The rate of false claims generated by leading AI models on news-related prompts nearly doubled in a single year. 18% in August 2024. 35% by August 2025. Not edge cases or obscure technical queries. News. The exact category of content that media companies are racing to automate.
The industry is speeding up while the accuracy is getting worse. That’s not a technology bet. It’s a credibility bet.
The numbers are worse than you think.
In Q1 2025, 12,842 AI-generated articles were pulled from online platforms for containing hallucinated content. Fabricated quotes. Invented sources. Statistics that never existed. Those are the ones that got caught. The actual number of hallucinated articles that slipped past editorial review and into real readers’ feeds? Nobody knows. That’s part of the problem.
A Deloitte Global AI Survey found that 47% of enterprise AI users admitted to making at least one major business decision based on hallucinated content. Nearly half. And the financial toll hit $67.4 billion in global losses tied to AI hallucinations in 2024 alone. That number is growing alongside adoption.
These aren’t projections. This is where we are right now. And it raises a question that most organizations deploying AI content tools haven’t answered: if you don’t know your hallucination rate, how do you know your credibility is intact?
The hallucination rate isn’t even the scariest part. It’s how AI gets it wrong.
When a journalist isn’t sure about something, you can usually tell. They hedge. They attribute. They write “according to” or “reportedly.” Uncertainty leaves fingerprints in human writing.
AI doesn’t leave fingerprints. It leaves the opposite.
Language models are actually more confident when they’re wrong, using phrases like “definitely,” “certainly,” and “without doubt” at higher rates when generating false information than when stating verified facts. This isn’t a bug. It’s baked into how these systems are trained.
The benchmarks that shape AI behavior reward confidence and penalize uncertainty. Saying “I don’t know” literally scores lower than hallucinating a confident wrong answer on nearly every major evaluation framework. Only one benchmark in wide use, WildBench, gives any credit for expressing uncertainty. And even WildBench scores “I don’t know” lower than a hallucination. The entire system is set up to produce false confidence.
It gets worse during post-training. The “alignment” process that makes AI sound helpful and conversational? It degrades the model’s ability to express doubt. Human evaluators prefer confident, decisive language, so models learn that hedging gets punished. The more polished the output sounds, the less you should trust it.
In one widely cited test, researchers asked 12 leading language models to name Mongolia’s bordering countries. Nine out of twelve confidently listed Kazakhstan. Kazakhstan does not border Mongolia. They didn’t hedge. They didn’t say they weren’t sure. They stated it as fact, same tone as telling you water is wet.
Now put that kind of false confidence on an earnings report. A breaking news story. An investigative piece. The AI doesn’t know what it doesn’t know, and it has no mechanism for telling you.
This is already causing real damage.
Deloitte submitted a report to the Australian government that contained fabricated academic sources and a fake court quote. A law professor catalogued roughly 20 errors in the document. Cost of the engagement: A$440,000. This wasn’t some scrappy startup cutting corners. This was one of the Big Four.
Stanford researchers asked leading language models about legal precedents. The models invented more than 120 court cases that never happened. Complete with realistic names like “Thompson v. Western Medical Center (2019)” and detailed legal reasoning. Convincing enough to cite in a brief. Entirely fabricated.
At ICLR 2026, one of the most prestigious AI conferences in the world, GPTZero found more than 50 submitted papers with hallucinated citations. Each paper had already passed through three to five expert peer reviewers. Most of the reviewers missed the fakes. Some of these papers had average ratings of 8 out of 10 and would have almost certainly been published without intervention. If domain experts doing careful peer review can’t catch this reliably, think about what that means for a newsroom pushing out dozens of pieces a day.
Audiences have already made up their minds. Consumer preference for AI-generated content dropped from 60% in 2023 to 26% today. Readers aren’t waiting around for the industry to fix this. They’re leaving.
“The next model will fix it.” No, it won’t.
This is the part where the conventional thinking breaks down.
Most people in media assume hallucination is temporary. That the next generation of models will be accurate enough to trust. The data tells a different story.
OpenAI’s newer reasoning models, the ones built specifically to think more carefully, hallucinate more than their predecessors. Not less. The o3 model hallucinates 33% of the time on person-related queries. That’s double the rate of the earlier o1. The smaller o4-mini is even worse at 48%. More sophisticated architecture, worse factual accuracy.
This makes sense once you understand what language models actually do. They don’t look up facts. They generate statistically probable text. Making them more capable makes them better at producing text that sounds right, which can make their mistakes harder to spot, not easier.
And there’s a compounding problem. As the internet fills up with AI-generated content, models are increasingly training on synthetic data, including other models’ hallucinations. Researchers have documented progressive quality degradation across model generations trained this way. The training data is getting polluted, and the pollution speeds up with adoption.
Waiting for a model that doesn’t hallucinate is waiting for something the architecture can’t deliver. If your content strategy depends on next year’s model being reliable enough to trust with your brand, you’re making a bet with very unfavorable odds.
The question everyone should be asking.
Most of the industry is focused on the output side, trying to catch and correct hallucinations after they’ve already been generated. Better fact-checking tools. Human review layers. Disclosure labels. These help, but they’re incomplete. You’re playing defense on the wrong end of the field.
What happens when you move quality control upstream? Instead of generating content and then checking whether it’s true, what if you controlled what the AI had access to in the first place? Find the data yourself. Validate it before it enters the system. Structure it so the AI works from verified signals instead of the open web.
That doesn’t replace editorial judgment. But it changes the equation. Hallucination becomes structurally difficult instead of statistically inevitable. The output starts from a foundation of trust instead of hoping to arrive at one.
This isn’t theoretical. It’s an engineering decision that some companies are already making. And it changes the math on what AI content can actually be.
The hallucination problem isn’t going away. But the companies that solve it won’t be the ones waiting for better models. They’ll be the ones who decided to control the inputs.
