AI’s hallucination problem: Why smarter models are making more mistakes

The news: As generative AI (genAI) tools evolve from chatbots to smarter reasoning models, they seem to be generating more errors with hallucinations—incorrect or fabricated outputs. Error rates are as high as 79% in some tests, more than double the error rate of previous models, per The New York Times.

AI’s reality check: Increased AI hallucinations raise concerns as companies rewrite their entire business models around AI. For example, an AI bot from vibe-coding standout Cursor fabricated a policy change, triggering user backlash and subscription cancellations.

• Tools from DeepSeek and OpenAI are now more prone to factual errors than they were a year ago, researchers found.
• OpenAI’s o3 and o4-mini hallucinated 33% to 79% on benchmark tests, per the Times. 
• Errors are tied to reinforcement learning and a shift to deeper reasoning paths that can amplify step-by-step mistakes.

Potential problems: Data training sources are drying up, resulting in regurgitation of data as AI companies struggle to find new sources for model training and methods to make AI smarter.

• Research from Paper describes “model collapse,” a phenomenon where large-language models (LLMs) trained on data generated by other LLMs (rather than original human data) gradually lose touch with the true underlying data distribution. 
• This leads to an increase in repetitive, less accurate, and sometimes outright incorrect outputs over generations, even if the models appear increasingly fluent and confident.

Brand marketers are worried—35% of them around the globe cite concerns about the reliability of genAI, particularly hallucinations, as the greatest challenge to using genAI in marketing, per Econsultancy.

Our take: Marketers should treat generative AI as a tool, not a source of truth. Verify outputs, use human oversight, and choose platforms built for accuracy. Trust will drive adoption—not hype.

AI’s promise is speed and scale, but without accuracy, it erodes credibility. As genAI transforms customer service, content, and search, catastrophic errors stemming from AI overreliance will hinder adoption and diminish trust.