searches on "Kılıçdaroğlu".
- Smear and withdrawal. Muharrem İnce withdrew his candidacy three days before the election, citing a fake sexually explicit video alleged to show him.
- An unproven Russia claim. Kılıçdaroğlu accused Russia of spreading doctored and deepfake content. No evidence supporting this claim was shared with the public.
- Twitter's compliance. The day before the election, Twitter blocked access to some accounts in Turkey at the government's request. Musk defended this by saying "the choice is have Twitter throttled in its entirety or limit access to some tweets" (JURIST, The New Arab). This threat had a legal basis: Law No. 7253, passed in 2020, allows the bandwidth of non-compliant platforms to be throttled by up to 90 per cent.
- Imbalance in the mainstream media. Manipulation on social media must be read within this picture. Between 1 April and 11 May, TRT gave Erdoğan 49 hours of airtime, Bahçeli about 29 hours and Kılıçdaroğlu 32 minutes. For the opposition, social media was not a choice but the only channel left. And we saw that this channel too could be narrowed on the eve of the vote.
Bandwidth throttling has become a habit
After the 2023 earthquake, and in March 2025 when İmamoğlu was detained, access to social media was slowed for hours. İmamoğlu's X account was also blocked in Turkey. These did not happen during an election period. But the result is the same: the channel through which a candidate reaches their voters can be closed by a decision of the state. Let us restate our line here once more: we do not defend İmamoğlu, but everyone needs the rule of law. The method applied to a mayor today is applied tomorrow to a union, a newspaper, a left-wing party.
The data infrastructure is ready
In 2016, the identity data of roughly 50 million people was leaked from MERNİS, the central civil registry system. This data has been in circulation for ten years, and "query panels" are built on top of it (our KVKK piece, The Aydo Software Operation). Political parties have legal access to voter lists. The mass text messages that land on phones during election periods are the result. There is no documented micro-targeting operation on the scale of Cambridge Analytica in Turkey. But all the materials needed for one are lying in plain sight.
What we do not know
In Turkey there has been no leak of internal messages like the one in Brazil. There is no document showing that Meta's or X's algorithm has favoured a candidate in Turkey. We cannot say this has not happened, but it is certain that we do not know. Platforms do not publish a detailed election transparency report for Turkey. Not knowing is itself the absence of oversight.
Two Models: Brazil and Turkey
Placing Brazil and Turkey side by side, we see the second ballot box working in two different ways.
| Brazil | Turkey | |
|---|---|---|
| Relationship between platform and state | Conflict: the judiciary tries to rein in the platform | Alliance: the state decides, the platform implements |
| Source of the intervention | The platform's own systems (recommendation, delivery, filters) | The state's demands (Article 8/A, access bans, bandwidth throttling) |
| Visibility | Invisible; exposed only by a leak | Visible; applied openly with the "withheld in Turkey" label |
| The platform's stance | "Technical error" | Silence and compliance (X implements more than 90 per cent of orders) |
| Legal counterweight | The electoral court issued an equality order and a fine | The blocking is approved by a criminal judgeship of peace, as always |
| Common outcome | Whoever holds the channel of the election holds the power to influence it |
What this table teaches is that "is the platform bad, or is the state bad?" is the wrong question. In Brazil the platform acts against the state; in Turkey it acts together with the state. In both cases, the communication channel between voters and candidates is outside public oversight. Monopoly and state may confront each other or join hands. But in no case does this channel belong to the people.
What Should We Expect in Future Elections?
According to the calendar, the next general election is in 2028. For the president to be able to stand again, however, Parliament must vote for an early election with 360 votes. Whenever the election is held, the ground is already being laid.
The censorship infrastructure is being built before the election
According to the records of the Freedom of Expression Association (İFÖD), at least 1,270 X accounts were blocked in September 2026 alone. These accounts had a combined 35 million followers. X implemented at least 90 per cent of these orders (HRW). Law No. 7590, which came into force on 31 July, transferred the Article 8/A blocking power from the Information and Communication Technologies Authority (BTK) to the Cyber Security Directorate, which reports directly to the Presidency. On 24 September the Ministry of Justice announced that it had set up special prosecution bureaus to combat disinformation. On 29 September T24 was shut down. On 14 October, the X account of the Workers' Party of Turkey (TİP), a legal political party, was blocked under Article 8/A (Medyascope). Even the account of EngelliWeb, which counts access bans, was made invisible from Turkey (Medyascope).
If a political party's account can be shut down outside an election period, it is not hard to guess what will happen during one. The tools are ready, and the trial runs have been done.
Anonymity is being abolished
With the regulation coming into force on 1 November, social media accounts are being tied to age verification through e-Devlet, the state's e-Government portal. The deadline for existing accounts is 31 March 2027 (Diken). The justification for the regulation is protecting children. But as İFÖD has warned, for it to be enforced, adults must also verify their identity. The result: which account belongs to whom will be gathered in a single state system. The pattern we showed in Whose Family Is Safe? is at work here too: a justification no one would want to oppose (children) is chosen, the tool is normalised, and its scope expands.
For everyone who posts ahead of an election to have their identity in the hands of the state means people will be more careful about what they write. The most efficient form of censorship is self-censorship.
AI fakery is getting cheaper
The doctored video of 2023 was crude work, and it was quickly recognised as fake. In the next election, voice clones and realistic video will be in everyone's hands. We expect two results: fake content will increase, and genuine recordings can also be dismissed as "generated by AI". Both effects serve whoever has the strongest resources, media and judiciary.
The platforms will keep quiet
X's owner is close to the Trump administration. Meta has left even HRW's letters unanswered. It is not realistic to expect resistance from companies that do not want to lose the Turkish market and are under threat of bandwidth throttling. No counterweight mechanism like the TSE ruling in Brazil is in sight in Turkey. In Brazil, the court can demand equality from the platform. In Turkey, it is the state itself that demands inequality from the platform.
The real intervention may not be on social media
All the attention of this piece is on social media. But we should not exaggerate. In Turkey, who can stand as a candidate is determined in the courts before it is determined on social media: a candidate's diploma case, the CHP congress case to be heard again on 21 October, the trustees appointed to municipalities. Social media manipulation complements these processes. Its main function is less to persuade voters than to prevent the rival from reaching their voters and to spread doubt about the result.
Empirical research supports this: the persuasive power of individual disinformation campaigns is lower than assumed. The real effect comes not from any single campaign but from algorithms that, over years, reward anger, polarise and shape the public sphere. An election begins years before that election.
The Official Narrative / A Class Reading
| Official narrative | Class reading |
|---|---|
| "It was a technical error, and it has been fixed" (Meta) | An error is political too: which error is fixed how quickly is the system owner's priority |
| "Our systems were not designed to favour anyone" | Design is measured by outcomes, not intentions; the outcome ran in one direction in the most critical week of an election |
| "We are fighting disinformation" (the state) | What is being fought is not disinformation but voices outside the government: newspapers, unions, parties, rights organisations |
| "Age verification to protect children" | What is protected is not the child but an unobjectionable justification for abolishing anonymity |
| "Foreign states are interfering in our elections" | Foreign interference is real, but the biggest interference is domestic: the monopoly that owns the channel and the state that directs it |
| "The platforms are complying with the law" | The law being complied with is one that postpones judicial review; the platform is the state's bailiff |
| "Social media is a free public sphere" | The space that looks public is private property; the landlord closes the door and dims the light whenever he likes |
| "Elections are won at the ballot box" | Elections are also won at the second ballot box, which decides who sees what on the way to the first |
Concrete Tasks
Keep records. Systematically record blocks, drops in reach and ad rejections. In Brazil the only source of information was a leak. In Turkey the only source of information is volunteer archives such as EngelliWeb. If you work on a campaign or in an organisation, keep reach data every day with screenshots and dates. Interference that becomes countable becomes contestable.
Monitor ad archives. Tools such as the Meta Ad Library partly disclose who paid for political ads, how much was spent and which audience they were shown to. Monitoring these archives regularly during the election period makes some of the dark posts visible.
Build fact-checking networks now. Build ties with outfits such as Teyit and Doğruluk Payı. As an organisation, decide in advance whom to contact, through which channel and how quickly when a fake audio recording or video emerges. A fact-checking network set up in the last week of an election is already too late.
Build channels that do not depend on the platforms. Mailing lists, your own website, RSS, Signal or Matrix groups, printed bulletins. These are not nostalgia but a backup strategy. An organisation whose only communication channel sits on one company's servers is that company's hostage.
Do not leave ballot security to social media. Volunteer ballot-monitoring organisations such as Oy ve Ötesi (Vote and Beyond) are the most concrete safeguard against a result being clouded by doubt. The murkier the second ballot box, the more important the observer at the first.
Sharpen the list of demands. Write the following into the programme of your union, your party, your professional chamber:
- independent, real-time oversight of algorithmic distribution during election periods,
- an obligation for platforms to publish country-level election transparency reports,
- a ban on political micro-targeting and dark posts,
- a ban on access blocking and bandwidth throttling during election periods,
- no tying of social media accounts to the identity system,
- in the long term, taking the public communication infrastructure out of the hands of private monopolies: open protocols, publicly or commonly owned platforms.
If you are a computing worker, take on the technical load. Keeping the website of a union, an association or a party up and running, setting up mirror sites, taking archives, measuring reach: this is our job. Every one of the tools described in this piece is written, run and configured by our colleagues. In Brazil, too, it was Meta's Brazilian employees who raised the alarm. If you work at a platform, do not forget that being able to say "this setting works in one direction" is only possible as an organised worker. We discussed this at length in The Computing Worker's Handbook.
Dear Young Comrades,
The most striking aspect of the Brazil case is not that Meta restricted Lula. It is that this could only be learned once a few employees' messages leaked. Whether a country's election was fair was known not to that country's citizens, but to the internal messaging system of a company in another country.
In Turkey the situation is even plainer. Here there is not even any need to hide it. Decisions are taken, companies implement them, accounts disappear under the "withheld in Turkey" label. In neither country is the key to the second ballot box in the hands of the people.
A century-long struggle was waged to protect the first ballot box: the secret ballot, independent election boards, observers, counting records. Each was once a demand dismissed as "unnecessary" or "unworkable". For the second ballot box, the same struggle has only just begun. Oversight of the algorithm, making the channel public, protecting identity look like distant demands today. So, once, did the secret ballot.
Whoever decides what voters see before they go to the ballot box is the real winner of the election. Taking that decision away from the owner of the box is the task of democracy in this century.
This piece was written on 17 October 2026. It will be updated after the 21 October make-good deadline in Brazil, the ruling of the full TSE and the run-off on 25 October.
Related Pieces
Censorship and digital surveillance
- From Obscenity to National Security: The Anatomy of a Forty-Eight-Hour Wave of Censorship: the arrival of the Cyber Security Directorate on the scene, and the thesis that "withheld" is an API call
- They Shut Down What They Cannot Buy: the story of the T24 ban and the class roots of censorship
- Whose Family Is Safe?: moral panic and the law of expanding scope
- The Panopticon's Doorkeeper: surveillance transparent downwards, opaque upwards
Data and privacy
- KVKK's "80s Trend" Warning and Privacy in the Digital Age: our data in the hands of the state
- The Aydo Software Operation: query panels and union data
- Whose Eye Are the Glasses?
Technology monopolies
- What Does Mark Zuckerberg Actually Want?: Meta's feed algorithm and its power to shape our goals
- The Revolt of Crystallized Labor: the concept of digital enclosure
- The Computing Worker's Handbook
Law and politics
Sources
Brazil
- Al Jazeera, "Meta employees say Lula's accounts restricted in Brazil election: Report", 16 October 2026
- ICL Notícias, "In internal messages, Meta admits Lula's accounts were harmed by restrictions", 16 October 2026
- Revista Fórum, "Meta's internal messages reveal restrictions on Lula and irregular recommendation of Flávio Bolsonaro", 16 October 2026
- Contee, "Meta admits restrictions on Lula's accounts and irregular recommendation of Flávio Bolsonaro"
- BPMoney, "Meta's internal messages acknowledge that Lula's accounts were affected"
- France 24, "Brazil's Lula accuses social media giant Meta of election interference", 11 October 2026
- WION, "Lula campaign accuses Meta of election interference after social media content blocked", 11 October 2026
Historical background and research
- Eady et al., "Exposure to the Russian Internet Research Agency foreign influence campaign on Twitter in the 2016 US election and its relationship to attitudes and voting behavior", Nature Communications, 2023
- Information, Communication & Society, "Fixing election interference, safeguarding election integrity? Developments and dynamics of election-related policies on social media platforms", 2026
- Brennan Center, "Digital Disinformation and Vote Suppression"
Turkey
- Human Rights Watch and 20 organisations, "Türkiye: End Internet Censorship", 12 October 2026
- Medyascope, "Access ban imposed on TİP's social media account", 14 October 2026
- Medyascope, "EngelliWeb's X account made invisible from Turkey", 14 October 2026
- Euronews Turkish, "The X account of Yaman Akdeniz, who tracks access bans, has been blocked", 24 September 2026
- Diken, "Age limit comes to social media on 1 November: verification via e-Devlet"
- Freedom of Expression Association (İFÖD), EngelliWeb
- GMF Alliance for Securing Democracy, "With Traditional Media in the Tank for Erdoğan, Social Media Became a Crucial Battleground in Turkish Elections", 2023
- JURIST, "Türkiye users' access to Twitter restricted ahead of presidential election", May 2023
- The New Arab, "Musk defends move to 'censor' Twitter in Turkey during elections", May 2023
- Hürriyet, "CHP congress irregularities case postponed: defendants to appear before the judge on 21 October"
This piece was published on Knowledge Commons. You may reproduce it under the CC BY-SA 4.0 licence, with attribution. Reproduce it.
