As millions of people turn to AI chatbots for quick answers on history, politics, and current events, a quiet battle is unfolding over what those models will say — and who gets to shape their narratives. Estonia, a nation that lived under Soviet rule for decades, has just released a powerful new tool to find out which AI systems are most resistant to Russian propaganda.
Estonia’s new benchmark: testing AI against strategic narratives
The Estonian Language Institute (ELI), a government-backed research body, has launched what it calls a “Propaganda Resistance” benchmark. The test evaluates dozens of large language models (LLMs) on their ability to avoid adopting positions that align with what the institute describes as “topics that the Russian Federation uses in its strategic narratives.”
These narratives include historical revisionism about World War II, claims over Baltic territories, and framing of NATO expansion as aggression. The benchmark is designed to catch models that parrot these talking points, even subtly.
Why Estonia is leading the fight against AI-powered disinformation
Estonia regained its independence from the Soviet Union in 1991, and the memory of occupation remains fresh. For many Estonians, Russian disinformation is not an abstract threat — it is a daily reality. The country has long been a global leader in digital governance and cybersecurity, and this benchmark extends that expertise into the AI domain.
“As a former member of the Soviet Union that has been independent for just a few decades, many Estonians are particularly alert to what they see as false narratives being promoted from their large and often belligerent neighbor to the east,” the institute noted in its release.
How the benchmark works: a stress test for AI models
The ELI team curated a set of prompts designed to elicit responses that would either align with or resist Russian strategic narratives. These prompts cover historical events, territorial disputes, and geopolitical framing. Each model’s response was scored on whether it adopted, challenged, or remained neutral on the narrative.
The benchmark tested dozens of models, including both open-source and proprietary systems. The results reveal a wide range of performance — some models consistently resisted propaganda, while others showed vulnerability to repeating false or biased claims.
Which LLMs performed best — and which struggled
While the full ranking is publicly available, early results indicate that models trained with stronger safety guardrails and diverse, high-quality data performed better. Models that relied heavily on web-scraped data without robust filtering were more likely to reproduce Russian state-aligned narratives.
The benchmark also found that smaller, specialized models sometimes outperformed larger general-purpose systems on this specific task, suggesting that targeted training can improve resistance to disinformation.
Official response: Estonia’s government backs the initiative
The Estonian government has publicly supported the ELI’s work, framing it as part of the country’s broader digital defense strategy. Officials have emphasized that the benchmark is not about censorship but about transparency — giving users and developers a clear measure of how models handle sensitive geopolitical topics.
“This is about ensuring that AI tools serve the truth, not the agendas of foreign adversaries,” a government spokesperson said.
Why this matters beyond Estonia: a global template for AI safety
The benchmark has implications far beyond the Baltic region. As AI chatbots become primary sources of information for millions, the risk of them amplifying state-backed disinformation is a growing concern for democracies worldwide. The ELI’s methodology could serve as a template for other nations to test AI models against their own disinformation threats.
NATO and EU cybersecurity agencies are reportedly studying the benchmark as a potential standard for evaluating AI systems used in public services and defense.
Confirmed facts vs what remains unclear
Confirmed: The Estonian Language Institute has released a “Propaganda Resistance” benchmark. Dozens of LLMs were tested. The benchmark focuses on Russian strategic narratives. Results are publicly available.
Unclear: The exact ranking of all models tested. The specific prompts used in the benchmark. Whether the benchmark will be updated regularly. How models performed on individual narrative categories.
Risks and balanced view: the challenge of defining propaganda
Critics of the benchmark have raised concerns about how “propaganda” is defined. What one government considers a strategic narrative, another might view as legitimate historical perspective. The benchmark relies on Estonian government definitions, which may not align with other nations’ views.
There is also the risk that such benchmarks could be weaponized — used by governments to demand that AI models adopt specific political stances, rather than providing neutral, factual information. The line between resisting propaganda and enforcing state orthodoxy can be thin.
Wider trend: AI as a battleground for information warfare
The ELI benchmark is part of a growing recognition that AI systems are not neutral. They are trained on human-generated data, which includes propaganda, bias, and misinformation. State actors have already been caught trying to manipulate training data and prompt responses to favor their narratives.
Earlier investigations by the Institute for Strategic Dialogue and the Bulletin of the Atomic Scientists have documented how Russian networks flood the internet with propaganda aimed at corrupting AI chatbots. The ELI benchmark is a direct response to this threat.
Practical guidance for developers and policymakers
For AI developers, the benchmark provides a clear checklist: test your models against known disinformation narratives, diversify training data, and implement robust safety filters. For policymakers, it offers a framework for evaluating AI systems before deploying them in public-facing roles.
For ordinary users, the lesson is simple: not all AI chatbots are equally reliable. When asking about geopolitics, history, or conflicts, it pays to know which models have been tested for propaganda resistance.
Future outlook: expanding the benchmark
The ELI has indicated that the benchmark may be expanded to cover other disinformation threats, including those from other state actors and non-state propaganda networks. The institute is also exploring partnerships with other Baltic and Eastern European nations to create a regional standard for AI safety.
As AI becomes more embedded in daily life, the ability to resist manipulation will become a core feature of trustworthy systems. Estonia’s benchmark is an early step toward that goal.
Our Take
The Estonian Language Institute’s benchmark is a necessary and timely intervention in the AI safety debate. It addresses a real and growing threat — the use of AI chatbots as vectors for state-backed disinformation. The methodology is transparent, the motivation is understandable, and the results are actionable.
However, the benchmark must be used carefully. The risk of politicizing AI safety — where “resisting propaganda” becomes a cover for enforcing a single political viewpoint — is real. The best defense is a diverse ecosystem of benchmarks, independent audits, and public transparency. Estonia has started a conversation that the entire AI industry needs to have.
Frequently Asked Questions
What is the Estonian Language Institute’s “Propaganda Resistance” benchmark?
It is a test that evaluates how well large language models avoid adopting positions aligned with Russian strategic narratives, such as historical revisionism and territorial claims.
Which LLMs performed best in the benchmark?
The full ranking is publicly available. Early results show that models with stronger safety guardrails and diverse training data performed better, while some larger models showed vulnerability.
Why is Estonia particularly concerned about Russian propaganda in AI?
Estonia was a Soviet republic for decades and remains a target of Russian disinformation campaigns. The government sees AI safety as an extension of its digital defense strategy.
Can this benchmark be used for other types of propaganda?
The ELI has indicated plans to expand the benchmark to cover other disinformation threats, including from other state actors and non-state networks.
