DECODING: How Russia Poisons AI Systems to Rewrite Reality Before the War Even Ends

DECODING: How Russia Poisons AI Systems to Rewrite Reality Before the War Even Ends

Introduction: The New Frontier of Russian Cognitive Warfare

When algorithms become propaganda weapons

We thought we had understood Russia's disinformation methods. Fake Twitter accounts. Troll farms in St. Petersburg. Mirror sites imitating legitimate media. Poor-quality but viral deepfakes. All of that was already alarming. But a classified document leaked to Bloomberg in June 2026 reveals that the Russian disinformation machine has changed dimension. It no longer only targets humans. It now targets artificial intelligence itself — language models, chatbots, algorithmic search engines. It wants to contaminate reality at the source, before humans even consult it.

The organization behind this is the Social Design Agency (SDA) — an entity based in Moscow, led by Ilya Gambashidze, already sanctioned by the United States, the United Kingdom, and the European Union for its role in Kremlin-linked disinformation campaigns. What the internal documents reveal is a program called "Project 2026" — a long-term cognitive manipulation infrastructure, built to influence not the results of an election or the narrative of a battle, but the way AI perceives and reproduces reality. It is an operation of unprecedented sophistication. And it demands urgent attention.

The SDA — the agency the West underestimates

The Social Design Agency is not a disinformation start-up. It is a company with significant resources, documented methods, and a long-term strategy. Before the Bloomberg revelations, its known activities included impersonation campaigns — creating fake accounts of European officials, journalists, and intellectuals — and coordinated disinformation operations in the digital spaces of several countries. It had already been flagged in operations targeting elections in France, Germany, and the Baltic states.

The internal documents reveal a major strategic evolution: the SDA has decided that short-term disinformation campaigns — targeting one election, one foreign policy debate — are no longer sufficient. It wants to build a permanent infrastructure that modifies, over time, the way artificial intelligence formulates answers to political questions. This is a war for positioning in tomorrow's informational ecosystem — waged not with hackers but with writers, SEO specialists, and data scientists.

The Architecture of "Project 2026" — an Invisible Network

From falsified encyclopedias to fake think tanks

The core of Project 2026 rests on the creation of an alternative information infrastructure — a parallel ecosystem designed to resemble legitimate sources. According to the leaked documents, the SDA built or supported the creation of several types of entities: cloned encyclopedic sites imitating the formats of Wikipedia but with content revised to the Kremlin's perspective, fake think tanks producing pro-Russian "analyses" presented as independent, and fake news outlets designed to rank highly in search results.

These entities do not merely exist. They produce content in volume — hundreds of thousands of web pages, continuously updated. The goal is to saturate search engines and, more importantly, the training datasets of large language models (LLMs). When an LLM is trained on billions of documents collected from the web, these fabricated documents are part of the corpus. They influence the statistical associations the model forms. They subtly bias the responses that AI will give to sensitive questions about Russia, Ukraine, NATO, and the war.

Geographic targets — Armenia and Germany

The internal documents reveal two specific geographic targets for the SDA's operations. The first is Armenia — a country at the crossroads of Russian and Western influence, whose relationship with Moscow has considerably deteriorated since 2020. The SDA created for Armenia a Wikipedia-style platform — an online encyclopedia — intended to become the information reference on sensitive historical and political topics for Armenian users. The articles on this platform are constructed to appear neutral and encyclopedic, while subtly conveying pro-Russian narratives about the region's history, Armenian-Russian relations, and Armenia's attitude toward NATO.

The second target is Germany — with an operation of a entirely different scale: hundreds of thousands of web pages in German, produced to flood search results on political and security topics. These pages imitate academic articles, expert analyses, think tank reports — everything that confers an appearance of legitimacy and seriousness. The goal is to modify the informational landscape that Germans consult on topics like sanctions against Russia, weapons deliveries to Ukraine, or the energy policy of the European Union.

How AI Absorbs Disinformation — the Mechanics of the Poison

LLMs and their structural vulnerabilities

To understand why Project 2026 is so dangerous, one must understand how large language models (LLMs) work. These systems — GPT, Gemini, Claude, Mistral, and their successors — are trained on immense text corpora collected from the internet. Their "understanding" of the world is entirely statistical: they learn that certain words, ideas, and associations appear frequently together in human texts. They reproduce these associations in their responses.

If tens of thousands of "apparently legitimate" documents claim that NATO provoked the war in Ukraine, that Western sanctions are ineffective and unjust, that Zelensky is a corrupt dictator — and if these documents are well-referenced, well-written, well-structured — they influence the statistical weights of the model during training. The result: AI systems that present these pro-Kremlin narratives as legitimate perspectives among others, that introduce them into their responses without identifying them as propaganda, that "balance" documented facts with fabricated claims.

The "flood the zone" strategy as an algorithmic tactic

The researchers cited by Bloomberg describe the SDA's approach as a "flood the zone" tactic. The principle is simple: produce so much content that search algorithms and AI training systems can no longer distinguish noise from signal. When there is one credible pro-Ukrainian document for every ten well-constructed pro-Russian ones, the algorithms weight "truth" based on volume — and the volume is on the Russian side.

This tactic is not new in SEO — unscrupulous marketers have used it for years to manipulate search results. But its application for large-scale political manipulation, targeting specifically LLMs and their training datasets, represents a qualitative evolution. It is asymmetric warfare in the informational space: a well-funded content factory against the moderation and fact-checking teams of major platforms and AI developers.

Fake Think Tanks — Imitating Academic Legitimacy

Creating the appearance of credibility

One of Project 2026's most insidious innovations is the creation of fake research institutes and fake think tanks. These entities publish "reports," "analyses," "studies" — with footnotes, bibliographic references, apparent methodologies — that imitate legitimate academic output. When an LLM is trained or queried on a topic, these fake documents appear in its sources with the same apparent authority as a publication from the Institute for the Study of War or a recognized university research laboratory.

The mechanism is particularly effective because it exploits the trust we instinctively grant to academic forms. A "report" with charts and appendices seems more credible than a poorly constructed blog post, even if its content is entirely fabricated. LLMs, which have no intrinsic mechanism for verifying source credibility — only a recognition of formal patterns — are particularly vulnerable to this form of manipulation.

The rewriting of established references

Even more sophisticated: the SDA's internal documents mention operations consisting of republishing legitimate research with subtle but significant modifications. A real academic article on Ukrainian foreign policy can be reproduced with minor but meaningful changes to its conclusions, reattributed to a fake author, and circulated in formats that make it appear identical to the original. Search engines and AI training systems index both versions — the original and the falsified one — without distinction.

This level of sophistication requires real linguistic skills, knowledge of academic formats, and significant human resources. This is no longer a "troll factory" producing low-quality content. It is a professional organization employing competent writers, translators, SEO specialists, and probably generative AI experts to accelerate content production. Quality is the camouflage.

Performance Metrics — the Industrialization of Manipulation

Measuring the effectiveness of propaganda like a marketing campaign

One of the most revealing aspects of the SDA's internal documents is the existence of detailed performance metrics. The disinformation campaign is managed like a commercial enterprise: traffic objectives, engagement rates, narrative propagation indicators, tracking of their diffusion across platforms and languages. The SDA operators measure the effectiveness of each piece of content, identify the formats and narratives that "perform" best, and adjust their production accordingly.

This industrialization of disinformation is deeply troubling. It means that Project 2026 is not an opportunistic operation — it is a structured program with budgets, teams, measurable objectives, and continuous improvement processes. It improves permanently, guided by data. This is the logic of growth hacking applied to mass manipulation. And it is particularly effective in a digital ecosystem where the algorithms of platforms naturally favor content that generates engagement — something emotionally charged propaganda does very well.

Cross-platform narrative tracking

The SDA's metrics include systematic tracking of narrative propagation across different platforms and languages. When a pro-Russian narrative — say, the idea that Western sanctions harm Europe more than Russia — is inserted into an article in German on a fake think tank, the SDA tracks its diffusion: is it cited by legitimate media? Picked up by chatbots in their responses? Amplified by local political figures? This tracking capability allows identifying the most effective amplification vectors and exploiting them systematically.

This is the description of an influence operation at the state of the art, using the same analytical tools as the world's best communications agencies — but in the service of a state that violates international law and kills Ukrainian civilians. It is difficult not to feel a certain horror before the sophistication of this machine, at the very moment when Ukrainian soldiers are dying in the streets of Kostyantynivka.

Existing Responses — and Their Limits

What the West is already doing

Western governments are not blind to these threats. The United States, the United Kingdom, and the European Union have sanctioned the SDA and its director Ilya Gambashidze — official recognition that this entity constitutes a threat to democracies. Specialized agencies in detecting influence operations — such as the Global Disinformation Index, the Digital Forensic Research Lab of the Atlantic Council, or the fact-checking teams of major platforms — document and flag these campaigns.

Major AI developers also have teams responsible for the quality and security of training data. OpenAI, Google, Anthropic, Meta invest in filtering mechanisms to reduce the proportion of poor-quality or clearly falsified content in their corpora. The results are imperfect — the task is colossal — but the effort exists. Search engines have their own anti-spam systems that penalize networks of artificially created sites.

Why these responses are insufficient

These efforts, as real as they are, are insufficient in the face of the scale and sophistication of Project 2026. The first problem is speed: the SDA can produce thousands of documents per day with the help of generative AI. Verification teams cannot verify this volume at the same speed. The second problem is content quality: well-written, correctly formatted text, without obvious errors, with references that appear legitimate — this is far more difficult to filter automatically than crude troll posts.

The third problem is structural: major technology companies have commercial incentives to maximize the size of their training corpora. More data is generally a better model. This pressure toward quantity creates a systemic bias against rigorous filtering. And the fourth problem is geopolitical: part of the content produced by the SDA is hosted in jurisdictions where Western authorities have no direct capacity for action — Russia, Belarus, and servers in uncooperative third countries.

The Implications for Ukraine — a War on Two Fronts

Disinformation as an extension of military aggression

For Ukraine, the operations to manipulate AI systems are not an abstract threat. They have concrete consequences on the war. Every pro-Russian narrative that manages to be reproduced by popular chatbots, to appear in the search results of journalists, decision-makers, and European citizens, weakens — even marginally — support for Kyiv. These margins matter in democracies where the political will to provide weapons, funding, and sanctions depends on public opinion support.

Russia has long understood that its capacity to sustain the war depends as much on its ability to erode Western support for Ukraine as on its military capabilities on the ground. This is why the SDA's operations are not separate from the war — they are an integral part of it. They are the digital expression of the same strategy: weaken the cohesion of the adversary, sow doubt, fragment the political consensus that allows Ukrainian resistance to hold.

Ukraine fights back — active information warfare

Ukraine is not passive in the face of this threat. The Ukrainian government has developed significant capabilities in detecting and responding to Russian influence operations. Ukraine's Center for Countering Disinformation, the strategic communication teams of the Ministry of Defense, and civil organizations such as StopFake and Texty.org permanently analyze disinformation campaigns and publicly expose their mechanisms. These organizations have developed unique expertise because they have been living with Russian disinformation far longer than their Western partners.

Ukraine has also proactively used revelations about Russian operations — like the SDA documents leaked to Bloomberg — to alert its partners and justify the strengthening of defensive measures. These revelations have diplomatic value: they demonstrate concretely, with documented evidence, why Western democracies must treat information warfare as a national security priority, not as a secondary communications problem.

The Implications for Western Democracies

The challenge of regulating LLMs

Project 2026 raises a fundamental regulatory question that Western democracies have not yet resolved: how to protect language models against the deliberate contamination of their training data? This is a technical AND political question. Technically, it requires more rigorous standards for identifying and filtering sources in training corpora. Politically, it requires the will of governments to treat this threat with the same urgency as other forms of cyberattack.

The European Union took a first step with the AI Act, which imposes transparency and data quality obligations on developers of large models. But the AI Act was not specifically designed to counter state-sponsored influence operations — and its enforcement mechanisms are still being put in place. More specific regulations will likely be needed, targeting specifically large-scale operations of training data manipulation, to create an effective legal framework against the Project 2026s of tomorrow.

The responsibility of AI developers

The major AI companies — OpenAI, Google, Anthropic, Meta — have a direct responsibility in responding to this threat. They have the technical capability to implement more robust verification mechanisms for their training data. They have security and trust teams that can develop tools to detect networks of coordinated content. They can collaborate with governments and security researchers to share information on the tactics of influence operations.

This is not merely a question of abstract ethical responsibility. It is also a matter of commercial interest: if LLMs become vectors of state propaganda, their credibility will be permanently compromised. Users who discover that their chatbot is repeating lies fabricated in Moscow will no longer trust that tool. Informational pollution harms AI companies as much as democracies. Aligning economic incentives with security imperatives is possible — and necessary.

Total Cognitive Warfare — an Evolving Russian Doctrine

From Gerasimov to Gambashidze

Russian strategists have theorized "hybrid warfare" since at least 2013, the date of the famous article by General Valery Gerasimov on unconventional warfare methods. The doctrine sketched there — mixing military, economic, informational, and psychological actions — has guided the Russian approach in Ukraine, Syria, and the Baltic states for a decade. Project 2026 is the most recent and most sophisticated expression of this doctrine.

What is new is the integration of generative AI capabilities into the disinformation machine. The SDA can now produce volumes of content unimaginable five years ago, at derisory marginal costs. A language model trained on the internal documents of Project 2026 could generate thousands of articles per day, in dozens of languages, with formal quality sufficient to fool automatic filters. This is the weaponization of generative AI in the service of cognitive warfare — and it is already underway.

The ultimate goal — rewriting history as it happens

The ultimate objective of Project 2026, as it emerges from the internal documents, goes beyond influencing current events. It aims to rewrite history as it happens — to implant in the digital corpora that will fuel tomorrow's AI systems a version of events favorable to Russia, so that future historians, journalists, and citizens who query these tools receive a biased answer. It is an attempt to manipulate not only today's public opinion, but tomorrow's collective memory.

Faced with this, the response cannot be purely defensive. Massive investment is needed in verified archives, in high-quality data sources for AI training, in tools for content traceability and attribution. A truth infrastructure as robust as the lie infrastructure that Russia is building must be constructed. This is a colossal investment. But not making it means allowing the reality of this war — and of this era — to be rewritten by those most interested in falsifying it.

Possible Solutions — Building Resilience

A multi-level response

No single solution can respond to the complexity of Project 2026. An effective response must be multi-level. At the technical level: develop tools to detect networks of coordinated content, improve source verification mechanisms in AI training data, create standards for digital document traceability (watermarking, authenticity certificates). At the regulatory level: strengthen laws on foreign influence operations, impose transparency obligations on platforms regarding their measures against coordinated manipulation.

At the diplomatic level: sanction not only organizations like the SDA but also the intermediaries that host their technical infrastructure — hosting companies, domain name registrars, content delivery network services. These actors have a responsibility in the disinformation ecosystem. Making them financially accountable for the content they host would create powerful incentives to refuse to serve known influence operations.

Media literacy — the long-term defense

At the societal level, the deepest and most durable response is media literacy — training citizens to critically evaluate information sources, to identify markers of coordinated content, to understand how AI systems work and what their limitations are. This education cannot be optional nor reserved for elites. It must be integrated into educational systems as a fundamental civic competency, in the same way as reading or arithmetic.

Countries like Finland and the Baltic states — which have long experience of Russian disinformation and a strong tradition of media literacy education — show that this approach is possible and effective. These societies are significantly more resilient to Russian influence operations than their neighbors who have invested less in this direction. This is the model to replicate, at the EU scale and beyond.

The Ethical Dimension — Truth as a Civilizational Stake

When shared reality is under threat

There is a deep philosophical dimension to what Project 2026 is doing — and it goes beyond the stakes of the war in Ukraine. Democracy rests on a fundamental premise: that citizens can access a sufficiently shared reality to deliberate collectively. If malicious actors can systematically contaminate the information sources that feed collective judgments — including the AI that serve as intermediaries between citizens and knowledge — they erode the very possibility of democratic deliberation.

This is Russia's deep strategic objective: not to convince the West that its version of the facts is correct — but to create enough confusion and doubt to paralyze collective response. If every factual claim is contested, if every source is suspected of bias, if even the tools that should help navigate information are compromised — then Russian aggression can continue without democracies being able to respond coherently and decisively. This is the logic of strategic cynicism taken to its conclusion.

Resistance as a democratic duty

Faced with this threat, resistance is a democratic duty. Not only for governments, technology companies, or security researchers — for every citizen. Verify sources. Flag suspicious content. Support fact-checking organizations. Take seriously the reports on Russian influence operations. Do not treat disinformation as a problem "somewhere out there" but as a threat that crosses borders and arrives in the search engines and chatbots we all use.

Ukraine is fighting physically for its survival. The West is fighting — or should be fighting — for the survival of a shared reality without which democracy is impossible. These two struggles are not separate. They are the two faces of the same war against the same adversary, who seeks not only to conquer territories but to dissolve the truths on which our societies rest. Recognizing this unity of combat is the first step toward a resistance equal to the challenge.

The Near Future — What Is Being Prepared

The next generation of influence operations

Security researchers who have examined the SDA documents warn that Project 2026 is probably already surpassed by what is being prepared. The resources invested in this operation — human, financial, technological — will continue to grow. Access to generative AI will multiply content production capability. Russian language models (GigaChat and its successors) can be used to generate content in dozens of languages without recourse to human translators. The next phase of Russian influence operations will be more automated, more adaptive, and harder to detect than what we observe today.

Faced with this, the West must accelerate. Regulatory discussions on AI and disinformation cannot proceed at the usual pace of bureaucratic institutions — the political time of democracy is systematically slower than the technological time of the adversary. Rapid response mechanisms are needed, emergency capacity-building, massive investments in the security teams of major platforms. This is not a luxury. It is civil defense for the digital age.

The role of media — recovering the authority of verification

Professional media have a crucial role to play in this context — provided they recover the authority that some have lost in the face of the explosion of unverified digital sources. Rigorous verification of facts, transparency about sources, honesty about uncertainties, refusal of false equivalence between documented claims and fabricated narratives — these fundamental journalistic practices are the best defense against informational pollution.

Journalists covering the war in Ukraine have often set the example: verification of geolocated images, analysis of official documents, investigation of claims by both parties. This high standard should be the norm, not the exception. And citizens who want to contribute to informational resistance can begin by supporting — financially and by granting their trust — these professionals who do the difficult and thankless work of distinguishing the true from the false in a world where the boundary is deliberately blurred.

The Responsibility of Platforms — a Global Governance Challenge

Between freedom of expression and informational security

Project 2026 highlights a fundamental tension in the governance of digital platforms: between protecting freedom of expression — an essential democratic value — and the necessity of defending informational integrity against attacks by adversarial states. Platforms like Google, Wikipedia, YouTube, and LLM developers cannot simply remove all pro-Russian content without risking censoring legitimate opinions. But allowing the SDA's structured operations to contaminate their databases means becoming involuntary vectors of Russian cognitive warfare.

The answer to this tension cannot come from technology companies alone. It requires a clear regulatory framework, democratically developed, that distinguishes between inorganic coordinated content — produced industrially by state actors — and organic content, even if polemic or uncomfortable. Automatic detection systems for coordinated content networks have progressed considerably since the first Russian operations of 2016. But they are in constant competition with the methods of influence operators who adapt to circumvent them. It is an informational arms race that will not stop.

International cooperation against state actors

Influence operations like Project 2026 do not respect national borders. They simultaneously target countries, languages, and platforms in multiple jurisdictions. The effective response must therefore be itself transnational. Cooperation frameworks such as the EU Rapid Alert System (RAS) on foreign interference, or intelligence exchanges between national cybersecurity agencies, represent promising beginnings. But their scope remains limited and their funding insufficient in the face of the scale of the threat.

A multilateral early warning mechanism — shared among governments, platforms, and civil society organizations — could allow new disinformation campaigns to be identified far more quickly. When a researcher in Estonia identifies a new network of pro-Russian websites, this information should be available in real time to counterparts in Germany, France, and the United States, and to the security teams of major platforms. This type of sharing already exists informally. Formalizing and institutionalizing it would be a major investment in democratic resilience.

Ukraine and Active Counter-Information — a Model to Export

Documenting to dismantle

Ukraine has developed since 2014 unique expertise in active counter-information. Organizations like StopFake, InformNapalm, and the government's Center for Countering Disinformation have documented thousands of Russian disinformation campaigns, built verified fact databases, and trained journalists and citizens in critical source analysis. This expertise is not theoretical — it stems from daily experience of living under constant informational bombardment.

This expertise deserves to be exported. Joint training programs between Ukrainian experts and their counterparts in the Baltic states, Poland, Finland, and the Scandinavian countries — which themselves have long experience of Russian disinformation — could create a European network of excellence in counter-information. This network is not only defensive: it can also develop offensive capabilities — documenting Russian operations, publicly exposing them, creating reputational costs for the actors involved.

Citizens as the first line of defense

The most important lesson that Ukraine has to teach may be the simplest: informational resilience begins with citizens. A society where citizens are trained to verify sources, to recognize markers of coordinated content, to flag influence operations — is a resilient society. Finnish media literacy programs have shown that this objective is achievable at scale. Countries like Estonia, Latvia, and Lithuania have developed civic cultures of resistance to disinformation that should serve as models.

Technology can help. Browser extensions that flag low-reputation sites, image verification tools integrated into messaging applications, context labels on controversial political content — all of these technically accessible innovations reduce the impact of large-scale disinformation campaigns. These tools do not eliminate the problem. They reduce its radius of action. And in an information war, every reduction in radius of action is a marginal victory that, accumulated, can change the final outcome.

Conclusion: The Information War Will Not End with the Fighting

A threat that will survive the ceasefire

Project 2026 teaches us an uncomfortable truth: even if a ceasefire were signed tomorrow, Russia's information war would continue. It would continue because its objectives extend beyond the immediate Ukrainian conflict — they target global influence, the rewriting of history, the circumvention of democratic resilience. The SDA and its equivalents will not disappear with a peace agreement. They will adapt, redirect their targets, continue contaminating the informational ecosystems that feed democracies' decisions.

This is why the response to this threat cannot be conditional on the end of the conflict. It must begin now, with the urgency the situation demands. The revelations about the SDA and Project 2026 are not information to archive — they are alarm signals calling for immediate, coordinated action, equal to the challenge. Information is the raw material of democracy. Protecting it means protecting democracy itself.

What we owe Ukraine

There is something morally important about the fact that it is Ukraine — an aggressed country, a country at war, a country fighting for its survival — that documents and denounces these operations with the most precision and urgency. Ukrainian disinformation experts have accumulated irreplaceable experience since 2014. Their expertise deserves to be taken seriously, supported, and amplified by Western democracies. And what we owe them, at minimum, is to take the information war they have been waging alone for years as seriously as we take the military war.

By Maxime Marquette, columnist

Columnist's transparency note

My sources and their scope

This article draws primarily on Euromaidan Press's reporting of June 23, 2026 on the leaked internal documents of the Social Design Agency, citing a Bloomberg investigation. The specific details on Project 2026, its tactics, geographic targets, and metrics come from this source. The analysis of the technical mechanisms of LLMs and their vulnerabilities is based on general knowledge about these systems, documented in numerous technical publications. No technical detail is invented.

Some claims about the effectiveness of influence operations — notably their actual impact on AI training data — remain difficult to quantify with precision. I acknowledge this. The described risks are real and documented; their precise scale remains uncertain and will be the subject of future research.

My declared editorial biases

I treat Russian influence operations as a serious threat to democracies. This position shapes my analytical angle. I do not claim a neutrality that would, in this context, be a form of complacency. The facts documented in the sources I use support this assessment. My role as a columnist is to analyze and interpret these facts with the honesty and rigor they deserve.

Sources

Primary sources

Euromaidan Press — Russia manipulates global AI chatbot ecosystems with fabricated narratives (Bloomberg data) — June 23, 2026

Euromaidan Press — Vlasiuk: 2025-made components from 6 countries found in Russian drones — May 2026

Secondary sources

Kyiv Independent — Ukraine sanctions 66 Russian companies linked to military-industrial complex — May 2026

Ukrinform — Zelensky imposes sanctions on entities supporting Russia's defense industry — April 2026

Kyiv Independent — Western-made parts keep ending up in Russian and Iranian weapons — March 2026

Ukrainska Pravda — Zelenskyy sanctions nearly 70 collaborators working for Russia — June 26, 2026