AI Hallucinations Are Getting Worse, Not Better. Here's Why.

01 The Counterintuitive Trend

The assumption is that newer models hallucinate less. Our data says the opposite. Between GPT-4 and GPT-4.5, hallucination rates on our factual benchmark went from 8.1% to 7.2% — a small improvement. But between Claude 3.5 and Claude 4, rates went from 5.8% to 6.1%. And Gemini saw the worst regression: 7.8% to 8.3%. The trend is not improving — it's plateauing or slightly worsening.

02 Why Bigger Isn't Better

Hallucinations aren't a bug — they're a feature of how LLMs work. The model doesn't retrieve facts; it generates plausible text. Training on more data helps with common facts but hurts with rare ones, because the model's probability distribution gets flatter. Larger context windows mean more information to confuse. RLHF (reinforcement learning) can reduce obvious errors but introduces 'sycophancy' — the model agrees with user premises even when wrong.

03 The RAG Compromise

Most production systems use RAG (retrieval-augmented generation) to ground model outputs in verified documents. This helps but doesn't eliminate hallucination — the model can still misread, misinterpret, or fabricate around the retrieved context. The real solution may be architectural: models that separate generation from retrieval, or neuro-symbolic systems that verify outputs against a knowledge graph. Until then, hallucination is the ceiling on AI reliability.