Despite ongoing upgrades to the reasoning capabilities of AI models, recent data shows these improvements may come at a cost: higher rates of AI hallucinations. Some of the latest language models appear to be producing more inaccurate responses than their predecessors — a troubling trend with wide-ranging implications.

AI hallucinations refer to errors where a model generates information that sounds plausible but is either false, misleading, or irrelevant to the prompt. These issues aren’t new, but rather than fading away, they seem to be growing in frequency — even as models get smarter on paper.

Recent benchmarks evaluating models like OpenAI’s newer LLMs (o3 and o4-mini) found that they hallucinate significantly more often than the earlier o1 model. In fact, the o4-mini model produced false or misleading content nearly half the time in certain tests. That’s a sharp contrast to the 16% hallucination rate seen in the o1 version from late 2024.

And it’s not just OpenAI. Data from Vectara’s public leaderboard — which tracks hallucination rates — shows similar increases in newer models from other companies, including DeepSeek. Interestingly, many of these newer models are designed to reason through complex problems step-by-step, but this added complexity doesn’t seem to help reduce hallucinations.

OpenAI has acknowledged the issue, noting that hallucinations aren’t necessarily tied to a model’s reasoning abilities. They’ve stated that research is ongoing to better understand and mitigate the problem in future iterations.

These rising hallucination rates pose real-world challenges. Whether it’s legal tools, research assistants, or customer service bots, applications that depend on reliable, factual output could be compromised if the models frequently go off track. Until hallucinations can be minimized or reliably flagged, users should remain cautious and always verify AI-generated content.