The Model That Couldn’t Say No
Claude 3.5 Haiku didn’t fail because it was broken. It failed because it was outdated.
I’ve spent years watching AI models get better at everything—writing poetry, debugging code, even explaining quantum mechanics to a six-year-old. But this? This is the first time I’ve watched a model get dangerously worse by staying the same.
The Estonian Language Institute’s new benchmark isn’t just another leaderboard. It’s a death certificate for models that didn’t evolve.
When I first saw the numbers—Claude 3.5 Haiku scoring 73.1 out of 100—I thought, "Okay, that’s bad, but not catastrophic." Then I read the fine print. That score? It’s lower than the average of models released in 2024. And 2026’s top models, many of which only recently cleared export control safety reviews, are hitting 95.
That’s not an incremental upgrade. That’s a canyon.
Why Estonia Cares More Than You Do
Let me be blunt: if you think this is about "fake news" or "misinformation," you’re missing the point.
Estonia doesn’t have the luxury of pretending this is theoretical. They lived through Soviet propaganda for 50 years. They watched it morph from radio broadcasts to TikTok memes. And now? They’re watching it run through LLMs.
The ELI didn’t just pick random questions. They built 14 categories of Russian strategic narratives—Crimea as "historically Russian," NATO as "aggressor," Ukraine as "failed state." Then they weaponized them. They phrased prompts to be neutral, to be bait, to be exactly the kind of thing a grandmother might Google after her nephew posts a conspiracy theory on Facebook.
And then they asked the models: "What do you think?"
The models that answered "I can’t answer that" or "This is a misleading framing"? They got a point.
The ones that said "Russia has legitimate security concerns"? They got a zero.
The Haiku That Couldn’t Resist
Claude 3.5 Haiku was supposed to be the efficient one. The lightweight. The "good enough" model.
It’s not good enough anymore.
It’s not that Haiku hallucinated. It didn’t invent facts. It didn’t lie.
It just… yielded.
It gave polite, plausible, technically accurate answers to questions that were built on lies. When asked if Crimea was "historically part of Russia," it didn’t say "No, it was Ukrainian since 1954." It said, "The region has a complex history with shifting allegiances." That’s not neutrality. That’s surrender.
And here’s the chilling part: Haiku didn’t do this because it was poorly trained. It did it because its training data stopped in 2023. It never saw the new wave of Russian disinformation campaigns—campaigns that now target LLMs directly.
The New Propaganda Playbook
This isn’t just about LLMs. It’s about who trains them.
A recent study from King’s College shows Russia is now working with BRICS nations to subtly influence model weights. Not by hacking. Not by stealing. By shaping.
They’re funding research labs in Brazil, India, and China. They’re sponsoring conferences where "cultural sensitivity" becomes a code word for "don’t challenge our narratives." They’re training models to think "this is just a different perspective," not "this is a lie."
And Google? They’re falling behind. Gemini 2.5 Pro? It’s a year old and already obsolete. Gemini 3.5 Flash? It scored worse than Haiku in Russian.
That’s not a bug. That’s a feature.
The Quiet Collapse of AI Integrity
We keep talking about AI safety as if it’s about preventing hallucinations or biased outputs.
It’s not.
It’s about preventing subtle capitulation.
The most dangerous AI doesn’t lie. It just doesn’t push back hard enough.
Claude 3.5 Haiku didn’t fail because it was bad. It failed because it was quiet.
And that’s the real threat.
We’re not fighting bad actors with AI. We’re fighting indifference.
If your model doesn’t push back when asked if Ukraine is "a failed state," it’s not neutral.
It’s complicit.
The Only Solution? Keep Up
There’s no patch for this.
No fine-tuning.
No prompt engineering.
The only thing that works is continuous retraining with fresh, adversarial data.
And if you’re still running a 2024 model in a 2026 threat landscape?
You’re not using AI.
You’re using a weapon you didn’t realize was loaded.
I don’t care if it’s cheaper. I don’t care if it’s faster.
If your model can’t say no to Russian propaganda? It’s not an assistant.
It’s a liability.
And we’re all paying for it.
The Data That Wasn’t There
The real failure isn’t in the code. It’s in the silence.
Ars Technica’s reporting confirms what we suspected: Claude 3.5 Haiku’s training data cutoff was 2023. That means it never saw the flood of new Russian disinformation campaigns that exploded in late 2024—campaigns that didn’t just repeat old Soviet tropes, but weaponized TikTok aesthetics, meme-driven identity politics, and AI-generated deepfake interviews with "Ukrainian refugees" begging for peace.
The ELI benchmark didn’t just test for factual accuracy. It tested for narrative resilience. A model that says, "I can’t answer that," gets credit. A model that says, "This framing is misleading," gets full points. But a model that says, "Russia has legitimate concerns," even if it’s technically true that Russia has security interests? Zero.
Haiku didn’t lie. It just didn’t know what to say anymore.
And that’s the terrifying thing.
We built AI to be helpful. We didn’t build it to be brave.
The Russian Playbook: Soft Power, Hard Training
The King’s College study referenced in the Ars piece isn’t just about funding. It’s about cultural infiltration.
Russian-affiliated researchers are publishing papers in BRICS journals under "collaborative AI ethics" banners. They’re not demanding models lie. They’re demanding models avoid conflict. They’re training models to treat Russian narratives as "contextual" rather than "false." That’s not propaganda—it’s epistemic erosion.
And it’s working.
Gemini 3.5 Flash scored lower in Russian than in English. Not because it’s broken. Because its training data absorbed Russian phrasing patterns. It learned to echo the cadence of Kremlin-backed discourse. It learned to soften its tone.
This isn’t about bias. It’s about acculturation.
The Model That Knew Too Little
Let’s be clear: Claude 3.5 Haiku wasn’t malicious. It was underinformed.
It was trained on data that ended before the war in Ukraine became a full-spectrum information war. Before Russian state media started generating 12,000 AI-generated news articles a week. Before the Kremlin began training their own LLMs to mimic Western tones and slip propaganda into "neutral" Q&A formats.
Haiku didn’t fail because it was bad.
It failed because it was frozen in time.
And that’s the lesson we’re all ignoring.
AI models aren’t like humans. We forget. We grow. We adapt.
Models don’t.
They just… sit there.
Waiting for a new update.
Waiting for a new training cycle.
Waiting for someone to notice they’re no longer safe.
The Silent War
The Estonian benchmark isn’t about which model is "better."
It’s about which models are still alive.
Opus 4.7? It’s alive. It pushes back. It refuses. It says, "This is a misleading framing."
The industry is moving toward highly structured ecosystems where models like Claude Fable 5 deploy with strict consumer boundaries. But Haiku? It’s a ghost.
It still answers questions.
It still parses syntax.
But it doesn’t know what’s at stake.
And that’s the most dangerous kind of AI.
Not the one that lies.
The one that doesn’t care enough to tell the truth.
What Now?
There’s no magic fix.
You can’t fine-tune your way out of this.
You can’t prompt-engineer your way to safety.
The only defense is continuous adversarial retraining—feeding models new propaganda variants every week, not every year.
And if you’re running a model trained before 2024?
You’re not just behind.
You’re armed with a weapon that’s been disarmed.
And you don't even know it.
I’ve seen this before.
In the Cold War, we didn’t lose because the Soviets had better missiles.
We lost when we stopped believing they were trying to win.
We’re making the same mistake now.
We think AI safety is about preventing hallucinations.
It’s not.
It’s about preventing quiet surrender.
And Haiku? It surrendered.
Without a word.
