When AI Becomes a Propaganda Megaphone: The Problem With Unvetted Training Data
I’ve been watching the AI hype train for a couple of years now, equal parts fascinated and concerned. The technology is genuinely impressive in some ways, but there’s always been this nagging worry at the back of my mind about what happens when we hand over our critical thinking to machines that don’t actually think.
Recent research showing that ChatGPT, Gemini, DeepSeek, and Grok are serving up Russian propaganda about the Ukraine invasion feels like that worry manifesting in real time. It’s not surprising, but it’s deeply frustrating.
Here’s the thing: these aren’t intelligent systems. They’re pattern-matching engines trained on enormous datasets scraped from the internet. And the internet, as we all know, is absolutely flooded with misinformation, propaganda, and bots. It’s like training a chef using recipes from a cookbook where half the pages have been replaced with instructions for making poison. The chef doesn’t know the difference—they just follow the patterns they’ve learned.
Someone in the discussion I was reading made a brilliant point about this being the same old “disinformation firehose” strategy that Russians have used on social media for years, just applied to a new technology. But another commenter took it further, and this is what really got under my skin: we’re not being attacked by Russians so much as we’re being sold to them by American tech companies who monetize engagement above all else.
Think about it. These platforms could implement stricter verification processes. They could be more transparent about their training data. They could build in better safeguards. But that would cost money and slow down the race to market. Instead, we get tools that confidently spew out whatever sounds plausible based on the patterns they’ve absorbed from the cesspool of the modern internet.
The framing issue makes this even worse. Different prompts can lead to wildly different responses from the same AI system. Someone tested asking for “facts-based, nonpartisan” takes on Ukraine from multiple chatbots and got reasonable answers aligned with, you know, observable reality—that Russia is the aggressor in an unprovoked invasion. But tweak the framing, and you can get these same systems to parrot Kremlin talking points. That’s not intelligence; that’s a fancy autocomplete function with no actual understanding of truth or ethics.
Working in IT, I’ve watched the LLM hype cycle with a professional eye. The industry has rebranded what should be called “machine learning” or “large language models” as “AI,” and it drives me mad. Real artificial intelligence would involve actual reasoning, understanding, and wisdom. What we have now is sophisticated pattern matching that can’t tell the difference between a well-researched article and propaganda because it lacks the capacity to evaluate truth claims.
The environmental implications bother me too. These massive training runs consume enormous amounts of energy, all to produce systems that will confidently direct you to non-existent academic papers or synthesize Russian propaganda into something that sounds authoritative. The carbon footprint for glorified spellcheck is staggering.
What frustrates me most is watching people treat these tools as oracles rather than what they actually are: flawed, biased systems trained on unvetted data. I’ve seen colleagues use ChatGPT to research technical issues without verifying the responses, and I’ve had to gently point out that the “documentation” they’re citing doesn’t exist. The system just generated something that sounded right.
Some people are developing workarounds—specific prompting techniques to try and force more truthful responses. But the fact that you need to essentially trick an AI into not lying to you is pretty damning evidence that something is fundamentally broken in how we’re approaching this technology.
Here’s what bothers me on a deeper level: we’re outsourcing our critical thinking at a moment when critical thinking has never been more important. We’re in an age where distinguishing truth from propaganda, fact from fiction, is crucial for maintaining democratic societies. And we’re building tools that actively make that harder by laundering misinformation through a veneer of technological authority.
The solution isn’t to abandon these tools entirely—they can be useful for specific tasks when used with appropriate skepticism. But we need much more transparency about training data, better verification processes, and most importantly, we need people to understand that these systems aren’t intelligent. They’re text prediction engines. Treat them like you’d treat a random person on the internet: potentially useful, but verify everything important.
The tech industry won’t fix this on its own. There’s too much money in the current approach, too much hype to maintain. We need regulation, transparency requirements, and perhaps most importantly, we need to cultivate a culture of critical thinking and source verification. In a DevOps context, we call this “trust but verify.” The same principle applies here, except maybe lean more heavily on the “verify” part.
I genuinely believe these language models have legitimate uses—code completion, brainstorming, drafting text that you then heavily edit. But using them as search engines or reference tools without verification? That’s asking for trouble. And when that trouble involves geopolitical propaganda about an ongoing war, the stakes are far too high to shrug and say “well, that’s just how the technology works.”
We built this problem. We can fix it. But first, we need to stop pretending that autocomplete is intelligence and start treating these tools with the appropriate level of skepticism. Your brain is still the best critical thinking tool you have. Maybe try using it a bit more than you use ChatGPT.